A MiniMax H3 studio workflow for text-to-video, image-to-video, and reference-to-video. One fused turbo model, 8 steps, and one video file per run.
Pick a mode and queue.
- T2V: prompt only
- I2V: unmute First Frame. Turn on Use Last Frame if you also want an end frame
- REF2V: add pictures and tag them <Picture 1> and <Picture 2>. For a clip, unmute Reference Video and Unpack Reference Video together and tag <Video 1>
Grey nodes are optional and stay muted until you have a file for them. Duration is 5–15 seconds. Stay at or under 0.98 MP (1344×768). 0.8 MP is the faster everyday size.
SeedVR2 and X2 Detail are toggles. Leave both off for the raw generation. Turning either one on still saves a single file. (Gone in version 2.0, replaced with something else.)
Model links and the custom nodes used here are listed inside the workflow. Tested and built for a 16 GB card.
If you're going to ask why 0.98 MP and not 1.0 MP, honestly i have no idea... but if for some reason you are interested in finding out, here's the link: https://platform.minimax.io/docs/guides/local-deploy-h3
Description
# MiniMax H3 Studio 2.0
Changes since the first public release (v1.0).
## New
- Continue a clip (V2V). Load the previous clip, set how much of it to carry over, and the new clip picks up the motion. This mode did not exist in v1.
- First + last frame is a mode of its own now. In v1 the last frame was an extra option on I2V.
- More reference slots: 4 pictures, 3 videos, 2 audio files. v1 had 2 pictures, 1 video, 1 audio.
- Optional upscale and detail refine passes, and an optional 48 fps interpolation step.
- Two speed modes, TeaCache and Spectrum. Both are bypassed when you open the workflow, so quality is unchanged unless you turn one on.
## Changed
- The mode dropdown is gone. Turn on the boxes you need and the matching path is used. Anything left switched off is ignored.
- Main model is now Minimax-h3_Singularity_x_10Eros-Max_PRISM_w6a8. The v1 fused turbo8 file is not used.
- Defaults: 8 seconds, 9:16 at 0.4 MP, random seed, empty prompt box. v1 opened at 5 seconds, 0.98 MP, with an example prompt filled in.
- One file per run, saved to output/video/minimax_h3_*.mp4 at 24 fps. The SeedVR2 and X2 save chain was removed. (way to slow)
- The notes on the canvas were rewritten: install steps and model links, prompt advice, how reference tags are numbered.
## If you are updating from v1
1. Download the new model into models/diffusion_models/minimax/. (subfolder it's optional)
2. Run Manager > Install missing custom nodes, then restart ComfyUI.
3. v1 workflows do not open cleanly in 2.0. Keep the old file if you still want to use it.