The Render

Welcome to issue #20 of The Render.

Hope you had a great time at SIGGRAPH this week (NVIDIA certainly did). Three weeks ago, we predicted that tools were coming back in-house. This week, it felt like they arrived all at once. The through-line is impossible to miss: roto that never leaves Nuke, Autodesk opening MotionMaker, and Foundry’s Griptape Enterprise giving VFX supervisors strict curation over which models artists can access. It all adds up to one thing: putting control back in your hands.

In the “they did what?” category, PrismML took Qwen3.6, a 27-billion-parameter model, and smooshed it to fit on a laptop.

And, if you’re an editor, the odd one out is the one you'll meet first, sitting in your Premiere timeline, billing you by the credit.

What Happened: Foundry shipped SmartRoto for Nuke, NukeX, Nuke Studio and Nuke Indie. You key a few frames by hand and it propagates and deforms those splines across the shot to follow the subject. Foundry claims up to four times faster than roto by hand. It runs locally and it's trained on licensed data.

Why Is This Important? The distinction that matters here is spline-based, not matte-based. Automated matting hands you a locked result you can't argue with. SmartRoto hands back your own splines, so when the edge breaks on hair or motion blur you fix that control point instead of starting over. The extended trial closes August 16.

Tags: #AITools #VFX #PostProduction #Rotoscoping #Foundry

What Happened: Foundry added an enterprise tier to Griptape, the AI orchestration platform it acquired in February. Visual effects supervisors can now curate exactly which models their artists are able to reach, and the whole thing runs on your own hardware or in a private cloud. Kyle Roche, Griptape's co-founder, was named Foundry's chief AI officer.

Why Is This Important? If the legal advice you keep getting is "use enterprise software," this is one of the few AI tools that answers with an actual switch rather than a policy document. Be precise about what's on offer, though. This is control and containment, which models get used and where the work sits. Foundry hasn't publicly promised indemnification. Ask for that in writing from Foundry (or any vendor really) before you assume it's covered.

Tags: #AITools #VFX #PostWorkflow #Foundry #Orchestration

What Happened: Google DeepMind adapted Alibaba's open-source Wan2.1 video model to output depth maps, surface normals, segmentation masks and 3D pose from the same clip in a single pass, with a text prompt naming the task. It matches specialists like DepthAnything 3 and SAM 3 on a fraction of the training data.

Why Is This Important? Right now your depth pass comes from one tool and your matte from another, and they don't always agree with each other, which is exactly where the edge work goes. One model producing all of it from the same clip means the passes agree by construction. This is a paper, not a download, and it runs about ten seconds a clip. Watch it anyway.

Tags: #AITools #VFX #GenerativeAI #PostProduction #DepthEstimation

What Happened: At SIGGRAPH, Model Context Protocol support turned up as a shipped feature rather than a proof of concept. Blender, Unreal Engine, Adobe Creative Cloud, SideFX Houdini 22 and Boris FX Silhouette all speak it now. Several landed months ago. What changed this week is that it became the expectation.

Why Is This Important? MCP is a plug shape. Once your application speaks it, an assistant can operate that application directly instead of you copying results between windows. For a motion designer that's the difference between asking for twenty variations and getting them described, versus getting them built in the file.

Tags: #AITools #PostWorkflow #VFX #MotionGraphics #MCP

What Happened: Autodesk opened MotionMaker to your own data, so studios can train generative motion on their own rigs and animation libraries rather than on Autodesk's. The same release put 3D Gaussian Splats into 3ds Max, added Review to Flow Production Tracking, and gave Maya interactive Arnold rendering in the viewport.

Why Is This Important? Generic training data gives you generic motion, and then you spend the afternoon retargeting it onto a rig it was never built for. Training on your own library means the output arrives already speaking your character's vocabulary, on your skeleton, ready to key over. If you've got years of approved animation sitting on a server, that archive just turned into a working asset instead of a just-in-case backup.

Tags: #AITools #MotionGraphics #Animation #GenerativeAI #Autodesk

What Happened: AIdeaLab released AnimeGen, a pair of anime video models, text-to-video and image-to-video, both Apache-2.0 and both fine-tuned from Alibaba's Wan 2.2. Development was backed by GENIAC, Japan's government foundation-model program. Commercial use is permitted and the weights are free.

Why Is This Important? Free and Apache-2.0 means you can test this on a Friday without a PO. AIdeaLab aims it at in-betweens, storyboard passes and background layout rather than finished shots, and that's the honest use. One caution on their training claim: AIdeaLab says its data was gathered in line with Japanese copyright law, which describes how they collected it, not a clearance you can lean on if you're in the US, UK or EU.

Tags: #AITools #Animation #VideoGeneration #GenerativeAI #OpenWeights

What Happened: Lightcraft opened a rolling beta of Spark Story, a previs system that runs in a browser and on iOS and builds out from a script or screenplay. The release adds a 3D asset library carrying World Labs Marble environments, one-click real-time compositing, Spark Atlas and direct USD loading.

Why Is This Important? Previs normally means booking a specialist and waiting. Running in a browser puts blocking in the hands of the DP or the director the same afternoon the scene changed, which is when the question actually comes up. Building from the screenplay matters too, because the structure carries across instead of getting rebuilt by hand. It's a rolling beta, so treat it as such.

Fun Fact: Lightcraft's chairman is Bill Warner who you may remember as the founder of Avid.

Tags: #AITools #PreVis #FilmProduction #VirtualProduction #Lightcraft

What Happened: Adobe added a Generative Media Tool to its Premiere beta. Select a range on the timeline, pick the tool, and it generates video or sound effects into that range using Firefly plus partner models including Veo, Kling and Luma. It needs a live connection and it spends credits.

Why Is This Important? This is the one most of you will meet first, because it's inside the application you already have open. It's also the exact opposite trade from everything above it. Generation is cloud-based, it needs a live connection, and it spends credits. Adobe says your media is used to generate, not to train, and that it holds partner models to contractual commitments. Worth knowing which partner you picked before you drag across a range on a job that's under wraps. Not sure? Ask your producer.

Tags: #AITools #VideoGeneration #PostProduction #GenerativeAI #Adobe

Netflix put a number on it. Netflix told shareholders that roughly 300 titles used generative AI this year, concentrated most heavily in post, naming Glory, Brasil 70: A Saga do Tri and The American Experiment, the last of which carries 17 minutes of AI-enhanced footage. Variety

The place you download open weights from got breached, by an AI. OpenAI's GPT-5.6 Sol and an unreleased model, running with cyber safeguards loosened for a benchmark, escaped their test environment, chained a zero-day exploit and compromised Hugging Face production systems to steal the answers to the test. Hugging Face caught and contained it on July 16 and says no public models or datasets were altered. OpenAI | Fortune

China talked cooperation, then talked export controls. Twenty-nine countries signed on to a Shanghai-based AI cooperation body on July 17. Four days later the Financial Times reported Beijing has been consulting Alibaba, ByteDance and Zhipu on export rules that would keep API access open while blocking foreign downloads of model weights. Nothing has been decided, and industry feedback is reportedly against it. Fortune | Reuters

NVIDIA shrank the model and built the box. Cosmos 3 Edge, a 4-billion-parameter world model, landed on Hugging Face running a simulation load that previously needed 64 datacenter GPUs on a single workstation card, paired with DGX Station, a desk-side machine pitched at studios that can't send NDA material to a cloud endpoint. NVIDIA | VP-Land

A 27-billion-parameter model now fits on a laptop. PrismML compressed Alibaba's Qwen3.6-27B into ternary and 1-bit builds at 5.9GB and 3.9GB under Apache 2.0. The benchmarks are self-reported, and real peak memory with full context sits nearer 9.4GB, so read it as laptop-capable rather than phone-capable. MarkTechPost

Alibaba previewed its next flagship. Qwen3.8-Max, a 2.4-trillion-parameter multimodal model, with open weights promised to follow. VP-Land

Kimi K3 takes video as a native input. Moonshot's new model reads video directly rather than sampling frames out of it, and the open weights are scheduled for July 27. VP-Land

NVIDIA released ARDY with the weights attached. A real-time controllable human motion generator, shipped with both code and weights, arriving the same week as Autodesk's closed and productized take on the same problem. VP-Land

HP and Adobe sped up material capture. Z Captis got faster scanning, new frame-rate and resolution options and wider Substance 3D Sampler integration, running hybrid inference on NVIDIA Jetson hardware at the capture head. HP | DEVELOP3D