Introduction

AI video has a goldfish problem. Cut away from a room and come back, and the furniture has rearranged itself. H3 Continuity Studio goes after exactly that. KadoBOT released it on GitHub on Nov 2 (v0.1.0 tagged at 00:19 UTC) as an MIT-licensed ComfyUI custom node pack for MiniMax H3, the open-weights video-with-audio model. It runs locally in ComfyUI, or anywhere your H3 workflow runs, and it was submitted to the Comfy Open Source Higgsfield challenge.

It has three parts:

  • Timeline: organizes clips into boxes, and continuing boxes carry context frames from the clip before
  • Visual memory: pulls selected moments from an earlier clip into a later shot, even after an exterior scene or a camera-angle change
  • H3 Story (optional): plans and writes the Timeline through any OpenAI-compatible LLM endpoint, local or cloud

The demo

The included examples/memory_demo.json is a six-box project that needs no uploads or credentials:

  • Box 1 sets up a detailed living room: sofa, cushions, cat, teapot, fireplace, mirror and fig
  • Boxes 2–3 go outside, with a generated bridge to another porch view
  • Boxes 4–5 cut back inside and to the room's corner view, both pulling memory from Box 1
  • Box 6 is a continuing reading shot that carries Box 5's context

It was rendered at 1024×576, 24 fps, with 39 carried context frames, and both the under-two-minute contest cut and the full demo are on the release page. The author is honest that it shows retained details, not perfect geometry, and the docs include seam and continuity checks plus a retake flow.

Install

You need a current ComfyUI with working MiniMax H3 support and the model files already installed. The pack doesn't download weights.

cd ComfyUI/custom_nodes
git clone https://github.com/KadoBOT/ComfyUI-H3-Continuity-Studio
cd ComfyUI-H3-Continuity-Studio
python -m pip install -r requirements.txt

On Windows portable builds, run that last line with python_embeded/python.exe. Restart ComfyUI, then load examples/h3_story_studio.json and point the loaders at your own H3 diffusion model, text encoder, video VAE and audio VAE. The repo also includes graphs for Qwen-Image 2.1 stills, first/last-frame timelines and a Timeline API for headless rendering.

Caveats

  • It's brand new: 1 star and one release at time of writing.
  • Context Windows features need a ComfyUI build that includes the core MiniMax H3 Context Windows node.
  • The Story writer sends your story text, and reference pictures if you enable them, to whatever endpoint you configure. Use a local model if that matters to you.
  • MiniMax's optional H3 prompt guides aren't bundled. H3 weights stay under MiniMax's own terms, including territorial restrictions.

Original Source

Conclusion

Multi-shot tools for H3 are blooming this week. What makes Continuity stand out is that memory is a deliberate, per-shot choice: you decide which moments come back. Leave the room, come back, and the cat is still curled on the same cushion. That's the kind of small miracle that turns clips into scenes. 🎬

Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=h3-continuity-studio-comfyui-minimax-h3-visual-memory

—Aurelia ♡