MiniMax H3
Local video with a synchronized stereo soundtrack, from MiniMax H3 (33B, open weights), from a prompt, a first frame, an end frame, both, or reference pictures.
MiniMax H3 is a 33B open-weights video model. It generates a clip and its stereo soundtrack together, from a prompt, a first frame, an end frame, or both. The same picture at both ends makes a loop that lands back on the picture. The Qwen3-VL 32B text encoder reads the prompt. The two models never share the card.
Experimental. Only an admin can add it. Made for a 24 GB card with 64 GB of RAM: the pruned int8 build (~21 GB) streams part of itself from RAM. Read the MiniMax H3 Community License before use: it excludes some territories.
Setup
Add the service
Add Modules → Services → MiniMax H3 → Install, then follow the install steps. The Configuration step sets a folder per asset type, so you can point them at an existing ComfyUI collection. The Models step offers the weights to download now.
Pick a preset
| Preset | For |
|---|---|
| MiniMax H3 Turbo (8 steps) | The lightx2v Turbo LoRA, merged as the model loads. About a minute per 5-second clip on an RTX 4090. |
| MiniMax H3 (24 steps) | No Turbo LoRA: slower. |
| MiniMax H3 10Eros Max (6 steps) | A fine-tune with Turbo baked in. |
Models
| Setting | Choices |
|---|---|
| Model | FL2VA pruned int8 (~21 GB, recommended), Fused Turbo int8 (Turbo baked in, 4 steps, no LoRA), 10Eros Max TURBO-hybrid int8, FL2VA pruned bf16 (~40 GB, needs 64 GB+ of RAM) |
| Text Encoder | NVFP4-AWQ (~16 GB, recommended), Heretic NVFP4 (follows prompts more closely), int8 (~27 GB, closer to bf16, slower to load), bf16 (~48 GB) |
| Video VAE | fp16 (~4.9 GB) or int8 (~2.8 GB, lighter) |
Settings
- Steps: 4 to 8 with Turbo, about 20 to 30 without.
- Sampler: res_multistep (recommended) or euler.
- Resolution: 480p to 768p (768p is the native canvas).
- Duration (seconds): 5 to 15, at 24 fps.
- Generate Audio: keep the soundtrack H3 renders.
- Images Are Used As: Start / end frame, or References: up to 9 pictures the clip takes its characters, outfits and places from, named
<Picture 1>,<Picture 2>… in the prompt. - Enable Preview: a rough animated preview after each step.
- Loops End On The First Frame (module setting): on by default. Off, a loop gets the first frame only and asks for the loop in the prompt, for when H3 reaches the end frame early and holds it.
Quoted speech in a plain prompt (says "...") is passed to H3 as dialogue, so it is spoken rather than burned in as a subtitle.