MiniMax H3 makes great 5-second clips. The hard part is making a longer film where the room, the props and the characters still look the same after a cut.
H3 Continuity Studio is a new ComfyUI node pack built for exactly that. It gives you a timeline of clips, four ways to join them, and a “visual memory” that feeds earlier moments back into H3 so later shots remember what came before.
This guide covers what it does, how to install it, how the join types and memory modes work, a first multi-clip film step by step, and the gotchas to watch for.
- What: ComfyUI-H3-Continuity-Studio by KadoBOT (MIT code), released 2-3 Nov 2026.
- Timeline: clips with their own prompt, pictures and settings; multiple takes per clip.
- Joins: Continue, Cut, New scene and Bridge (a short generated transition).
- Memory: earlier moments go to H3 as a native video reference (
<Video N>), straight from the saved latent, with no extra model. - Install: clone into
ComfyUI/custom_nodes/,pip install -r requirements.txt, restart, loadexamples/h3_story_studio.json. - Needs: a working MiniMax H3 setup and a ComfyUI build with the core MiniMax H3 Context Windows node.
What Is H3 Continuity Studio?
H3 generates about five seconds per pass, so a longer video is always a chain of clips. Every join is a chance for things to drift: a lamp moves, a shirt changes colour, a face looks different.
Continuity Studio adds a Studio panel on top of a normal H3 workflow: a timeline, a viewer and a clip inspector. Each clip keeps its own prompt, reference pictures and settings, and you can render one clip or the whole sequence.
| Feature | What it gives you |
|---|---|
| Timeline | A list of clips you can reorder, split and zoom |
| Takes | Several candidate renders per clip; keep the best |
| Join types | Continue, Cut, New scene, Bridge |
| Visual memory | Earlier moments passed back to H3 as a video reference |
| Dependency tracking | Re-render a clip and later clips that depend on it are flagged, with the reason in a tooltip |
| Story writer (optional) | An LLM plans and writes the timeline for you |
The pack moved fast: five tagged releases on 2-3 November added resizable panes, timeline zoom, split-at-guide and saved stories.
Install It in ComfyUI
First make sure H3 itself works. If not, set it up with my MiniMax H3 ComfyUI guide; on 16 GB, the H3 GGUF workflow is the practical route.
- Update ComfyUI. You need a build that includes the core MiniMax H3 Context Windows node.
- Clone the pack:
cd ComfyUI/custom_nodes git clone https://github.com/KadoBOT/ComfyUI-H3-Continuity-Studio - Install requirements with ComfyUI’s Python:
pip install -r ComfyUI-H3-Continuity-Studio/requirements.txtOn Windows portable:
..\..\python_embeded\python.exe -m pip install -r ComfyUI-H3-Continuity-Studio\requirements.txt. - Restart ComfyUI.
- Load
examples/h3_story_studio.jsonfrom the pack’s folder. - Point the loaders at your files: H3 diffusion model, text encoder, video VAE, audio VAE and any LoRA. The example only stores the author’s file names and downloads nothing.
Your H3 files stay in the usual folders: models/diffusion_models/, models/text_encoders/, models/vae/ and models/loras/. My ComfyUI model folders cheat sheet lists them all.
Warning: do not install a second copy next to an existing one. Both register the same nodes and routes. Back up your projects before replacing an install.
The Four Join Types
| Join | What happens | Use it for |
|---|---|---|
| Continue | Carries the previous clip’s context forward | The same shot going on longer |
| Cut | Starts a new target from a new prompt | A new angle in the same scene (where drift usually sneaks in) |
| New scene | Starts a new scene but keeps the film’s visual history | Moving to another location with the same characters |
| Bridge | Renders a short generated transition between two clips | Smoothing a jump without re-rendering either shot |
The pack’s own advice: render the first clip, add a continuing clip, render it, and check the join before building further.
Visual Memory Modes Explained
This is the clever part. Earlier moments are passed to H3 as a native video reference, which you refer to in the prompt as <Video N>.
Each remembered moment is a 5-frame window taken from a saved clip’s original video latent. It goes straight to the model with no decode and re-encode, so no quality is lost and no extra model is needed.
| Memory mode | What the next clip receives | Best for |
|---|---|---|
| Same scene | The scene’s establishing view plus a midpoint from the previous clip | Most cuts inside one location |
| Selected moments | Up to three views you pick from any earlier clip, each with a description | Bringing back a prop, room or outfit from far earlier |
| Cast | Each scheduled character as they look now, plus the place | Character consistency (needs a story timeline) |
| Off | Nothing; a plain H3 generation | Deliberate style breaks, or comparing results |
At a cut, every memory mode except Off also passes the full previous clip’s latent.
Note: H3 snaps each moment to a 17-frame boundary, so a picked moment can start slightly before the time you chose. If a source clip is missing or out of date, the clip stops with “render the source first”.
Make Your First Multi-Clip Film
The pack ships examples/memory_demo.json, a six-clip film rendered at 1024×576 and 24 fps that needs no uploads. It is the best way to see each feature working.
- Load the demo and point the loaders at your H3 files.
- Render clip 1 (the living room) and check it looks right.
- Render clip 2 (a Cut outside) and look at the join.
- Render the Bridge to the porch view.
- Render the Cut back inside, which reuses moments from clip 1 through Selected moments.
- Compare the last clip, which has memory Off, with the others to see the difference memory makes.
Then build your own: start with one Continue and one Cut, keep each clip short, and only add more once the joins hold. My H3 prompt guide helps with the per-clip prompts.
For a quick comparison, the author suggests rendering the same seeds three ways: memory off, automatic scene memory, and a hand-picked view. Compare detail, motion at the joins, time and VRAM.
The Optional Story Writer
A Story node can plan the film and write the timeline for you. It talks to any OpenAI-compatible endpoint, local or cloud.
- Local and private: point it at Ollama or LM Studio’s server. See my best Ollama models list for a model that fits next to H3.
- Cloud: any OpenAI-compatible API works, but the story text (and reference pictures, if enabled) is sent to that provider.
- Optional prompt guides: the MiniMax prompt-guide text is not bundled; the pack runs fine without it and just reports it missing.
Tip: running a local LLM and H3 on one GPU will fight for VRAM. Unload the LLM before rendering, or run it on a second GPU (see my ComfyUI multi-GPU guide).
Tips for Better Continuity
- Keep clips short. The pack gives no VRAM figures; start with short clips that fit your setup.
- Describe what to keep. In Selected moments, write what matters (“the green sofa and the lamp on the left”).
- Use Same scene for most cuts and switch to Selected moments only when something far back needs to return.
- Use Bridge instead of re-rendering when two good shots do not quite flow.
- Keep the seed and LoRA steady across a scene for a consistent look.
- Finish later: upscale with Tessera or reach 60 fps with my H3 1080p 60 fps guide.
Troubleshooting
Missing “MiniMax H3 Context Windows” node
Your ComfyUI is too old. Update ComfyUI (not just the node pack) and restart.
“Render the source first”
A Selected moment points at a clip that is not rendered or is out of date. Render that clip, then try again.
Later clips marked as needing a render
You re-rendered a clip they depend on. Hover the clip to see why, then re-render it so its memory matches.
Nodes or routes registered twice
You have two copies installed. Remove one folder from custom_nodes and restart.
Out of memory
Shorten the clip, use a GGUF H3 model, or lower resolution. My H3 out of memory guide covers 16 GB setups.
Details still drift after a cut
Switch the clip to Selected moments, pick a clear earlier view of the object, and describe it. The author notes memory keeps details but does not guarantee identical geometry.
My Take
H3 Continuity Studio is the first tool I have seen that treats H3 like an editing suite rather than a clip generator. Feeding earlier moments back in at the latent level is a smart, lossless way to fight drift.
It is very new, and the author only checked the package on a CPU startup, not with full test renders. Start with the six-clip demo, then build your own film one join at a time.
FAQ
What does H3 Continuity Studio do?
It lets you build a longer MiniMax H3 film from many clips on a timeline, with join types and a visual memory that keeps scenes and characters consistent.
Does it need an extra model?
No. Memory uses H3’s own video reference input. You only need your normal H3 files.
How many earlier moments can a clip remember?
In Selected moments mode, up to three views from any earlier clip, each with a description.
Is H3 Continuity Studio free?
The code is MIT licensed. The H3 model itself follows MiniMax’s licence terms.
Can I use a local LLM for the Story writer?
Yes. Any OpenAI-compatible endpoint works, including local servers like Ollama or LM Studio.
How much VRAM does it need?
The project gives no figures. It needs whatever your H3 setup needs, so start with short clips and a GGUF model on 16 GB cards.