One MiniMax H3 workflow for ComfyUI, in four windows. Free and open source (MIT).
- ▶ The 0.15 video: FastH3 Omni vs PDMD in 48 clips (10 min): https://youtu.be/k5xSKv_cNIQ
- ▶ Walkthrough of all four windows (22 min): https://youtu.be/Xeyb0Y9bvho
- ❤ Patreon: https://www.patreon.com/Stuubzzz (new versions reach patrons first)
### What it does
- Setup decides how the video is made. Five modes: Speed (≈480p, 8 steps), Hybrid (≈540p, 10 steps), Quality
(≈768p, 20 steps), PDMD (≈768p in 4 steps) and FastH3 Omni (≈768p in 8 steps, Reference mode).
- Director holds what the video is: Reference or First/Last Frame mode, shots on a timeline, style, sound and
music, the cast with reference pictures, voices and clips, and LoRAs per chunk.
- Seeds plays 1–4 full seeds with sound side by side. You pick the winner.
- Output finishes only the winner: up to 4K, any frame rate from 24 to 120 (RIFE), RTX VSR or DLSS5 + Color Lock,
Quality polish and de-stutter.
### Also inside
- Videos longer than one chunk, rendered chunk by chunk, with a choice of how each chunk joins the one before
(picture and sound, fresh sound, last frame only, or a new shot).
- Takes: render one chunk again, or hunt it with seeds of its own. Every take is kept.
- Up to eight middle frames in First/Last Frame mode.
- Characters made from many pictures (RefMods), on every chunk.
- Voices: a recorded line, a voice sample, Lip Sync.
- Continue a clip, and pin sounds and clips on the timeline.
- Your own prompt: paste a whole H3 prompt (written by an LLM, for example) into the prompt override.
### New in 0.15
- FastH3 Omni mode: FastVideo's 8-step distilled Reference model, with its sparse attention built in. About 10 %
faster than PDMD on 15 s clips in my test, and a voice clip reaches every chunk.
- The prompt override in the open, under the Director, next to the compiled prompt.
### Install (the workflow needs its node pack)
1. Update ComfyUI (0.34 or newer; FastH3 Omni mode needs 0.42 or newer).
2. ComfyUI Manager → Install via Git URL → https://github.com/stuubszzz/Stubelius-Ultimate-H3 , then restart.
3. Load the workflow, then Manager → Install Missing Custom Nodes.
4. Models from [Comfy-Org/MiniMax-H3](https://huggingface.co/Comfy-Org/MiniMax-H3): the ref2va and fl2va pruned int8
models, the Qwen3-VL text encoder, the video and audio VAEs and the turbo LoRA. The full list with direct links is
in the [README on GitHub](https://github.com/stuubszzz/Stubelius-Ultimate-H3).
Developed on an RTX 5090 (32 GB). With less VRAM, turn on the low-VRAM switches in the Setup window, or use GGUF
models.
### Credits and license
Built on Muse Minimax Director V1.2 by Muse Collective. MIT licensed. PDMD mode uses the 4-step LoRA by Zimo Wang et
al. (Apache-2.0). FastH3 Omni mode uses FastH3 Omni by FastVideo (Hao AI Lab) and Nuva Lab. Powered by MiniMax H3,
used under the MiniMax H3 Community License.