Skip to content

← The Lab

Markerboard draws whiteboard explainers to your voiceover, plus a Blender face-capture stack, ComfyUI blockout tips and Kevin Ngo's JavaScript drawing prompt

A brand-new open skill turns a topic into a marker-on-whiteboard explainer timed word by word to the narration. Also today: a Claude-scripted Blender face-capture stack, ComfyUI's advice on how detailed a blockout should be, the prompt behind Kevin Ngo's viral JavaScript drawing, and LTX's write-up of its SDR-to-HDR research.

Markerboard (arifbuildsai, MIT, published Nov 13). Open the folder in Claude Code or any agent that runs commands and give it a topic. The skill asks one question first: are you supplying the voiceover and word timings, or should it produce them? Then it writes a beat-numbered script showing what gets drawn on each beat, and stops until you approve it, because a re-record would break every timing. After that it makes the voiceover (your ElevenLabs key, an MCP voice tool or a local model), transcribes it for word-level timestamps, and renders one continuous board where a marker draws each mark on its spoken word while the camera follows. Eight automated checks run before you get the MP4, usually about 30 seconds long. It needs Node, a Chromium browser and ffmpeg, and runs on macOS, Windows and Linux. It's hours old with no stars yet, so treat it as early.

A Blender face-capture stack. @thegreatest_sv posted a cartoon face that blinks, smirks, winks and laughs in sync with audio, with no hand-set keyframes (about 200 likes). Per the post, a script Claude wrote reads a real face on video and drives the eyes, brows, mouth and tongue frame by frame. The pieces are all existing open tools: MCP for Blender for scene control, MediaPipe face tracking (52 blendshapes), BlendArMocap to turn that tracking into Blender animation, and Rhubarb Lip Sync for mouth shapes from the voice track. The useful idea is to capture your own performance and let the agent do the retargeting. Track only performers who've agreed to it.

ComfyUI on blockouts. In a short explainer, ComfyUI says there are two blockouts that work with video models like Seedance 2.5 or MiniMax H3. One is simple shapes that only fix composition, with the action described in the prompt so the model is free to animate it. The other is fully animated, precisely timed motion when you need control. Avoid the middle: a stiff, half-animated blockout tends to come through stiff. Separately, ComfyUI shared a Comfy Cloud workflow that pushes the camera from a museum wall into a painting and on through it as a 3D space (about 120 likes, 100+ bookmarks), a ready-made 'step inside the image' transition.

Kevin Ngo's drawing prompt. His Nov 11 'JavaScript drawing by Claude Fable 5.2' passed 2,500 likes. On Nov 13 he posted the prompt (about 490 likes, 650 bookmarks). The structure is worth borrowing. Point the agent at a folder of your own reference art and have it study it. Let it choose 3 to 5 dry media. Insist that it look hand-drawn, 'not like an SVG'. Constrain it to JavaScript on a 16:9 canvas. Give a long emotional story it must evoke without depicting anything from it, and name a colour era ('90s anime') while asking for something original. He also said to keep subagents on the same model and skip headless Chrome to save laptop memory.

LumiVid from LTX. LTX's research team posted a write-up of LumiVid, the method behind its SDR-to-HDR tool. A small LoRA on a frozen LTX-2.5 model, trained on about 300 paired clips, turns 8-bit footage into scene-linear half-float EXR you can re-expose and grade. The paper reports beating the strongest dedicated baseline on ARRI footage (the authors' PU21-PSNR numbers). The open IC-LoRA weights have been on Hugging Face since late September under LTX's licence, and the hosted version is the SDR to HDR endpoint we covered when LTX opened its VFX API on Nov 10.

Signals: Markerboard is the author's own launch post only. The face-capture stack is a creator post that took off modestly. The ComfyUI and LTX items are official posts, and Kevin's prompt is a tracked creator's post. None meets two HOT signals.

Skills & repos in our directory

More from the Lab

Topics: ComfyUI · Audio, voice & music · Open-weight video & audio models · Video MCP servers