Skip to content

← The Lab

Claude Code Video Toolkit makes 3D character films in Blender from one JSON file; plus Higgsfield's Claude-built Blender cities, Mesh2Motion Release 14 and HyperFrames 0.9.73

Digital Samba's open-source Claude Code Video Toolkit (MIT, about 2.7k stars) can now put a 3D character in your video. One JSON file holds the cast, the set, the shots and every move, timed to the narration, and Blender renders it with no window open. The same day, Higgsfield showed Blender cities Claude built and rendered on its GPU Harness, and Mesh2Motion shipped a big free rigging and animation update.

What the toolkit is. You tell Claude Code what video you want, and the toolkit handles the script, the voiceover, music, images, clips and a Remotion render, mostly on open models (Qwen3-TTS, FLUX.2, ACE-Step) deployed to your own cloud GPU account. Per the README, its own 90-second explainer cost $0.33 in compute. /setup configures the GPU provider, storage and voice, and /video starts a project from a template. The hello-world example renders an MP4 with no API keys.

The new 3D part (v0.28.0 Nov 12 to v0.29.1 Nov 14, explained in a Nov 20 video). You write shots by type (wide, medium, close-up) and the acting as verbs like walk, sit_down, carry and sip, with no keyframes or coordinates. Every beat lands on a word of the narration, so when the script changes, the acting follows. A check command builds the scene in about five seconds and fails on problems like a hand inside the body, a limb through a table, a sliding foot or a mug that tips, before any frame renders. Output can be a whole film with sound, one clip per shot or transparent clips that play inside a Remotion video. For your own character, a page walks you through eight webcam reference photos (they stay on your machine), an image editor makes a character sheet, an image-to-3D service makes the model, and fit seats it on the shared rig with blinking eyes and fingers. Today's merge (PR #257) improves fitted faces and hands in close shots and adds docs for 'painted world' looks.

Requirements and honesty notes, per the project. Blender 4.2 or later is optional, and nothing else in the toolkit needs it. With a GPU, a frame takes a second or two, so a 27-second example is about 25 minutes. Run uv run tools/blender_render.py doctor first to time a frame on your machine. Fitted characters, the kitchen and the lecture hall are labelled experimental, it's been run in full on macOS, its checks pass on Linux and Windows is untested. The author pitches it as the repeatable partner to a live Blender MCP session: try a look live, then keep it in the spec so the same file gives the same film next month or in CI.

Higgsfield's Blender cities (Nov 20). A day after launching GPU Harness, Higgsfield posted a Shibuya Crossing it says Claude built in Blender, with 9,662 people, about 2,200 vehicles and 4,261 buildings from the real street map (11:22 AM MST, about 250 likes). Later it posted a Miami-style town of 2,538 assets rendered in Unreal Engine on GPU Harness in 74 minutes, plus a domino version of The Great Wave. These are vendor showcases tied to a Black Friday GPU sale (up to 63% off, per Higgsfield), so read them as what's possible, not typical times.

Mesh2Motion Release 14 (notes Nov 19; @scottpetrovic Nov 20, 6:18 AM MST, about 250 likes). Scott Petrovic's free, open-source take on Mixamo: import a GLB/GLTF model, fit a skeleton and export animations. This release adds props that export with your character (poles, knives, swords, a shovel), manual weight-painting cleanup, remixing two animations into a new one, 50+ new animations, Mixamo-compatible animation import and a fix so exports face the right way in Unity. Code is MIT and the art assets are CC0, so it's a cheap way to get rigged, moving characters into Blender, Unreal or a code-driven video pipeline.

HyperFrames 0.9.73 (Nov 20, 4:15 PM MST). Preview copies of long clips with sound no longer show black footage, saving history to a drive that stops responding no longer freezes the app, audio automation no longer crashes the preview renderer, and remote fonts embed instead of warning that they're missing. Renders now refuse more than 100,000 copied elements and cap remote downloads at 1,000 per compile.

Signals: project video and release notes, vendor posts and an author post that did well. Nothing here hit GitHub Trending or a tracked curator, so none of it is HOT by our rule.

Related tools on ReelStack

Skills & repos in our directory

More from the Lab

Topics: HyperFrames · Remotion · ComfyUI · Higgsfield