What it is
MiniMax-H3 makes picture and sound together, and this workflow puts the whole thing in one node.
Most H3 workflows lose their users before the first render: a missing custom node, a model in the wrong folder, a number the model quietly refuses. This one is built the other way round — one pack to install, models ComfyUI downloads for you, five numbered sections on the canvas, and nothing that can be set to a value the model rejects.
What you can make
Image to video. Drop in a picture; the clip starts there.
First AND last frame. Give it both and it renders the journey between them.
Text only. Switch both image loaders off and it still generates sound.
2, 5, 10 or 15 seconds at 24 fps, portrait or landscape.
Built to be quick to test
Draft, fast and full-quality tiers. Try ideas on a draft canvas for a fraction of the render time, then switch to full quality for the take you keep.
"Match my image" takes the shape from your own picture, so nothing is stretched or cropped — a landscape photo does not come out squashed inside a portrait canvas.
Steps default to 14, and 8 is plenty while you are still deciding what you want. Start on the 2 s quick test to prove your install before spending a long render.
Customisable where it actually matters
Swap the checkpoint. The model arrives on a link, so any H3 fl2va file works — a smaller community quant if 21 GB is too much for your card, a fine-tune, whatever you prefer.
Two LoRA slots, shipped switched off, with nothing chosen for you. Pick your own file and press Ctrl+B to turn one on. Stack two, or copy the node for more.
Everything else is fixed on purpose — sampler, scheduler, cfg, sigma shift, negative prompt. Those are the dials whose only realistic effect here is to break a render that already works.
Two things worth knowing
Say what you want to HEAR. Sound is generated with the picture, so name it in the prompt: voices, footsteps on gravel, rain on a window, room tone, a distant foghorn. That is what separates H3 from an ordinary image-to-video workflow, and a prompt that ignores sound wastes half the model.
The info box tells you what happened. Section 5 prints the frame count, the canvas, which checkpoint loaded, which LoRA was active and how your pictures were fitted. When something looks off, read that before anything else.
Install
ComfyUI Manager — Manager > Custom Nodes Manager > search SimpleShot > Install > Restart. (Just published, so if it is not listed yet, use the ZIP.)
Or a ZIP — download from github.com/manzakin/ComfyUI-SimpleShot (green Code button > Download ZIP), unzip, move the folder into ComfyUI/custom_nodes/ and rename it to ComfyUI-SimpleShot. Restart. It must end up as ComfyUI/custom_nodes/ComfyUI-SimpleShot/__init__.py — one folder too deep and ComfyUI will not see it.
Drag the JSON from the download onto your canvas, and say yes when ComfyUI offers to fetch the models.
Nothing to pip install. No second custom pack. No API key and no paid service — this runs entirely on your own machine.
Requirements, honestly
The four model files total about 42 GB, and the checkpoint alone wants a strong GPU. That part cannot be made small, and pretending otherwise would waste your evening. The mitigation is built in: point the diffusion loader at a smaller community model and the workflow does not care which one you use.
If a node ever refuses to run, its message is meant to be a complete instruction. If you get one that leaves you stuck, that is a bug worth reporting.
Which checkpoint — fl2va, not ref2va
MiniMax-H3 ships as two different trainings and this workflow is the fl2va one — the first/last-frame task. A ref2va checkpoint will not work here: it is built to render from reference photographs and cannot be aimed at a last frame, so it is the wrong tool rather than a worse one. Check that the filename says fl2va before you spend 21 GB of download on it.
If you do load one by mistake, the node stops and says so instead of burning a render on it. A ref2va workflow is a separate release, later.
Description
First release. Needs the SimpleShot node pack (one pack, from ComfyUI Manager or GitHub); the model files download themselves. fl2va only — a ref2va checkpoint will not work here. Fifteen nodes in five numbered sections, two optional LoRA slots, and draft / fast / full-quality render tiers.