NVIDIA puts AI agents inside your creative tools

In partnership with

Welcome to VP Land! Open source keeps reaching deeper into production, from calibrating LED stages to modeling a digital human's face.

Two for the calendar: PIXERA's HDR and Digital Color in Realtime Video Workflows runs this Friday in Santa Monica, with a free YouTube livestream. And for SIGGRAPH, we have the AI Workflows Summit, a free event this evening. I'll be moderating a panel on hybrid film production workflows with AI.

In today's edition:

ASC open-sources an LED-volume test kit
NVIDIA puts local AI agents across creative tools
Alibaba previews a 2.4T-parameter model
Google frees a scan-based head model

The ASC Is Open-Sourcing a Virtual Production Test Kit for LED Volumes

The American Society of Cinematographers is open-sourcing StEM3-VP, a set of reference assets for evaluating and calibrating LED volumes, and the material is joining the Academy Software Foundation's Digital Production Example Library.

Built for the LED-to-camera seam. The assets test the image path between the wall and the camera, where color, moiré, and brightness problems surface on a volume. A production can characterize and sign off a stage before principal photography, using open media instead of vendor-specific test patterns.
Openly licensed, no vendor lock. StEM3-VP sits in DPEL alongside other production-grade test content, giving cinematographers, DITs, and stage engineers a fixed target that does not belong to a single wall or processor manufacturer.
A two-decade lineage. It follows the ASC's original StEM from 2004 and StEM2 in 2022, extending that evaluation approach from digital projection and color pipelines to the LED stage.
Surfaced through ASWF's open program. We covered the Open Source Days schedule where StEM3-VP was slated to present, on a track spanning color-fidelity standards, OpenPBR, and studio-safe AI workflows.

SPONSOR MESSAGE

The best prompt engineers aren't typing. They're talking.

Power users figured this out early: speaking a prompt gives you 10x more context in half the time. You include the edge cases, the examples, the tone you want — because talking is fast enough that you don't skip them.

Wispr Flow captures everything you say and turns it into clean, structured text for any AI tool. Speak messy. Get polished input. Paste into ChatGPT, Claude, Cursor, or wherever you work.

89% of messages sent with zero edits. 4x faster than typing. Works system-wide on Mac, Windows, and iPhone.

Start flowing free

NVIDIA Launches Local AI Agents and More

NVIDIA used SIGGRAPH to move AI agents directly into the tools artists already use, with a cluster of releases aimed squarely at creative and production work.

Local agents in Blender. An open Agent Toolkit lets teams run AI agents inside Blender on their own hardware, keeping unreleased assets, scripts, and client material on-site instead of routing them through a third-party API.
Agents across more creative apps. NVIDIA lined up a wave of creative software behind MCP, the open standard that lets an agent read and act on an application's live state, extending the same automation well beyond a single tool.
A synthetic-video detector for newsrooms. A new Synthetic Video Detector microservice flags whether a clip contains synthetic content, a sign that provenance is becoming part of the same creative stack.

Alibaba Previews Qwen3.8-Max, a 2.4-Trillion-Parameter Multimodal Model With Open Weights to Follow

Alibaba's Qwen team opened preview access to Qwen3.8-Max, a 2.4-trillion-parameter multimodal model it calls its most capable yet, and the latest sign that China's open-weight labs are shipping frontier-class systems right behind Kimi and Fable.

In paid preview now. It runs through Alibaba's Token Plan, Qoder, and QoderWork at 10 percent of standard pricing and takes images, video, and documents as input, with open weights promised at a later date.
The ranking is a vendor claim. Qwen says it trails only Fable 5 among frontier systems, but it has published no benchmarks and has not disclosed the active-parameter count or its mixture-of-experts setup, so 2.4 trillion is total size, not per-query compute.

Google Open-Sources GNM Head, a Parametric 3D Head Model With 250+ Identity Controls

Google open-sourced GNM Head, a scan-based parametric 3D head model, under a commercial-friendly Apache 2.0 license, moving a studio-grade building block out of the research lab and into production pipelines. The model exposes 636 controls spanning identity and expression, down to fine detail in the eyes, teeth, and tongue, so you can dial in a specific face or drive a performance without hand-sculpting every blendshape. A free Blender 5.1+ importer ships alongside it, turning those controls into live sliders in the viewport with no plugin fees or research-only strings attached. For anyone rigging digital humans, populating background crowds, or building avatar systems, it is a rare genuinely production-ready open asset: commercially licensed, usable today in a tool most artists already run, and detailed enough to hold up in close-up work.

Neill Blomkamp’s NIGHTBORNE: a 13-minute Seedance 2.0 AI-film test

Neill Blomkamp’s 13-minute sci-fi horror film is the first release from Barley Studios, set in Peter Watts’ Echopraxia universe. Its production uses Seedance 2.0 to generate every frame, with Blomkamp directing shot by shot through prompts.

The “fully AI” label needs context: the project also uses real concept artists and licensed faces and voices from 32 human performers. That makes it a useful watch for the emerging hybrid model, where generative imagery carries the shots while artists and performers still supply core creative inputs.

Stories, projects, and links that caught our attention from around the web:

▶️ Alibaba's Wan-Streamer runs real-time, full-duplex video conversation at about 550ms and can embody any character.

🎬 AI video startup Invideo is making a feature film and soliciting scripts at [email protected].

👩🏻‍💻 A free, open-source DaVinci Resolve MCP hands an AI agent 37 live tools across color, sound, and timeline assembly.

We Watched Codex Rebuild a Real Theater in Blender From Phone Photos

In Denoised, we watch Codex reconstruct a real theater in Blender from phone photos in about 10 minutes. It is a useful demonstration of how quickly a filmmaker can turn casual reference images into a workable starting point for planning shots, blocking, or testing an idea in 3D.

The workflow is not a replacement for measured scans, clean production geometry, or an experienced artist’s judgment. But for early previs and location-driven exploration, it suggests a faster path from reference to scene.

Read the show notes or watch the full episode.

Watch/Listen & Subscribe

Spotify | Apple Podcast | YouTube