Aximmetry showed an upcoming MCP server at NAB that lets any AI agent operate its virtual production software through natural language. The company is targeting a summer release alongside the next version of Aximmetry, and the install is a single configuration file that drops into any MCP-compatible agent client.

Once installed, the agent understands the Aximmetry control board. Operators can ask it to build a basic virtual production scene, move talent inside the scene, draw a virtual path, or change other control board settings without opening the panel. The agent also goes beyond the manual UI, generating components like a transition effect between two clips on request. The result is a standard Aximmetry project file, so any change the agent makes can be opened and refined the way an operator would edit it by hand.

How the MCP Server Works

The Model Context Protocol is the standard Anthropic introduced for letting AI agents call external software through a documented tool interface. Aximmetry's implementation ships as a server file that any MCP-compatible agent client can register. After that, the agent treats Aximmetry as a tool it knows how to drive.

The scope of what the agent can do mirrors the application itself. Anything an operator can adjust on the Aximmetry control board is exposed as an instruction the agent can issue. That includes:

• Scene creation. Building a basic virtual production from a prompt.

• Talent positioning. Moving subjects inside the scene.

• Virtual paths. Drawing motion paths through the volume.

• Component creation. Building elements like a transition effect between two source clips.

Because Aximmetry projects remain standard project files after the agent finishes, a producer can hand a chat-built scene to a virtual production operator for refinement, or use the agent's output as a starting point and finish manually.

MCP Adoption Spreads Across Production Tools

Aximmetry's server lands as the broader VFX and post toolchain begins exposing itself to AI agents through the same protocol. We previously covered Foundry's Griptape acquisition, which brings agent orchestration into Nuke and Nuke Stage. The pattern is consistent: vendors skip bolt-on AI features and expose their existing software interfaces to whatever agent the user already trusts.

For virtual production crews not yet running a dedicated technical director, an MCP server changes the practical entry point. A chat interface can stand up a workable scene before someone with deep Aximmetry experience does the polish pass. That fits the company's broader accessibility push, which previously brought Aximmetry Instant to NAB as a streamlined version of the platform aimed at users without prior virtual production experience.

Image Gen Lives Inside Aximmetry Today, 3D Generative Backgrounds Still Under Review

The MCP server is one piece of a longer roadmap. Aximmetry has already integrated 2D image generation into the platform: operators can ask the AI to generate a still or video for placement on a virtual screen, or have it format incoming data and render it inside the scene. A future build is expected to embed an agent inside Aximmetry itself, so users can interact conversationally without bouncing to an external client.

In the interview, Aximmetry said it is continuing to evaluate generative 3D systems for backgrounds and environments, including World Labs and NVIDIA's generative models, but has not integrated any yet. The company plans to add a 3D system once it identifies one capable of meeting its requirements.

The MCP server is targeted for release alongside Aximmetry's next major version this summer.

DreamWorks Animation's MoonRay renderer — the path-tracing engine behind every DreamWorks feature film since 2019 — has joined the Academy Software Foundation as an official hosted project. For VFX and animation professionals, this moves a battle-tested studio renderer into neutral, community-governed territory, opening the door to serious adoption well beyond DreamWorks' own walls.

MoonRay has been open source for a while. The meaningful shift here is the step up to ASWF hosting — the same governance model that underpins tools like OpenColorIO, OpenEXR, and MaterialX, all now standard fixtures across the industry's pipeline.

What's in the Vault: MoonRay has real production credentials — the renderer has a direct line to The Wild Robot, The Bad Guys 2, and every DreamWorks animated feature for the past six-plus years.

Under New Management (Sort Of)

The practical change is how MoonRay will be developed and governed going forward.

According to the announcement, MoonRay will now be maintained and further developed under ASWF with contributions from the global open source community. DreamWorks Animation — already an ASWF member — will continue providing ongoing support and dedicated engineering resources.

That last part matters. Rendering tools that lose their core maintainer team tend to stagnate quickly. Knowing DreamWorks remains actively involved while ASWF provides neutral stewardship addresses the biggest concern that comes with depending on a studio-originated tool.

For production teams evaluating external dependencies, ASWF hosting means:

• Open development processes with transparent roadmaps and community-driven priorities

• Vendor- and studio-neutral governance, reducing the risk of one organization's internal priorities derailing the tool

• Lower friction for contributions — pipeline engineers can upstream patches, tool vendors can target the renderer without negotiating direct relationships with DreamWorks

• Educational alignment — academic programs building curriculum around real production tools now have a clearly sanctioned, neutral home to point students to

What MoonRay Actually Is

MoonRay is a production path-tracing renderer — a CPU and GPU-capable system that handles both stylized and photorealistic output, with the full set of features production teams expect: AOVs, LPEs, distributed rendering, and a Hydra Render Delegate (hdMoonRay) that plugs into any DCC tool with USD Hydra support for interactive preview rendering.

That last point is particularly relevant for studios already invested in USD-based pipelines — Hydra compatibility means MoonRay can slot into existing DCC workflows without requiring custom integrations from scratch.

What separates MoonRay from other open source render options is simply where it came from: it was built inside a major Hollywood animation studio and refined over years of demanding production use, not assembled by committee or maintained primarily as a demo tool.

Where This Fits in Your Pipeline

The honest answer is: it depends heavily on your setup.

For large facilities and studios:

• MoonRay's multi-year feature film track record gives it credibility most open source renderers lack at this scale

• The realistic adoption path is probably a secondary renderer — test it on a single sequence, evaluate pipeline compatibility, then scale from there

• Studios that want to own their rendering stack without building from scratch now have a strong starting point, with an active core team still in place

For mid-sized studios and indie pipelines:

• No per-core licensing is a meaningful cost advantage for smaller operations

• Integration, however, is non-trivial — you'll need pipeline engineering time and rendering expertise to get meaningful results

• MoonRay is most likely to click immediately for teams already running Linux-heavy, open source stacks

For tool vendors and developers:

• ASWF hosting makes MoonRay a viable rendering backend to support without the dependency risk of a single-studio project

• The Hydra delegate path provides a clean integration surface for DCC and pipeline tools

For specifics on supported platforms, licensing terms, and documentation depth, the openmoonray.org project site is the authoritative starting point.

How to Train Your Renderer

The second part of the news may ultimately be more useful for practitioners than the press announcement itself.

DreamWorks Animation CTO Bill Ballew will keynote ASWF's Open Source Days in Los Angeles on July 19–20, with a talk titled "How to Train Your Renderer: MoonRay's Journey from DreamWorks' Dragons to the ASWF."

According to the announcement, Ballew will cover:

• The technical challenges of "breaking out of the studio" — decoupling a proprietary renderer from internal build systems and dependencies

• Strategy and lessons learned in building a sustainable open source community around a production tool

• Why ASWF's role is considered "vital" for MoonRay's path forward

For anyone dealing with rendering pipelines, tool development, or open source strategy in animation and VFX, this session has the potential to offer concrete, hard-won insight that a press release can't. It's also likely to reveal roadmap clarity and community-building plans that will determine how fast MoonRay picks up non-DreamWorks contributors.

Reading the Credits

The real measure of this announcement won't come from the press release — it'll come from job postings and show credits over the next few years.

ASWF hosting makes serious, broad adoption of MoonRay possible in a way that ad-hoc open sourcing does not. Whether that potential converts to actual adoption depends on how quickly the documentation matures, how actively DreamWorks and new contributors maintain the roadmap, and whether mid-sized facilities start testing it in production.

The tools to watch it happen are all public: ASWF project communications, the MoonRay GitHub repository, and Ballew's Open Source Days keynote in July. If you're in Los Angeles that week, it's worth the room.

Canon announced the EOS R6 V, the first full-frame body in its video-focused EOS V series. The camera pairs a 32.5MP sensor with 7K Open Gate capture, built-in active cooling, and a flat, rig-friendly body that drops the EVF and mechanical shutter in exchange for longer recording times and a lower price.

Key specs:

• 7K Open Gate (3:2) up to 30p in 12-bit RAW and MP4, plus 7K DCI 60p in RAW Light

• Uncropped 4K 120p for slow motion, with oversampled 4K Fine up to 60p from the full 7K readout

• Active cooling fan rated by Canon for roughly 120 minutes or more of continuous 7K 60p RAW Light, limited by power or card capacity rather than overheating

• Body-only price: USD $2,499, with shipping in late June 2026

A Video-First Body in the EOS Lineup

The EOS V series began with the smaller PowerShot V1 and EOS R50 V in 2025, aimed at creators stepping up from phones and entry-level cameras. The R6 V pushes that line into full-frame territory and positions itself below Canon's dedicated C80 6K full-frame cinema camera as a hybrid option for solo shooters, event videographers, wedding crews, and streamers who need cinema-style capture in a smaller package.

The design choices reflect that video focus. There is no electronic viewfinder. The shutter is electronic only. The body is flat with minimal protrusions, built around a large vari-angle rear LCD, an integrated zoom lever, an in-body tally lamp, and vertical shooting support with an optional vertical grip. The result is a camera that drops onto a gimbal or into a cage without fighting the rig.

7K Sensor, RAW Pipeline, and Internal Capture

The 32.5MP full-frame CMOS sensor reads out at roughly 6960 x 4640 for Open Gate work, with a latest-generation DIGIC-class processor handling the data. Internal recording covers 7K Open Gate up to 30p in 12-bit Cinema RAW Light and MP4, and 7K DCI (17:9) up to 60p in 12-bit RAW Light. Oversampled 4K DCI and UHD "Fine" record up to 60p from the full 7K readout without a crop, and uncropped 4K runs up to 120p for slow motion.

Canon Log 2 and Log 3 are both onboard, alongside Custom Picture controls and user LUT support. The camera includes dual base ISO in Log, around 800 and 6400 for RAW C-Log2/3, plus an HDR Movie Mode up to 4K DCI/UHD 60p, Slow & Fast Motion recording, and 4-channel audio.

For external work, the full-size HDMI Type-A output supports uncropped 7K 30p ProRes RAW or cropped ~4.3K 60p ProRes RAW with compatible Atomos recorders. Third-party coverage from CineD reports dual card slots: CFexpress Type B and SD UHS-II.

Active Cooling Built for Long Takes

The headline change for long-form shooters is the integrated cooling fan. Canon rates the camera for roughly 120 minutes or more of continuous 7K 60p RAW Light, or uncropped 4K DCI/UHD Fine 60p, with the limit being card capacity and battery rather than thermal cutoff. That puts the R6 V into territory typically reserved for boxier cinema bodies, while staying in a mirrorless form factor.

Stabilization is handled by a 5-axis in-body system rated up to 7.5 stops at center and around 7 stops at the periphery with compatible lenses, with coordinated control across IS-equipped optics. Autofocus has been tuned for video work, with smooth focus transitions and tracking aimed at handheld and gimbal use.

Stills Toolkit Keeps 40fps Burst but Loses Mechanical Shutter

Despite the video orientation, the R6 V keeps a hybrid stills toolkit. The electronic shutter runs up to roughly 40 fps with AF/AE tracking, with a buffer of around 330 JPEGs or 150 RAW frames at that speed. A pre-continuous shooting mode captures up to 20 frames in RAW, HEIF, or JPEG before the shutter is fully pressed.

The tradeoffs are real, though. Dropping the mechanical shutter means relying on the electronic shutter's rolling-shutter characteristics for fast action and flash sync, and the missing EVF takes bright outdoor stills work into rear-LCD-only territory.

RF 20-50mm Power Zoom and Accessories

Canon paired the camera with the RF 20-50mm F4 L IS USM PZ, the first full-frame L-series RF lens with a built-in Power Zoom. The lens offers a constant f/4 aperture, optical image stabilization, and motorized zoom designed for smooth ramps and remote control during video shoots.

The launch also includes a Tripod Grip HG-200TBR, the Wireless Remote Control BR-E2, and the Macro Lite Adapter Set AD-M1.

What This Means for Hybrid Shooters

At $2,499, the R6 V slots in as a lower-cost, video-optimized alternative to the EOS R6 Mark III, trading the EVF and mechanical shutter for 7K Open Gate, active cooling, rig-friendly ergonomics, and long-form capture rated past two hours. Canon already covers high-end full-frame video with the C80 in the Cinema EOS line; the R6 V offers a different bet, that solo operators and small crews want most of those capabilities in a mirrorless body they can hand-hold, mount on a gimbal, or drop into a cage without adapting a cinema chassis.

For working professionals, the practical question is workflow. The internal Cinema RAW Light pipeline, full-size HDMI for ProRes RAW out, and Canon Log 2/3 support put the R6 V into the same post pipeline as Canon's cinema cameras. With body shipping in late June 2026, the next signals to watch are real-world rolling-shutter performance on the electronic shutter, sustained-recording behavior in hot environments, and how the RF 20-50mm PZ holds up as the first full-frame Power Zoom L-series lens.

Sony announced the Alpha 7R VI, the sixth generation of its high-resolution full-frame mirrorless line. The camera moves the R series onto a 66.8MP fully-stacked back-illuminated Exmor RS sensor and adds the new BIONZ XR2 processor with a dedicated AI unit, delivering 8K 30p internal recording, uncropped 4K 120p, blackout-free 30 fps stills, and IBIS rated up to 8.5 stops at center.

Key specs:

• 66.8MP fully-stacked Exmor RS sensor with up to ~16 stops of dynamic range per Sony

• 8K 30p with 8.2K oversampling in 4:2:2 10-bit, plus full-frame 4K 60p and uncropped 4K 120p

• 30 fps blackout-free continuous shooting with full AF/AE tracking and up to 60 AF calculations per second

• Body: USD $4,499.99, shipping June 2026

A High-Res Body That Stops Trading Resolution for Speed

The previous generations of Sony's R line have traded speed for pixel count, leaving sports and action shooters reaching for the lower-resolution Alpha 1 or A9 bodies. The A7R VI closes that gap. The new fully-stacked sensor architecture reads out around 5.6x faster than the previous model, which is what makes blackout-free 30 fps shooting and 8K 30p video possible from a 66.8MP chip.

Sony's PR materials position the camera at the photography end of the hybrid market, but the video spec sheet puts it directly into competition with Canon's EOS R5 Mark II for the pro-hybrid mirrorless segment, where high-resolution stills, internal 8K, and 4K 120p have become table stakes.

BIONZ XR2 and the AI Unit

The BIONZ XR2 image processor is paired with an integrated AI processing unit driving Real-time Recognition AF. Sony's subject detection now uses skeletal-based human pose estimation, allowing the camera to track a subject across complex movement and partial occlusion. AF calculations run up to 60 times per second during continuous shooting.

The same processor backs the camera's continuous shooting buffer. Sony rates the A7R VI at up to ~215 JPEGs or 65 compressed RAW frames at 30 fps when shooting to CFexpress Type A.

Sensor-Level Dual Gain and Heat Management

The video implementation pushes past what previous A7R bodies have offered. The A7R VI records 8K 30p with 8.2K oversampling in 4:2:2 10-bit, full-frame 4K 60p, and uncropped 4K 120p for slow motion. Sony added Dual Gain Shooting at the sensor level, a dual native gain implementation that optimizes noise, shadow detail, and dynamic range across the ISO range.

Heat management is the other shift. Sony quotes up to ~120 minutes of 8K recording, putting the camera into territory that previously required external recorders or stepping up to a dedicated cinema body. Stabilization has been redesigned, with IBIS rated up to 8.5 stops at the image center and around 7 stops at the periphery, plus a Dynamic Active Mode tuned for handheld video.

Viewfinder, Body, and Power

The EVF jumps to 9.44 million dots with HDR support and wide DCI-P3 color gamut coverage, paired with a 4-axis multi-angle rear LCD. The magnesium-alloy body includes illuminated rear buttons for low-light operation and dual USB-C ports that allow simultaneous charging and data transfer.

The camera is the first Sony body to use the new NP-SA100 battery, rated at approximately 710 shots via the LCD or 600 via the EVF on CIPA testing. Storage is handled by dual card slots with CFexpress Type A and SD UHS-II support.

XLR-A4 Adapter Adds 32-Bit Float Audio

The launch is paired with the new XLR-A4 XLR adapter, which enables 32-bit float audio recording for productions that need clipping-resistant capture without external mixers. The adapter is priced at USD $779.99 and ships alongside the body in June 2026.

Sony also released a set of accessories for the body: the Vertical Grip VG-C6 ($459.99), Battery Charger BC-SAD1 ($139.99), DC Coupler DC-C2 ($149.99), and the new NP-SA100 battery ($119.99).

What This Means for Hybrid Pros

At $4,499.99, the A7R VI is priced to compete head-on with Canon's R5 Mark II in the pro hybrid market, where the buying decision usually comes down to system lock-in, lens preference, and the specifics of each body's video pipeline. Sony's pitch is the stacked-sensor readout speed, the AI unit's skeletal tracking, and a video toolkit that now reaches uncropped 4K 120p and 8K 30p without dropping the high-resolution stills capability that defines the R line.

For working professionals, the practical questions are familiar ones. Real-world rolling-shutter performance at 30 fps electronic capture, sustained 8K recording behavior in production heat, the 32-bit float audio workflow through the XLR-A4, and how the new pose-estimation AF holds up on demanding subjects will all matter more than the spec sheet once cameras ship in June 2026. The hardware is now in place for the A7R line to be a serious option for action and video work, not only landscape and studio shooters.

Google has introduced Gemini 3.5 Flash, positioning it as the first model in what the company calls its next generation of Gemini — and the performance backbone behind a faster, more autonomous Gemini app. The announcement is light on technical detail, but the intent is clear: 3.5 Flash is built for speed, and it's powering at least one of Google's new agentic experiences, Gemini Spark.

For media professionals, the story here isn't a spec sheet. It's a positioning signal about where Google is taking its AI stack — and what that means for tools already integrating Gemini into creative workflows.

What Google Announced: The official May 19 blog post frames Gemini 3.5 Flash as a model that "combines frontier intelligence with lightning-fast action" — but stops well short of a technical reveal.

There are no benchmark numbers, no latency comparisons, no context window figures. What Google did share:

• Gemini 3.5 Flash is the first model in Google's next generation of Gemini models

• It's designed to make the Gemini app feel fast and responsive — specifically for agentic, proactive behavior

• Gemini Spark, a new agent experience inside the Gemini app, runs on Gemini 3.5

That last point is the most concrete thing in the post. Spark is framed as a more autonomous layer built on top of Gemini 3.5 — the "proactive, 24/7 help" use case Google has been building toward. Beyond that, the announcement is best understood as a product-positioning statement, not a technical disclosure.

Speed as Infrastructure: For agent-style tools to feel usable, they need to feel instant — and that's the niche 3.5 Flash is designed to fill.

The Gemini app has been evolving toward more proactive behavior: taking initiative, surfacing relevant information unprompted, and handling tasks on behalf of users rather than waiting for step-by-step instructions. For that to work without friction, the underlying model has to respond quickly enough that it doesn't feel like a bottleneck.

That's especially true for media teams working fast:

• Summarizing long documents during a call rather than after it

• Drafting production emails without switching context from a meeting

• Keeping up with live planning sessions where notes need to happen in real time

Speed is infrastructure in those scenarios. A fast-action model like 3.5 Flash is what makes the difference between an AI that feels like a live collaborator and one that feels like a slow search engine.

Gemini Spark and the Always-On Layer: Google's description of Spark in this announcement is brief, but the framing matters.

Spark isn't positioned as a chatbot or a prompt interface — it's described as an agentic experience that runs on Gemini 3.5 and delivers continuous, proactive help. Think less "ask a question, get an answer" and more "a system quietly working in the background while you're focused on something else."

For production environments, that maps to a growing set of workflow needs:

• Prepping agendas and call sheets from calendar and email threads

• Maintaining continuity notes across shared docs and chat histories

• Surfacing relevant reference material while you're still in a writing or review session

The blog post doesn't go deeper than the app-level framing — no API details, no specific integrations — but the direction is consistent with how Google has been threading Gemini into productivity and creative contexts.

How This Connects to Google's Broader Creative Push: We've been tracking Google's Gemini integrations into creative tools across several recent stories, and 3.5 Flash fits a clear pattern.

In our coverage of Gemini's conversational image editing in AI Studio, the appeal was real-time responsiveness — describing an edit and seeing it applied without breaking your creative flow. In our piece on Google's AI-powered pointer, Gemini was positioned as a live collaborator that acts on the interface in response to natural language — an experience that only feels usable if latency is low. And in our coverage of the Avid–Google Cloud partnership, Avid outlined plans to embed Gemini models and Vertex AI deeper into the editorial environment.

Gemini 3.5 Flash fits the same trajectory. Google appears to be building toward a tiered model structure where:

1. Flash variants handle fast interaction, orchestration, and background task execution

2. Higher-capacity models handle deeper reasoning and heavier content generation

If that pattern holds in media tools — including through the Avid integration — a model like 3.5 Flash would likely handle "always-on" assistant roles: fetching scene notes, filtering transcripts, checking continuity metadata. The heavier creative work would route to more capable (and slower) models. The official announcement doesn't draw that line explicitly, but the positioning points that way.

What's Still Missing: The gaps in this announcement are worth naming clearly, because they matter for anyone evaluating Gemini 3.5 Flash for production use.

• No benchmarks: There's no published data on how 3.5 Flash compares to earlier Flash-tier models in latency or throughput

• No context window or token pricing: For long-form media work — scripts, transcripts, edit logs — those specs are practical necessities, not footnotes

• No media-specific examples: The blog post stays at the level of general productivity, not creative production

Until Gemini 3.5 Flash appears in Vertex AI with documented limits and pricing, its viability for large-scale media workloads — batch transcript processing, automated metadata tagging, real-time collaboration features — remains an open question.

Watching the Stack Develop: Google's Gemini 3.5 Flash announcement is a positioning move, not a product launch in the traditional sense — but the direction it signals is worth tracking for anyone building or buying into Gemini-integrated workflows.

What to watch: whether partner tools like Avid start specifying which Gemini models handle which tasks; whether Flash-tier models appear as distinct options in creative tool menus; and whether Google publishes Vertex AI details that clarify 3.5 Flash's suitability for production-scale workloads. The fast-action layer of Google's AI stack is taking shape — how it lands inside the tools filmmakers use daily will be the real measure of its relevance.

Google's new image tool treats your photo like a set — not a canvas. Announced at I/O 2026, Google Pics is built around one central idea: every element in an image is an editable object, not a locked-in collection of pixels.

That framing matters more than the name.

The Object Problem: Prompt-only editors are fast but notoriously bad at surgical revisions.

The familiar loop goes like this: you generate something close, try to fix the 20% that's off, and the whole composition drifts. A new sky means a new image. A wardrobe swap becomes a regen. The art direction you finally landed? Gone.

Google Pics, announced at I/O 2026, is Google's direct answer to that. According to Google's official I/O notes, Pics "treats every element as an individual object rather than a flat static image." The result, Google claims, is the ability to create, swap, or refine specific details with more control — without blowing up the rest of the scene.

What Pics Actually Does: It's built for editing as much as generating.

Pics runs on Nano Banana, Google's latest Gemini 2.5 Flash–based image model. According to Google's Workspace update blog, the tool is designed to help you "create just about anything with precise creative control." In practice, that means:

• Generating from a blank canvas via text prompt

• Editing existing photos — not just AI-generated ones

• Targeting specific elements to swap or refine without touching the rest of the image

That last capability is where Pics separates itself from a standard text-to-image generator. We've covered how Google was already experimenting with conversational image editing in AI Studio — chat-based edits like "move the subject closer" or "darken the background." Pics looks like the next layer: those edits now anchored to object-level understanding from the model up, not bolted on top of a flat generation.

On Set, Not On Canvas: Object-awareness changes what creative teams can actually do with an AI image tool.

For filmmakers, designers, and production marketing teams, the practical applications map directly to real workflow pain points:

• Versioning without starting over — change a product color, swap a prop, or adjust background dressing for a regional campaign without regenerating the full composition

• Protecting art direction across edits — lock in the lighting and mood while only altering a specific costume piece, sign, or set element

• Working with real photography — Pics can edit existing photos, which opens it up for unit stills cleanup, rough location concepts, or quick set extension mockups where AI-generated assets would look out of place

None of this has been demonstrated publicly yet. Google hasn't released UI screenshots or a full walkthrough, so we don't know whether the interaction model leans toward Photoshop-style selections, a conversational edit flow, or something hybrid. What's confirmed is the underlying claim: the model represents images as distinct objects, not as captions on a flat bitmap.

Inside the Stack: Pics is the image layer in a Gemini creative ecosystem that now covers video, music, and stills.

Pics fits a pattern we've been tracking. When we covered Lyria 3, Google's music generator built SynthID watermarking and production-relevant controls into the Gemini ecosystem. Veo handles video. Now Pics handles image creation and editing. All three are being threaded through Workspace, AI Studio, and the Gemini app.

It's also worth noting the connection to 3D world-building workflows we've written about. Pics isn't positioned as a 3D tool, but object-level image representation is exactly the kind of foundation that matters if you're eventually bridging 2D concept work into 3D environments — something the broader industry is actively pushing toward.

What to Watch For: The object-aware claim is the right idea; real-world consistency will be the real test.

Until trusted testers share concrete examples, the questions worth tracking are:

• Granularity — Can Pics meaningfully target "the reflection in the window" or "the logo on the mic," or does control stay coarse?

• Consistency across versions — Can it maintain a character, product, or environment across multiple variants without drift? That's critical for campaigns and episodic key art.

• Real-photo behavior — Object awareness is easier to claim on synthetic scenes. How well does it parse and edit actual photography?

• SynthID watermarking — Google has applied SynthID to Lyria 3 outputs, but the current I/O sources don't explicitly confirm whether Pics outputs carry consistent watermarking. That gap matters for broadcast and campaign pipelines.

Getting Access: Pics is live for testers now, with a broader Workspace rollout due later this summer.

Per Google's I/O 2026 announcement:

• Now: Available to trusted testers (no public application or waitlist details provided)

• Later this summer: Rolling out to Google AI Pro and Ultra subscribers in Workspace

There's no mention yet of consumer-tier access through the standard Gemini app, standalone pricing outside Workspace, or regional restrictions.

Scene Direction, Not Spell-Casting: If Google delivers on object-aware editing, Pics narrows the gap between AI generation and real production tools.

The shift from "write a prompt and hope" to "select an object and change it" is exactly what creative teams have been waiting for. Whether Pics holds up in real preproduction and marketing workflows depends on how precise and stable that object-level control actually is once it's in testers' hands. We'll be tracking closely when examples start to surface — particularly around character consistency and how the tool handles real-world photography beyond synthetic scenes.

Google's Gemini Omni is no longer just a control layer—it's now a video creation environment, with the ability to turn text, image, and video prompts into what Google calls "cinematic, high-quality" outputs directly inside the Gemini app. The update, rolling out starting May 19, 2026 to Google AI Plus, Pro, and Ultra subscribers worldwide, marks the first time Google has positioned a Gemini model as a front-end, consumer-facing video tool with explicit cinematic ambitions—not a research demo, not an API embedded in someone else's product.

For filmmakers and media pros, the important question isn't whether Google can generate video. It's where this capability lives and what it connects to—and that's where this update is worth paying attention to.

What Gemini Omni Can Do with Video: A multimodal input → video generation and conversational editing workflow, with AI avatar creation built in. What's Still Missing: Output specs, usage limits, licensing terms, and watermarking details are all absent from Google's announcement. Why It Matters: Gemini is bridging its consumer app and professional infrastructure—a pattern we've been tracking across Lyria 3, the Avid partnership, and the AI pointer work.

From the Source Material: Google says Gemini Omni now handles video through four headline capabilities:

• Multimodal prompting: Feed it text, images, or existing video and it generates video outputs. That covers text-to-video, image-to-video, and video-to-video transformations—though Google provides no details on duration, resolution, or frame rates in its blog post.

• Conversational editing: Adjust your footage in natural language. Google specifically calls out cinematic zooms, background swaps, and built-in templates as available edit types.

• Custom AI avatars: Record or upload material of yourself and Gemini Omni builds a talking avatar version of you for video outputs—one that "looks and sounds like" you, according to Google.

• Built-in templates: Pre-built formats for faster output, though the blog doesn't describe what those templates contain or whether they're editable.

Prompt to Picture: Gemini Omni's multimodal input means you're not limited to describing a shot in text—you can hand it a still from a location scout, a reference frame, or existing footage and work from there.

The three input modes—text, image, and video—open up a few practical use cases even before knowing the technical ceiling:

• Generate a rough visual from a script page or treatment

• Animate a still reference into a moving shot to communicate tone

• Transform or reframe existing material via prompt

The catch is that Google's blog gives no technical floor or ceiling: no supported resolutions, no maximum clip duration, no frame rate options. Until those details surface, it's difficult to assess whether outputs are useful beyond 1080p social content.

Talking Through the Timeline: Conversational editing is the feature with the most direct relevance to production workflows—if it works reliably.

The ability to request a cinematic zoom or background swap via natural language, rather than jumping into a timeline, has real value as a concepting and previsualization tool. Consider:

• Drafting multiple framing options from a single generated shot, without touching a node graph or timeline

• Swapping location backgrounds to pitch client options before committing to a scout or VFX pass

• Roughing in camera moves for a storyboard or animatic review

What the blog doesn't clarify is how precise this control gets. Can you target a specific frame? Specify focal length or easing on a push-in? Maintain continuity across multiple shots in a sequence? Without those answers, "conversational editing" describes a broad range of capability—from highly useful to barely sketched.

The Avatar Question: The custom AI avatar feature is the one that will land differently depending on your job title.

For solo creators and internal communications teams, the ability to scale a single person's on-camera presence—training videos, internal updates, marketing—without booking studio time has real efficiency value.

For anyone working with talent, on-camera hosts, or brand campaigns, the feature runs directly into contract, consent, and union territory. Google's blog doesn't address:

• How much source material is required to build an avatar

• Whether outputs carry any form of watermarking (visible or metadata-based, like Lyria 3's SynthID approach)

• What disclosures or consent flows exist for third-party likenesses

• What license users hold over generated avatar performances

Until those questions have answers from Google, most production and agency environments will treat this as an internal-use feature only—not something that touches broadcast or client-facing delivery.

Where Omni Fits in Google's Creative Stack: We've covered Gemini's expanding role in media workflows across several fronts—and this update connects to all of them.

Our coverage of Google's AI pointer focused on Gemini Omni as a control interface—a way to talk to your computer and direct existing software. This update moves Omni from directing tools to being the tool. That's a meaningful shift in how Google is positioning the assistant.

On the professional infrastructure side, Google Cloud's multi-year partnership with Avid brings Gemini models into Media Composer workflows via Vertex AI—firmly in the high-end post environment. Gemini Omni's new video features sit at the other end of that spectrum: consumer-tier, app-based, accessible to anyone on a Google AI subscription.

Google is building toward a stack where the same underlying model family powers both professional editing infrastructure and a chat-style creation app. This update is a step toward closing that gap at the creative layer.

Before You Build a Workflow: Several questions need answers before Gemini Omni's video tools belong in any professional pipeline:

• Output specs: What resolutions and aspect ratios are supported? Are there per-clip duration limits or monthly generation caps by subscription tier?

• Commercial rights: Are outputs licensed for broadcast, streaming, or paid campaigns? Google's blog is silent on this.

• Watermarking and disclosure: Are generated videos or avatars tagged—visibly or via metadata—the way Lyria 3's audio outputs carry SynthID?

• Collaboration: Can teams share and iterate on projects? Is there any connection to Google Drive, Workspace, or the Avid integration already in place via Google Cloud?

None of these are answered in Google's announcement. For a production, agency, or studio environment, none of them are optional.

The Long Exposure: Gemini Omni's video tools are worth tracking because of where they sit in Google's ecosystem, not just what they do on day one.

Google has shown a pattern with its professional creative tooling: Lyria 3 launched with SynthID watermarking and explicit commercial-use framing built in from the start. The Avid partnership was structured around enterprise-grade workflow integration. If Gemini Omni's video features follow the same arc—adding licensing clarity, watermarking, and professional controls over time—what starts as a consumer sandbox could evolve into a legitimate previs and concepting environment.

For now, the most honest framing is this: Gemini Omni's new video tools are a capable-sounding ideation layer with too many unanswered questions to slot into a delivery pipeline. Use it to sketch, pitch, and explore. Hold on production-grade commitments until Google fills in the technical and legal details.

Google has upgraded Flow from an AI project canvas into an agent-powered creative workspace — and the most interesting part isn't the agent itself, it's what the agent lets you build.

At I/O 2026, Google announced that Flow is rolling out a new agent to all users today, alongside a feature called Flow Tools that lets you vibe-code custom creative tools directly inside the platform. Together, they push Flow closer to something filmmakers and media teams can actually build a pre-production workflow around — not just a chat window for early ideation.

The big caveat: pipeline integration details are still sparse, and the gap between "powerful demo" and "production-ready" remains wide.

The Agent in the Room: Flow's new agent is built to do more than answer questions about your project.

According to Google's I/O 2026 announcement, the agent is designed to:

• Plan and reason through complex, multi-step tasks — not just autocomplete single prompts

• Maintain a "deep understanding of a project" — meaning it should track context across scripts, references, notes, and creative decisions, rather than treating each session as a blank slate

• Operate "under user control" — Google is explicit that the agent doesn't act autonomously; it moves when you direct it

For production teams, "under user control" matters. It signals explicit approvals and checkpoints rather than an AI silently revising your script or recutting a timeline. That's the right call for any professional workflow, even if it limits how much heavy lifting the agent can do on its own.

The agent is powered by Gemini Omni, Google's newest model, which also brings natural-language video editing to Flow — letting you stack edits conversationally rather than jumping between settings panels.

Flow Tools: Vibe Coding for One-Off Creative Needs: The more practically interesting announcement is Flow Tools, which lets you build project-specific utilities inside Flow using plain language.

Google's examples include tools for:

• designing video effects

• hand-drawn animations

• layering text

The concept is natural-language-driven tool creation: you describe what you need, and the agent assembles something usable without you writing code. Think of it as spinning up a micro-tool purpose-built for a single project or sequence — a rough animatic generator calibrated to your storyboard style, an overlay tool mocked up for your show's lower-third treatment, a filter preset for testing a specific visual tone before committing it to your main pipeline.

According to coverage from PCMag's I/O live blog, Flow now supports multiple actions and complex edits alongside these vibe-coded tools, which suggests the editing layer and the tool-building layer are designed to work together.

For working creatives, the appeal is speed-to-prototype:

• Generate and test a look before touching After Effects or Nuke

• Build a text-layering template with your show's brand rules baked in, without wrangling software you don't have access to

• Iterate on motion language or animation style quickly, then hand off to whoever implements it at production level

What's not yet clear is how these tools surface (are they code cells, custom panels, templates?), whether they're shareable across teams and reusable across projects, and whether there are export paths into Premiere, Resolve, or any other NLE. That's the gap that will determine whether Flow Tools stay in the concept lab or start showing up in real pipelines.

Where Flow Sits in Google's Wider Media Push: Google has been building agentic AI into professional media tools on two parallel tracks — and Flow is the one Google owns outright.

We've covered how Avid Media Composer's Google Cloud partnership brings Gemini-powered semantic search, highlights creation, and agentic metadata work directly into editors' existing workflows. And we've covered how Google's AI reached into ComfyUI and cloud media pipelines at NAB, plugging models into node-based frameworks studios are already running.

Both of those strategies are about meeting professionals where they are — inside Avid, inside ComfyUI, inside cloud infrastructure they're already paying for.

Flow is a different bet: Google building its own creative environment and making it smart enough to earn a seat at the table alongside those tools. The Flow agent is what Google believes can justify that position. Where Avid + Google handles augmenting traditional editorial, Flow is pitched as the place to incubate new projects, looks, and workflows from scratch — now with an agent that can help architect those workflows as you go.

Whether a standalone AI studio can coexist with established DCCs and NLEs in a real production — or ends up siloed as a development and pitching environment — is still an open question.

The Two-Type Agent Test: If you've been tracking the AI agent space in creative tools, most implementations fall into two categories:

1. Task agents — contained, single-function actions like auto-tagging shots, cleaning up dialogue, or generating B-roll for a specific beat. Most of what Google's Avid integration demos show lives here.

2. Orchestration agents — agents that plan, delegate, and coordinate across a longer project arc. Flow's new agent is pitched in this direction: reasoning about what comes next, maintaining project context, and helping you define and build the tools you'll use.

The honest answer is that orchestration agents are much harder to get right — especially when projects get messy and real, with conflicting drafts, evolving briefs, and multiple collaborators pulling in different directions. Until there's hands-on user footage and case studies from actual productions, Google's claims here are intent, not proof.

Getting Into the Frame: Flow's agent update is worth paying attention to — particularly if you're building out a pre-production or visual development process and want a workspace that can hold the full shape of a project over time, not just respond to isolated prompts.

The near-term use case is clearest as a concept and development lab: a place to structure creative briefs, test visual directions, rapidly prototype tools for a specific campaign or series, and think through production planning before things get locked. If Google closes the integration loop — surfacing Flow-built tools in ways that talk to Avid, Adobe, Unreal, or cloud asset systems — it becomes something more central. For now, Flow is evolving from AI-assisted brainstorming toward agent-assisted creative production, and this I/O update is a meaningful step in that direction. The next step is seeing it hold up under the weight of a real project.

Google's Gemini app just got a persistent background layer — Gemini Spark turns the assistant from a chat interface into a cloud-based agent that keeps running tasks while you're away from your device. For production teams already embedded in Gmail, Docs, and Slides, the implications go well beyond personal productivity.

Google announced Gemini Spark on May 19 as part of what it's calling the "next evolution" of the Gemini app. Trusted tester rollout begins this week, with a U.S. beta for Google AI Ultra subscribers following next week.

What Spark Actually Is: The distinction matters here. Spark isn't a new product or a separate app — it's the agentic layer added to the existing Gemini app, running on Gemini 3.5 and Google's Antigravity harness, which is Google's internal orchestration framework for managing multi-step, multi-app agent behavior.

What makes Spark different from a standard Gemini prompt is persistence. It's designed to:

• Keep running in the cloud when your laptop is closed or your phone is locked

• Monitor and act across Workspace apps — Gmail, Docs, and Slides, plus "connected apps" (Google hasn't enumerated which ones)

• Chain actions across services — for example, read incoming emails → synthesize notes into a Doc → draft a follow-up email

• Ask for confirmation before high-stakes actions like spending money or sending emails on your behalf

Google's examples in the announcement are all personal-use oriented: parsing recurring credit card statements, tracking school emails, turning meeting notes into documents. The underlying mechanics, however, map closely to how production teams already use Google's tools.

What This Could Mean for Production Workflows: Sticking strictly to what Google has actually described, several applications translate directly to film and TV workflows.

The inbox monitoring capability is the most immediately useful. Many producers and coordinators already route project communications through Gmail. A Spark instruction like "watch for any emails with 'call sheet' in the subject, summarize daily changes into a Doc, and draft an update to the crew" follows the same logic as Google's school-email example — it's the same mechanic applied to production logistics.

The notes-to-document-to-email flow has a clear production analog as well:

• A raw transcript from a table read or mix review becomes a structured Google Doc

• From that Doc, Spark drafts follow-up emails to department heads

• That's a multi-app chain Spark is explicitly designed to handle

On the financial side, Google's recurring statement parsing lines up naturally with vendor invoice tracking and budget reconciliation — the kind of administrative work that eats coordinator hours on every production.

One important constraint: Google has said nothing about multi-user or shared workspace support. For now, Spark appears to be single-account assistance, which puts it squarely in the "personal productivity" lane rather than a team-level production coordinator.

Where Spark Fits in Google's Broader Push: This announcement doesn't arrive in isolation. Google has been building out multiple layers of Gemini-powered agents across different parts of the production stack.

We've covered Google's AI pointer project, which routes Gemini through motion, speech, and shorthand controls at the OS level. And we've covered Google Cloud's partnership with Avid, where Gemini models handle style matching, B-roll suggestions, and emotional cue detection inside Media Composer.

Spark occupies a different layer: it sits near the communications and planning infrastructure surrounding a project, not the media assets themselves. There's no indication in this announcement that Spark can read Drive files, understand media assets, or coordinate with tools like Avid, Adobe, or DaVinci Resolve. Google's integration list here is Gmail, Docs, and Slides — full stop.

The more interesting question, which nothing in this announcement answers, is whether these layers eventually connect. Google is building:

• UI-level agents (the pointer)

• Media-specific agents (Avid/Content Core via Vertex AI)

• Personal cloud agents (Spark, tied to communications and documents)

Whether those layers will eventually share context around a production pipeline is an open question. For now, think of Spark as the assistant for your production paperwork, not a collaborator in your edit.

The Trust and Control Gap: Agentic systems that operate while you're offline require clear permission boundaries — and the announcement leaves several important ones unaddressed.

Google confirms Spark will pause and ask before sending email or spending money. What Google has not detailed:

• Granularity of controls — Can you restrict Spark to draft-only mode, or limit it to specific senders?

• Auditability — Is there a log of what Spark did in your account overnight?

• Cross-account behavior — Can it operate in shared production company inboxes, or is it strictly personal?

For anyone handling contracts, rights clearances, or budgets, those questions matter more than convenience. The safe approach for early adopters is to test Spark in non-critical, non-financial contexts until the permission model is clearly documented.

Access and Availability: Rollout is staged and U.S.-only at launch:

• Trusted testers: This week

• Google AI Ultra beta (U.S. only): Next week

• Everyone else: Watch and wait

There's also no mention of APIs or admin controls for organizational deployment, which means studios and post houses looking to evaluate Spark at the team level don't have a clear path in yet.

The Long View: Google is methodically expanding where Gemini operates — from prompt responses, to interface-level control, to media workflows, and now to persistent background agents running in your cloud account. Spark is the latest extension of that pattern.

For creative professionals, the practical ceiling on Spark today is administrative assistance: smarter inbox monitoring, automated document synthesis, and structured follow-up drafts. That's genuinely useful for producers and coordinators buried in project communications. Whether it grows into something that meaningfully touches the production pipeline depends on integrations and enterprise controls that Google hasn't shipped yet.

Keep an eye on whether Drive access, media awareness, and team-level management land in future updates. When those arrive, the calculus changes significantly.

The latest Denoised covers the AI tools and stories shaping mid-2026 production: an open-source pipeline that builds full 3D scenes from a single image, Netflix's new in-house GenAI animation studio, a pair of DeepMind drops ahead of Google I/O, and the latest round of AI marketing backlash.

We dig into Image Blaster's Claude Code-driven 3D world generator, break down Netflix's INKubator job listings, run through Google's mouse-pointer demo and the Fabula story tool, and unpack why a real Monet painting got dragged for looking AI-generated.

Quick Take

Tooling is getting more specific while reception is getting more complicated. Image Blaster shows what a single developer can ship with Claude Code, World Labs, and fal.ai. Netflix is spinning up an entire animation studio for GenAI workflows. Google's DeepMind is pitching an AI mouse pointer and a story-structure sidekick. Meanwhile, a Monet painting labeled "AI" gets called slop, and Sony's Xperia camera launch melts in real time. The tools are arriving faster than the audience's willingness to accept them.

What We Explored: DeepMind's AI Mouse Pointer and Fabula Story Tool

Ahead of Google I/O, DeepMind dropped two demos that point at where Gemini is heading. The first is an AI-aware mouse pointer. We previously covered the demo, which pairs voice with cursor gestures so Gemini can act on whatever the user is pointing at. Point at a to-do list, say "add emojis," and the system runs the edit. Point at an address, ask for directions, and Maps loads with your location plugged in.

Why this matters for production tools: The bottleneck on agentic workflows in apps like DaVinci Resolve or Blender has been visual context. Claude Code can drive Resolve, but it cannot reliably see what it is doing, so it flies blind on anything beyond simple node graphs. A pointer-plus-voice combo is a step toward the two-way feedback loop those workflows need.

Addy's take: The same interaction model would unlock complex software like Blender for non-experts. Point at a vertex, say what you want, let Gemini handle the operator path.

The second drop is Fabula, a story-writing sidekick. We previously covered its research preview, developed in collaboration with playwrights and screenwriters and grounded in story-structure references. Fabula does not write stories for you. It builds the scaffolding: beats, character structure, hierarchy. Then you fill it in.

The fair critique, which we echo from Ben Affleck's comments on the Joe Rogan show, is that any tool trained on conventional story structure will pull toward the median. Asymmetrical work like Memento or Pulp Fiction does not come out of beat sheets. For YouTube verticals, episodic content, and any pipeline where volume matters more than originality, Fabula maps cleanly. For breaking the form, it is just the AI version of the screenwriting handbook.

What We Broke Down: Image Blaster's Single-Image 3D Pipeline

This was the segment that "blew my mind," in Addy's words. A developer named Nicholas Neilson released Image Blaster, an open-source toolkit that takes one input image and outputs a full navigable 3D scene with models, lighting, physics, and sound effects. The orchestration runs through Claude Code, with World Labs handling the 360-degree environment generation and fal.ai APIs hitting individual mesh and image generation models.

What ships with it:

• A companion app for viewing and posing the generated objects

• Built-in physics, so generated objects can be tossed around the scene and bounce off other geometry

• Sound effect generation tied to object interactions

• Open-source code, downloadable from the project repo

The practical angle: This is previs work for solo operators. Build the world in seconds, block the camera, then feed clean 3D-rendered passes into a video model for final pixel. The thesis we keep coming back to is that mid-2026, text-to-video alone does not give you the control real production needs. Traditional 3D control plus generative finishing is the workflow that actually clears the bar.

The bigger signal: A single developer shipping this with Claude Code, World Labs, and fal.ai is the more interesting story than any individual model release. The composability is what matters.

What We Debated: The Monet Test and Sony Xperia's AI Camera Mess

Two stories from the same week capture how audiences react to anything labeled AI, even when the label is wrong.

An X user (@dioscuri) posted a cropped image online and claimed it was AI-generated in the style of Monet, then asked commenters to detail what made it inferior to a real Monet. The replies poured in: harsh edges, no soft color blending, no symbiosis between elements, no sense of space, "borked nonsense." The image was a crop from an actual Monet. We previously covered the underlying research showing that viewers systematically downgrade aesthetic assessments of work labeled AI, regardless of what it actually is.

Sony Xperia walked into the same buzzsaw on purpose. The phone account launched its new AI Camera Assistant with a marketing post showing before-and-after photos where the "after" images looked dramatically worse: blown-out highlights, crushed shadows, the kind of overprocessed look that defines AI slop. The internet had a field day. The next day, Sony posted a clarification. We covered the walk-back: the feature does not actually edit photos after capture. It suggests four creative direction settings before the shutter fires. Auto-exposure with scene-aware suggestions, in other words. The clarification makes the feature defensible. The launch post had already tanked the perception.

The pattern across both stories: The AI label is now load-bearing on its own. The Monet test shows it negatively biases evaluation even on objectively skilled work. The Xperia launch shows that even useful features can read as slop when the marketing assets are wrong. For anyone shipping AI-assisted tools to creators, the lesson is that demos and copy matter as much as the tech.

The shift to watch: We see 2026 as a vibe shift, where most skeptics have moved from outright dismissal to wanting to understand what these tools actually do. The window for marketing AI features with carelessness is closing.

What We Questioned: Netflix Spins Up INKubator as an In-House GenAI Studio

The bigger institutional move is Netflix's. We previously covered INKubator, Netflix's new internal animation studio built around GenAI-native production pipelines. Most of what is publicly known has been pieced together from job listings: long-term technology strategy focused on "GenAI enabled workflows, artist tooling, and scalable, secure multi-show environments," with initial output aimed at animated shorts and specials using experimental GenAI-native pipelines.

What is significant about this: This is not a department inside an existing team. This reads as a full studio spin-up. Netflix already runs in-house production on a small percentage of its slate (Blue Eye Samurai is the example we keep returning to, with a reported $100 million-plus animation budget). INKubator looks structured to figure out the production methodology before pushing anything into the main streaming queue.

Why animation first: Animation is the easier proving ground for GenAI today. Live action carries more thorny problems (talent likeness, on-set capture, IP). But animation has its own AI-resistant challenges: squash and stretch, mouth exaggeration on dialogue, the kind of expressive performance that goes beyond puppeteering. Video-driven character control gets you a generic puppeted performance. Real animation needs training data and control surfaces that can exaggerate motion in service of a story beat.

The cost question driving this: If a series in the Blue Eye Samurai cost range can be made with a meaningful percentage of GenAI shots, the studio math changes. INKubator is where Netflix figures out that ratio.

What We Closed With: Disney's Muppet Mocap Animatronic

A small story that we couldn't not flag: Disney's rebrand of the Aerosmith Rock 'n' Roller Coaster into a Muppets attraction includes a new Scooter audio-animatronic, and the programming workflow is genuinely interesting. We covered the breakdown showing that Imagineering motion-captured a real Scooter puppet, performed by a human puppeteer, then mapped that mocap data onto the animatronic. The puppet wore a mocap suit and facial markers tracked the mouth and expressions.

The result is a permanent robot that carries the hand-driven performance characteristics of the original puppet. It is a clean example of using capture tech to preserve craft, rather than replace it.

Bottom Line: Specific Tools Are Winning, Marketing Is Still Hard

The episode's connecting thread is that specific, composable tools and methodologies are showing up faster than the audience and marketing layers can keep up.

• Image Blaster shows what one developer can ship by composing Claude Code, World Labs, and fal.ai. The composability is the story.

• DeepMind's Mouse Pointer and Fabula point at where Google is heading with two-way visual context and scaffolded creative tools. Useful for high-volume content; less useful for breaking the form.

• The Monet test and Xperia's backlash show that the AI label is doing heavy lifting on its own. Viewers downgrade work labeled AI; bad marketing assets can sink useful features.

• Netflix's INKubator is the institutional bet that GenAI animation pipelines are ready for a dedicated studio, starting with shorts and specials.

• Disney's Muppet mocap is the counter-example: capture tech in service of preserving a craft performance, not replacing it.

The pattern: the tools are getting more useful and more specific, but trust and reception are still the harder problems to solve.

Links from This Episode

Tools & Platforms:

• Image Blaster: Open-source single-image 3D pipeline using Claude Code, World Labs, and fal.ai

• Fabula: DeepMind's Gemini-powered story-structure assistant (research preview)

• Gemini AI Pointer: Google DeepMind's voice-plus-cursor interaction demo

Projects & Studios:

• Netflix INKubator: Netflix's new GenAI-native animation studio

• Disney Scooter Animatronic: Mocap-driven animatronic for the Muppets coaster rebrand

News & Analysis:

• Monet AI Bias Test: Viewers downgrade work when told it is AI, even for a real Monet

• Sony Xperia Walk-Back: Sony clarifies the AI Camera Assistant after launch-post backlash

Chinese short-video platform Kuaishou is preparing to spin off its Kling AI video generation business into a standalone company at a target valuation near $20 billion, according to The Information. The unit is in pre-IPO talks to raise roughly $2 billion, with Tencent Holdings among the reported potential investors, and an independent Hong Kong listing eyed as soon as 2027.

Three numbers anchor the deal for production teams:

• Kling AI's annual recurring revenue reached $500 million ARR as of April 2026, up from roughly $240M ARR in December 2025

• J.P. Morgan analysts project $1.3 billion ARR by Q1 2027

• A $20 billion valuation would equal nearly 70% of parent company Kuaishou's total market capitalization, per Pandaily's reporting

The Revenue Curve: Kling's monthly revenue more than doubled inside a single quarter.

The unit pulled in approximately $20 million in revenue during December 2025, equivalent to a $240M ARR run rate. That figure climbed past $300M ARR by January 2026 and reached the $500M mark by the end of April, according to BigGo Finance's breakdown. Kuaishou launched Kling in 2024 and has grown it into a direct competitor to Google's Veo and ByteDance's video tools. The growth curve is the clearest commercial benchmark yet for the AI video category, where most Western competitors do not disclose revenue.

Spinoff Mechanics: A standalone Kling changes how the product is funded, priced, and prioritized.

Kuaishou confirmed its board is evaluating a restructuring plan for Kling's assets and business, while cautioning the proposal remains preliminary with no agreements signed. A $2 billion raise at a $20 billion valuation implies roughly 10% dilution, leaving Kuaishou with a controlling stake while bringing in capital earmarked for compute, model training, and Western market expansion. The Hong Kong-listed parent stock surged on the reports, per the South China Morning Post. A 2027 IPO timeline gives the new entity roughly 18 to 24 months to build out enterprise infrastructure before facing public market scrutiny on growth metrics.

What Changes for Production Teams: An independent Kling will likely accelerate work on features that matter to working studios.

The current Kling 3.0 model generates 1080p clips at flexible durations up to 15 seconds, with multi-shot storyboarding supporting up to six camera cuts inside a single generation and native audio with lip sync, based on Morphic's product breakdown. These capabilities are aimed at narrative and commercial work, not social content. As a parent-company side project, Kling has competed for engineering resources with Kuaishou's core short-video app. Standalone status means dedicated roadmap control and a sharper enterprise sales motion. Production teams already running Kling through API access or third-party integrations should expect:

• Faster iteration on generation length, resolution, and audio sync as engineering headcount expands

• Pricing changes once the unit needs to justify its valuation to outside investors

• A more aggressive push into US and European enterprise accounts, where Veo and Runway currently lead

The Hong Kong Bet: A 2027 listing puts Kling on a public reporting timeline that reshapes competitive transparency.

If the IPO closes on the indicated schedule, Kling will be among the first AI video pure-plays to disclose audited financials at scale. Runway, Pika, and other Western competitors operate without public revenue obligations, so a public Kling will set the benchmark the rest of the market gets measured against. The financials also settle a longer industry debate: $500 million ARR climbing toward $1.3 billion suggests at least one company has converted AI video into durable enterprise spend, not just viral demos. Production teams evaluating long-term vendor commitments should watch how the round closes with strategic investors like Tencent: who leads it, what board seats come with it, and whether Kuaishou retains pricing or licensing veto. Each of those terms will shape what enterprise customers can negotiate over the next two years.

Two AI tools aimed at independent filmmakers launched at the Cannes Marché du Film from Chris Bird, the former managing director of Prime Video UK who stepped down after more than 14 years at Amazon.

The launches:

• CineMe AI generates photo-realistic visual storyboards from a script in seconds, co-founded with director Dan Hartley

• HawksHead AI uses proprietary databases to forecast how a project will perform with specific audience segments at the script or synopsis stage

• A charitable trust called the CineMe Future Fund holds 5% of CineMe's equity and provides enterprise-grade AI access to screen industry workers

Both products target capabilities that were once confined to major studios and streaming platforms. Bird's pitch is that the same analytics and visual-development infrastructure used by streamers should reach independent producers at the point where it matters: before a frame is shot.

Storyboards in Seconds: CineMe AI converts a screenplay into photo-realistic frames that producers, directors, production designers, cinematographers, costume designers, and VFX teams can review and iterate on before principal photography.

The platform sits in the visual-development gap between script and pre-vis. Hartley, who directed HBO and Sky's documentary David Holmes: The Boy Who Lived, put the case for it directly:

"Before CineMe you used to have to wait until you'd made a film before you could see it, now you don't."

CineMe is in beta with what the founders describe as confidential, high-profile productions. Generative AI VFX capabilities for explosions and large-scale set pieces are planned as the platform scales, though not yet available.

Audience Forecasting From a Synopsis: HawksHead AI analyses scripts or synopses against proprietary AI tools and private databases to predict performance with specific audience segments, then suggests adjustments to scripts and casting that could improve resonance.

A feature the company calls a "synthetic panel" lets creators test changes and receive feedback within hours rather than weeks. The pitch is aimed at independent producers who need to win commissions or secure broadcaster and streamer investment.

Bird framed the broader thesis in a Screen Daily interview:

"The old ways of commissioning, based on who you knew and what talent you could attach, are changing rapidly. Streamers and platforms have for years used sophisticated data analysis to make investment decisions, and that data was their exclusive preserve. Now, through HawksHead AI and CineMe, we're levelling the playing field, putting that same power into the hands of British filmmakers and content creators at the point where it matters most: before a single frame is shot."

Five Percent to a Charitable Trust: CineMe assigned 5% of company equity to the CineMe Future Fund, a charitable trust designed to provide enterprise-grade AI access to screen industry workers whose livelihoods have been disrupted by shifts in the production landscape.

The structure is unusual for an AI startup at launch. Most generative AI companies serving production charge subscription fees and offer free or discounted tiers for hobbyists. Routing equity to a trust that subsidizes access for working professionals is a different model, and one that distinguishes CineMe from peers competing on price and feature parity.

What's Worth Watching: Outside of confidential beta partnerships, details remain limited on training data, sample outputs, and pricing for either tool. The fidelity of CineMe's storyboards and the predictive accuracy of HawksHead's forecasts will only be measurable once beta projects ship.

For production teams, the practical evaluation points are whether CineMe's outputs reach the quality bar to replace concept art or scouting passes, and whether HawksHead's forecasts hold up against actual commission outcomes. Both will need open testing before the "levelling the playing field" framing becomes more than a launch pitch.

This Denoised covers the NAB tools that did not make our initial coverage and the wave of Claude integrations that landed across creative software. We walk through Lightcraft Spark's browser-based virtual camera, Selects by Cutback's agentic editing pipe, Martini's angle-repositioning feature, and the Anthropic donation that triggered a Blender backlash. We close on Kling's jump to 4K and the Academy's new AI rule for the Oscars.

Quick Take

A lot of NAB news happened off our initial recap. Lightcraft Spark turns the Vcam workflow into a browser tab. Selects by Cutback builds an MCP-based highway between an NLE and Claude. Martini lets you reposition a 3D camera against a 2D AI frame. Then Claude rolls into Adobe and Blender, the Blender community pushes back on Anthropic's enterprise donation, Kling ships native 4K, and the Academy formalizes its position on AI nominees. Different tools, one thread: how does a working filmmaker actually use any of this on a job?

What We Tested: Lightcraft Spark Story

Joey caught a prototype demo of Spark Story from Lightcraft, the team behind the Jetset iPhone camera tracker. Spark is a web platform where you scan a location as a Gaussian splat, drop the splat into the browser, and drive a virtual camera using an iPhone or the in-browser camera. Think Unreal Engine Vcam without Unreal Engine.

What it does in the browser:

• Imports Gaussian splat scans of real locations

• Connects an iPhone or virtual device as a camera input

• Records camera moves with a take recorder

• Brings in animated 3D objects you can trigger during a take

• Exports for previs, storyboarding, or as a base layer for AI video pipelines

Joey's take: "There's nothing to install or set up. It just syncs." On the NAB Wi-Fi the camera response felt real time.

Addy's add: "I've set up Vcam with Unreal before. It's gotten easier and easier, but all the take setup and file naming and the professional stuff around it, I think this offers as an add-on on top of that platform."

The pitch lands hardest for AI workflows: previs a scene in Spark, then style-transfer the recorded camera move through Kling or Seedance to get an AI-generated shot with real camera control. Lightcraft said Spark is targeting a launch later this summer. We covered the full Spark platform announcement when Lightcraft first laid out the pre-to-post production vision.

The connection question: What if a Vcam from Spark could drive the 3D camera in Martini's angle editor? More on that below.

What We Explored: Selects by Cutback and Agentic Editing

Selects is positioned as AI-assisted editing for podcasts, talking-head YouTube, and vlog footage. The surface-level pitch is auto-cut, auto B-roll. The interesting part sits underneath: Selects ships with Claude integration via MCP servers, so an LLM can sort, analyze, and assemble cuts across multiple Selects projects.

The demo Joey saw: A production company hits the per-project footage cap, so the MCP layer lets Claude reach into prior Selects projects, treat them as a searchable footage library, and assemble new cuts pulling clips across them.

Why this matters: Selects is one of the NAB tools building the highway between an NLE and an agentic LLM, rather than only adding AI features to its own UI. As Addy put it, "Once the highway has been established between an NLE to agentic LLM, it's only a matter of time before the LLM gets better and smarter and is able to make better editing decisions within your NLE."

Where it still needs a human:

• Creative editing with voiceover and narrative B-roll

• Anything beyond predictable formats like multicam podcasts

• The "craft of editing" once the first cut exists

The right read on Selects is not "AI editor" but "agent-ready footage layer." For broader context on how Model Context Protocol is connecting Claude to creative apps, we previously broke down MCP and the demos showing Claude controlling tools like Blender and DaVinci Resolve.

What We Tested: Martini's Angle Editor

Martini is a canvas-based AI filmmaking platform that runs all the major image and video models inside one workspace, with a built-in timeline that exports to your NLE. The unique feature is the angle editor: feed it a frame, the system analyzes the people in it, Gaussian splats the scene, and gives you a 3D space to reposition the camera in.

What stood out:

• Canvas, not nodes for laying out shots

• Multi-user collaboration in the same workspace

• Built-in timeline for rough edits and NLE export

• Angle editor that turns a single 2D frame into a navigable 3D scene

• Camera control beyond text prompts including over-the-shoulder repositions

Joey's note: "Some of the best camera control that I've seen."

We covered Martini's launch when the Y Combinator-backed team came out of the W26 batch with a "Figma for generative film" pitch. The angle editor is the piece that turns it from another model aggregator into something cinematographers can actually direct.

The combo question: Addy floated it directly. "Imagine tying a Vcam like a Lightcraft Spark to this system. In real time you have some level of interactivity. Now you're really getting somewhere with cinematography." Either Lightcraft adds 3D scene reconstruction or Martini adds physical camera input. Whichever side closes the gap first owns a workflow no single tool currently delivers.

What We Debated: Claude in Adobe and Blender

Anthropic spent the gap between our episodes announcing creative-tool integrations. Two are worth tracking.

Claude in Adobe Creative Cloud. Claude can now talk to Photoshop, Premiere, and the rest of the Creative Cloud stack. Addy's read on why this happened fast: "Adobe Creative Suite has a very robust API for third-party plugins. The connections already exist." The current limitation is one-way. Joey tested similar Claude integrations in DaVinci Resolve and the model cannot see what it is doing inside the project. As Addy framed it: "It's more of like a one-way street where Claude is just instructing the DCC, but it's not really feeding back." Useful for tedious batch work and project setup. Not yet useful for iterative real-time editing.

Claude Meets Blender. Anthropic announced a Blender Connector plus an enterprise-tier donation reportedly worth around €240,000 annually. The connector lets Claude build scenes, gray-box environments, and manage assets inside Blender via prompts.

Then the backlash hit. The Blender community pushed back hard enough that Blender said it would accept the donation once but not as a recurring annual contribution.

Joey's argument: The €240,000 represented roughly the cost of four full-time developers. Blender is free and open source because of donations like this. Anthropic was not asking for code changes, only sponsoring development.

Addy's read: "This is a big setback for Blender not taking that money, in my opinion."

The Blender use case for AI workflows:

• Build gray-boxed 3D scenes as motion base layers

• Feed those layers into Kling or Seedance for style transfer

• Hand off tedious tasks like importing assets and managing the outliner

• Stay relevant in a world where some 3D pipelines collapse into pure AI generation

• The pattern with both Adobe and Blender is the same. Claude is becoming a connective layer across creative software. The technical work is there. The community acceptance is uneven.

What We Questioned: Kling 4K and the Oscars Rule

Kling AI now supports native 4K output for its Video 3.0 series. Most other video models still rely on an upscale pass. Native 4K matters even if the final delivery is HD, because down-resing a 4K render gets you sharper details on skin, hair, and edges. The open question is wide shots, where AI video still struggles with pixel density. As Joey put it: "I'd be curious how this 4K works with wide crowd shots. Wide shots where those are still the trickiest with AI. There's not enough pixels, the details fall apart." We covered the Kling 3.0 launch including the native 4K capability and built-in audio generation.

The Academy formalized its AI rule. Only roles "demonstrably performed by humans with their consent" are eligible for Oscar nominations. Joey's read: this is a clarification, not a sweeping ban. AI as a tool inside a workflow remains fine. AI as the credited author or performer is out. Addy's broader frame: "AI is a tool for the creatives, and the creative is going to combine AI with traditional tools to make the thing that they make."

Joey's running joke that has a point: "The more important question is, will AI qualify for the Razzies?"

Bottom Line: The Connective Tissue Year

The throughline across this episode is how tools connect to each other. The standalone AI model race is leveling off. The next race is integration.

• Lightcraft Spark moves Vcam workflows out of Unreal and into the browser, with a clear handoff to AI video pipelines.

• Selects by Cutback builds an MCP highway between an NLE and Claude, betting on the agent-editor pattern.

• Martini's angle editor adds 3D camera control on top of 2D AI frames, opening a slot for physical camera input from tools like Spark.

• Claude in Adobe and Blender turns a chat model into a one-way controller for Photoshop, Premiere, and 3D scenes. Community acceptance is the bottleneck, not capability.

• Kling 4K raises the floor on AI video resolution, with wide shots as the remaining stress test.

• The Academy rule clarifies that AI in a workflow is fine, AI as a credited author or performer is not.

The combo that did not exist at NAB but should exist soon: a browser Vcam feeding a 3D-aware AI canvas, with an LLM running cleanup and assembly downstream. Spark plus Martini plus a Claude-aware NLE is the workflow worth building toward.

Links from This Episode

Tools & Platforms:

• Lightcraft Spark: VP Land coverage of the browser-based virtual camera and previs platform

• Martini: VP Land coverage of the canvas-based AI filmmaking workspace

• Selects by Cutback: AI-assisted editing with Claude and MCP integration

• Shade: cloud media asset manager with built-in AI search

News & Analysis:

• Adobe Creative Cloud integrates with Claude

• Claude launches Blender Connector

• Blender changes stance on AI donations after Anthropic backlash

• Kling AI launches 4K mode for Video 3.0

• Kling 3.0 native 4K coverage: VP Land breakdown

• OpenAI ChatGPT Images 2.0

• Academy bans AI from winning Oscars

• Model Context Protocol explained: VP Land context on Claude's creative-tool connectors

Channels:

• Theoretico: AI YouTube channel running historical-period POV vlogs

A virtual production studio that doubles as an AI workflow shop is pitching brands on the idea that a shoot should leave behind a reusable digital asset, not a torn-down set. We sat down with Door G Creative Director Justin Poirier and Virtual Production Technical Director Joe Ross at NAB 2026 to talk through the studio's AI workflows, custom Unreal Engine pipelines for clients like CVS, and why healthcare brands are landing on the LED volume.

Key takeaways:

• AI runs through pre and post, not on the wall. Door G uses generative tools for image previz, transcript analysis, and footage tagging while keeping final volume content out of the generative pipeline.

• A two-day on-location shoot collapsed into six hours when Door G reused pre-built environments on the volume.

• Custom Unreal blueprints built with Claude Code let the studio rebuild and re-dress client sets like CVS without starting from scratch.

• Healthcare is a growing vertical because virtual production sidesteps HIPAA, hospital access, and competitor IP exposure.

Where AI Fits, and Where It Doesn't: Door G keeps generative tools off the LED wall and runs them through the pipeline instead.

Poirier described Door G as a virtual production studio trying to build what he calls the "studio operating system of the future" by applying AI upstream and downstream while keeping the human element intact. In pre-production, the team runs creative briefs and client transcripts through AI to surface what a client actually said and generate image options that visualize a concept. In post, the same approach pulls quotes, finds themes, and analyzes footage. What they're not doing is using generative AI for final on-wall content.

"From what we've seen so far, it just isn't there yet," Poirier said, while leaving open the possibility for still backgrounds down the line.

Poirier said pulling a still from the Unreal Marketplace and using AI to animate it gives clients a fast read on how a video game engine asset will look in motion, which gives brands new to the format confidence it will work. We previously covered Door G's VP playbook for brands, including how the studio combines LED volumes, 2D plates, and still photography.

Why Brands Are Booking the Volume: Sustainability, scheduling, and a CVS store that doesn't get built and torn down.

Door G's pitch to brands often comes down to logistics. Poirier said one shoot that would have taken two days on location collapsed into six hours on the wall because the environments were already built in Unreal Engine. Sustainability is another driver: moving a crew between locations gets replaced by uploading a new environment.

The CVS work is the clearest example. Door G built a custom CVS store as a virtual asset after seeing brands repeatedly build physical sets for one-off shoots.

"They were going up to Canada and building CVS stores where there are no CVS stores," Poirier said.

Retail clients commonly travel to Europe to build American-style environments for tax reasons, then tear them down. A reusable digital store flips that math by turning the build into an owned asset.

Building Tools at the Speed of Production: Ross is using Claude Code to turn one-off Unreal builds into repeatable pipelines.

Ross handles the technical end at Door G and uses Claude Code to write custom Unreal blueprints that fit a specific client's needs, which expedites set construction on repeat engagements. The CVS build is the case study: the first version was a major undertaking, and since then Door G has built blueprints that let the team construct shelves, stock them with assets, swap products, and "messy them up a little bit" so the shelves don't look too perfect.

Poirier wants to extend the approach to his shot lists. He builds detailed shot lists for every shoot so production can continue if he disappears, and he's interested in using Claude Code to automate the repetitive parameters across a single video, with the caveat that too much automation costs him the visualization step.

"Striking that balance of automating it enough so that I can do it quickly, but keeping the human element in there enough so that I can still see everything," Poirier said.

Show Floor and What's Next: Avid's Gemini integration, consumer cameras edging into broadcast, and healthcare as a quietly growing vertical.

Ross flagged Avid's integration of Google's Gemini as an agentic tool for metadata tagging and generative frames; the studio has years of archived footage that's been hard to search. Poirier added that consumer-grade gear keeps edging into broadcast quality, pointing to Insta360, small Blackmagic cameras around $1,000, and GoPro adding a Micro Four Thirds mount.

Poirier expects more brand adoption because the format produces more content per shoot day; swapping backgrounds lets clients walk away with multiple looks from one session. Ross added that once a brand has a digital twin of its store or product, the asset travels across stages and campaigns.

Healthcare is the growth area Poirier called out. Door G has produced virtual production work for Samsung, Baxter, and Aetna. The case is compliance-driven: HIPAA restrictions, limited hospital room access, and the need to keep new products from competitors all push healthcare brands toward controlled, virtual environments rather than working on-site.

Axle AI rolled out its fourth-generation tagging engine at NAB, expanding semantic search across video, transcripts, and human-applied tags in a single index. The MAM and DAM platform also debuted a native Avid Media Composer panel and is now letting customers extend the product themselves using Claude. AI processing runs on-premise, a deliberate counter to the cloud-default tooling that dominates elsewhere in the category.

Bogoch also referenced the company's Portfolio DAM acquisition from Extensis, a stills-focused product line with customers including universities, libraries, and Lockheed Martin. The deal effectively doubles Axle AI's business and gives Portfolio users access to the same on-prem AI and workflow automation stack that powers the video product.

Fourth-Generation Axle Tags. The current version of Axle Tags, the company's tagging and search engine, expands semantic search from the video content itself to a combined index of video, transcripts, and human-applied tags. According to Bogoch, the engine "searches the whole enchilada" so a query hits captions, speech-to-text, and metadata in one pass.

• On-prem by default. Bogoch estimates 90% of AI tooling discussed in the industry runs on cloud infrastructure, which he said creates privacy, security, and cost issues that most of Axle AI's customers will not accept. He pointed to Amazon Rekognition's terms of service, which permit retaining uploaded content for training, as the kind of fine print that "a lot of studios would freak out" about.

• Petabyte-scale archives. Bogoch said 4K and 8K acquisition is pushing more customers into petabyte territory, where cloud egress and storage costs make on-prem hardware the more practical option.

• Hybrid storage reality. Post-COVID, customers have returned to a balanced model: storage and editors co-located in the office, with remote freelancers accessing assets through hybrid setups. We previously covered Axle AI's expansion to Mac mini and Mac Studio for on-prem processing.

Avid Media Composer Panel. New at the show is a Media Composer panel that lets editors search the Axle AI catalog and drag clips directly into the timeline, matching the integrations Axle AI already ships for Premiere Pro and DaVinci Resolve. Axle AI previously supported Avid OP-Atom MXFs at the file level, but the panel makes the workflow native inside Media Composer. We covered the Media Composer panel launch in greater depth.

According to Bogoch, the integration reflects what Avid customers are asking for: open systems that talk to Avid without locking them into a closed pipeline.

Agentic Workflows and Claude-Built Customization. Bogoch described agentic AI for video as being at "1.0" today, with most teams brainstorming rather than running production work through agents. Axle AI's existing Connector workflow tool can talk to agentic engines, and Bogoch expects the space to move quickly over the next six months.

The bigger structural shift, he said, is how customers extend the product themselves using Claude. Because Axle AI's stack is modular and exposes REST APIs, users can build the last 10% of functionality they need without waiting on the vendor roadmap. Internal teams are working the same way.

• ONNX model support. Customers can train and plug in their own machine learning models for category-specific recognition (Bogoch cited airline logos, flora and fauna, and undersea content as examples) when the default AI models don't cover their use cases.

• Customer-built integrations. A Portfolio customer built a WordPress publishing interface for pushing subsets of stills directly from the DAM, demoed on a recent Axle AI webinar. According to Bogoch, "a few days of vibe coding" can now replace a quarter of vendor roadmap waiting.

• Coders as product managers. Bogoch said Axle AI's engineers are shifting toward managing "Claude minions" rather than writing every line of code themselves, compressing what used to be a year of features into roughly two weeks.

Portfolio DAM and Workflow Automation. Portfolio handles stills at scale (millions of files), with on-prem deployment positioned as a way to control storage costs as file sizes climb. Now part of Axle AI, Portfolio users get the same Connector workflow engine that handles automation rules on the video side.

According to a Portfolio representative on the interview, common automated workflows include:

• Consent capture for biometrics. Triggered by the Biometric Information Privacy Act, a Connector flow can look up a subject's email from an identified face, send a consent form, and gate downstream facial recognition processing on the response.

• Metadata validation routing. Files arriving without complete metadata, or with AI-applied tags that need a human-in-the-loop check, can be flagged and routed to a reviewer before publishing. We covered Axle AI's Searchr portable media system for similar workflow patterns on the video side.

• Rule-based ingest. Incoming stills can be routed to specific Portfolio locations based on metadata values, keeping the interface simple while the automation runs underneath.

Facial recognition and object recognition models from the Axle AI side now run against stills inside Portfolio (transcription excluded, since stills don't have audio), supplementing human-applied metadata.

At NAB, Imaginario AI launched StoryLab, a feature that takes a natural-language creative brief and returns a multi-video rough cut by routing the job through three specialized AI agents. The platform sits on top of multimodal indexing that breaks footage down to the shot and scene level across vision, speech, and sound, then puts that index to work on automated curation tasks.

Co-founder and CEO Jose Puga positioned Imaginario in the editing tool stack: not raw-interview cleanup, but curation of finished content into highlights, summaries, and compilations.

From System of Record to System of Action. Puga framed the product shift in those terms. Imaginario started by improving video search accuracy, indexing content across visual shots, dialogue, music, sound effects, and ambient sound. According to Puga, the indexing combines those modalities into a unified understanding down to the scene level. With that foundation in place, the company is now layering curation features on top of the index, including highlights, clips, rough cuts, and insights.

The platform stays storage-agnostic. Imaginario supports Google Drive, Dropbox, Box, Frame.io, Backblaze, Wasabi, and AWS S3, with more integrations planned. Puga said the goal is to avoid forcing teams to subscribe to another MAM system or re-upload assets.

The Three Agents Behind StoryLab. A user types a creative brief in natural language, selects source assets, sets a duration, and picks moods or shot types. That activates a chain of three agents:

• Archive researcher. Searches the indexed footage for relevant visual shots, dialogue, and sound design, building a pool of candidate material against the brief.

• Copywriter. Takes the candidate pool and writes a story arc with an introduction, main actions, and resolution, organizing scenes into a coherent narrative.

• Assistant editor. Adjusts the start and end points of each selected scene, functioning as what Puga called a junior editor handling timing optimization.

In a demo using two Lord of the Rings films, StoryLab processed a 10-minute summary brief in 42 seconds. The output preview lets editors trim handles, reorder scenes, or swap selections before exporting.

Editing Suite and Distribution Integrations. StoryLab is targeted at finished content rather than raw interview footage, with summaries, highlights, and compilations in the 2-to-10-minute range as the current sweet spot. A separate Single Clip feature handles shorter cuts from finished pieces, automatically tracking speakers and adding subtitles for direct export to YouTube Shorts, TikTok, or Instagram Reels.

The platform currently integrates with Premiere and Resolve, with Avid support planned for Q3 2026. Puga said CapCut and Canva integrations are also on the roadmap as Imaginario expands beyond post-production into content marketing teams.

On-Prem and Enterprise Controls. Imaginario runs as a cloud app on AWS infrastructure (S3 storage, EC2 GPU and CPU instances) but Puga said an on-prem version is targeted for end of Q3 2026. The on-prem build is aimed at clients in industries sensitive to data training, including teams that want to buy their own hardware and process content locally. The intelligence layer may stay hybrid, with LLM calls from providers like Claude or OpenAI handling reasoning while indexing runs on-site.

On agent autonomy and MCP-style access, Puga was direct about enterprise constraints. He said open-source agent stacks are not ready for primetime in regulated industries like healthcare and financial services, citing cybersecurity assessments and the need for deterministic behavior. Imaginario's approach is to treat MCP as another context channel feeding its existing agent logic rather than handing control to a general-purpose orchestrator.

Vibe Coding the Roadmap. Puga, who also serves as Imaginario's chief product officer, said he prototypes weekly using the Imaginario API plus open-source models to test new verticals before committing engineering time. He cited a hypothetical CCTV use case (license plate or vehicle recognition) as the type of feature he can build a working demo for in days, then validate with prospects before greenlighting production work. According to Puga, the approach has saved months of misallocated engineering across the team.

Further out, the StoryLab roadmap adds a general researcher agent that ingests production notes, scripts, brand guidelines, and franchise context before the archive researcher runs. Puga also flagged plans for analytics-driven clipping, where agents pull YouTube performance data and trend signals from a target demographic to inform what gets cut and how. The 9:16 reformatting, caption templates, transitions, and music layers already exist elsewhere in the product and are slated to roll into StoryLab once the rough cut layer is stable.

Blackmagic Design walked us through the AI tool set in DaVinci Resolve 21 at NAB 2026, and the depth of the controls inside each plugin matters more than the headline count. AI CineFocus reconstructs a depth map from a flat shot and gives colorists a virtual cinema lens to refocus, change aperture shape, and animate a rack pull after the fact. IntelliSearch turns the media pool into something closer to a queryable database.

The AI plugins live in the Color page as Resolve Effects under the Refine section. Most build on tools introduced in Resolve 20, which we covered in our AI editing workflow breakdown.

Focus Pull, Reframed: AI CineFocus is the most camera-aware of the new tools, replicating a full optical lens rather than a simple defocus blur.

• Internal depth map. The plugin generates the depth map from the source clip on its own. If you already have a rendered depth pass, you can feed it into the alpha input instead.

• Virtual aperture. Controls cover aperture shape (spherical or anamorphic), number of blades, rotation, and chromatic aberration. The bokeh shape responds to those iris settings rather than being a generic blur kernel.

• Track Focus to Point. Enabling this mode exposes keyframeable X and Y controls. Pair it with Resolve's object tracker and the focal point follows a subject across moving footage without manual keyframes.

• No Foreground Blur toggle. Locks a foreground region as always-sharp so a focus pull behind a subject doesn't soften them by accident.

• Map level controls. Near and far limits, gamma, and post-process softening let colorists tune which parts of the depth map fall off and how aggressively.

The Blackmagic rep demonstrated the tool on a car shot where the original cinematographer pulled focus to keep both the car and the city sharp. In Resolve 21, the choice of which plane stays in focus becomes a node-level decision in post.

Faces, Edited in Place: Three plugins target on-camera talent and share the same face detection and tracking pipeline as the existing Face Refinement tool.

• Face Age Transformer. You enter the subject's actual age, then keyframe an age offset over the clip. The demo took 20 years off a 40-year-old subject between two keyframes. The tool requires a roughly front-facing angle and loses control past a three-quarter view.

• Face Reshape. Adjusts face shape, eye size, nose, mouth, and eyebrow position on a tracked face. The demo animated a mouth shape between keyframes to shift a performance from cranky to happier. Eyes, nose, and mouth controls are combined by default, with an option to split them.

• Blemish Removal. A single strength slider with a narrow detection focus. Pores and fine hair survive; spots and blotches get smoothed. Combine it with a Power Window or Magic Mask selection to target a region instead of the whole face.

Sharpening, Deblurring, and Up-Res: Two optical correction plugins handle different fixes that get conflated in marketing copy.

• Ultra Sharpen. Rebuilds edge detail and recovers fine texture, demonstrated on a bird where feather definition came back after the pass. Controls are a sharpened amount slider and a global blend opacity. Works well on slightly soft-focus interviews.

• Motion Deblur. Generates a new cleaned source clip rather than acting as a filter, then sits on a layer above the original for toggle comparison. Demoed on a moving sign with heavy motion blur.

• Different from Super Scale. Super Scale is the upscaling algorithm that resamples a clip to a new resolution. Ultra Sharpen works on the existing frame to recover edge detail that was obscured. Run a denoise node first to cache cleaner inputs for the rest of the pipeline.

IntelliSearch and Auto-Generated Smart Bins: The AI clip analysis pass is the workflow change colorists and assistants will feel most.

• Slate detection. Resolve drops duration markers wherever it sees a slate and extracts the data printed on it.

• Transcription with speaker detection. Runs alongside audio classification during the analysis pass.

• Persistent face identity. Detected faces get auto-generated smart bins. Rename a bin to a person's name and Resolve tags future media with the same face. The demo found "Sean Villon" across multiple shoots from a single search.

• Search to smart bin. An IntelliSearch query can be saved as a smart bin that keeps updating as new media enters the project. The face-to-speaker-transcript link is not wired up yet; the Blackmagic rep noted it as a feature request.

Editors who have been bolting on third-party search tools like Jumper for clip discovery now have a native option inside the same project that holds their timeline.

Availability: DaVinci Resolve 21 is in public beta and free to download from Blackmagic Design. Studio license holders get the AI plugins as part of their existing license. The Photo page ships in both the free and Studio versions, with Studio adding ResolveFX and Inspector tools on top of the shared photo grading interface.

A World Labs engineer has released an open-source toolkit called image-blaster that converts a single still image into a complete 3D environment, including meshes, physics, lighting, and audio, with Claude Code orchestrating the pipeline.

The project comes from Nicholas Neilson, a member of the World Labs team. Image-blaster is a personal experiment, not an official World Labs release.

Quick read:

• Single-image input: Feed the toolkit one image and it generates an explorable 3D scene

• Composable stack: Marble handles 3D generation, Claude Code drives the workflow, and fal supplies compute

• Side project, not product: A solo experiment from a World Labs employee, with no formal support or roadmap

Twitter tweet

The Stack: Image-blaster wires three AI services into one workflow.

Marble, World Labs' image-to-3D model, builds the geometry and environment from the source image. Claude Code, Anthropic's terminal-based coding agent, runs as the orchestration layer, coordinating calls between the other services and assembling the output. Fal handles the compute and model inference. The result is a scene with rendered meshes, physics-ready geometry, baked lighting, and generated audio.

Why Claude Code is the Glue: Neilson built image-blaster as a Claude Code toolkit, not a standalone app, putting the agent itself at the center of the creative loop.

A developer or technical artist runs Claude Code in their terminal, points it at an image, and the agent chains the API calls, handles errors, and writes the assembly script that produces the final scene. The toolkit form factor means anyone with Claude Code installed can fork the project and modify the pipeline.

This pattern reflects a wider trend in which AI coding agents are treated as a runtime for creative tools, not just IDE assistants.

The World Labs Connection: Marble is World Labs' generative 3D system, and Neilson's day job places him close to what the model can do.

The startup, founded by Fei-Fei Li, has focused on what it calls "world models," AI systems that produce navigable 3D scenes rather than flat images or video. We covered the company's web-streamable 3D Gaussian Splatting capability as another milestone in that direction. Image-blaster is not endorsed by World Labs. It is a side project from a team member exploring what the underlying model can do when paired with an agentic coding layer. Production teams evaluating Marble for pipelines should weigh the official product, not this experiment, against their needs.

Status check: The GitHub repository at github.com/nicholasneilson/image-blaster returns a 404 and may have been renamed, moved private, or taken down. Anyone tracking the project should watch Neilson's X account for updates.

One Image, Many Scenes: Image-blaster sits at the intersection of two trends: image-to-3D generation maturing into usable pipelines, and AI coding agents taking on roles beyond software development.

For VFX artists, previs supervisors, and indie creators experimenting with rapid scene building, the toolkit signals what composable AI workflows can look like outside the major DCC apps. Whether image-blaster itself sticks around or not, the pattern of agent-orchestrated, model-agnostic creative tools is worth tracking.

Goodnight, Lamby, a live-action and animation hybrid short film produced by Darren Aronofsky's Primordial Soup and built in part with Google DeepMind's generative AI tools, will premiere at the 79th Cannes Film Festival in the Cannes Classics section. The short marks the directorial debut of Brooklyn artist Dustin Yellin and features the voices of Paul Rudd, Chris Rock, and Yellin's daughter, Zia Copernicus Yellin.

• Selected for Cannes Classics as one of two contemporary works in the 2026 lineup, screening alongside restored prints and documentaries

• Produced by Darren Aronofsky and Justin A. Gonçalves under Primordial Soup, with Villavicencio Studio and Google DeepMind attached as production companies

• Live-action and animation hybrid built around Yellin's three-dimensional "frozen cinema" sculpture work, with an original song from Maggie Rogers

The short follows three-year-old Zia as she searches for her stuffed Lamby through a layered, surreal sculpture built by her father.

Rudd voices "Papa" and Rock voices "Copernipus" in a dreamscape that moves between oceans, caves, and outer space, according to Deadline. The piece grew out of Yellin's signature glass-and-collage sculptures, which he calls "frozen cinema," and turns those static objects into a moving narrative environment. Zia Copernicus Yellin, the director's daughter, voices the title role.

Executive producer Nick St. Pierre, who announced the project on X, framed the short as a live-action and AI film. Yellin and Aronofsky will both attend the Cannes Classics screening.

This is the second short to surface from the Aronofsky-led AI studio in twelve months and its first selection at Cannes.

Primordial Soup launched its partnership with Google DeepMind in May 2025, announcing three short films built around DeepMind's Veo video model and related generative tools. The first, Eliza McNitt's Ancestra, premiered at the Tribeca Festival in June 2025 and drew critical pushback over the training data and labor questions baked into generative video pipelines.

Goodnight, Lamby is the partnership's first project to land at a major European festival. Cannes Classics, traditionally a home for restorations and cinema-history documentaries, is folding two contemporary works into its 2026 lineup, per Cineuropa. That placement signals festival appetite for hybrid AI work even as the underlying tooling remains contested.

Yellin's gallery describes the short as a translation of his sculptural practice into film.

The Brooklyn artist's Almine Rech announcement frames Goodnight, Lamby as an extension of his "frozen cinema" practice, with the layered glass sculptures functioning as the film's primary set. The production stack pairs that physical work with generative video tooling from Google DeepMind, name voice cast, and a Grammy-nominated musician contributing original music. The credits include Primordial Soup, Villavicencio Studio, and Google DeepMind as production companies.

Cannes Classics is opening to AI-assisted narrative shorts alongside its archival programming.

For VP Land readers building hybrid pipelines, Goodnight, Lamby is a useful data point on what gets through the gate at a top-tier festival: a known producer, a credentialed first-time director with an existing fine-art practice, name voice talent, and AI tooling that functions as a component of the work rather than the marketing pitch. The 79th Cannes Film Festival runs May 12-23, 2026, and the reception from critics and trade press will set the tone for how the rest of Primordial Soup's three-film slate is received.