The same face in every MiniMax H3 clip, with no LoRA training. Part 1 turns 8 pictures of one person into one small face file (a RefMod). Part 2 reads that file and makes an H3 video with sound that keeps the face. Both are free ComfyUI workflows.
What is in the zip
- Part 1: the workflow that makes the face file from your pictures
- Part 2: the workflow that makes the video, for one person or two
- 8 sample pictures for each of the two people in the video
- The prompts and seeds from the video, and a QUICKSTART with every model file and where it goes
Settings that work
- Face file: 6 to 10 pictures of one person, same clothes in all, max tokens at 0
- Video: 4 steps with the 4-step turbo LoRA, 1536 x 864, 124 frames (5 seconds)
- Two people: one face file each, and "on the left" and "on the right" in the prompt. If two faces mix, change the seed.
Needs: ComfyUI with MiniMax H3 support, the free ComfyUI-MiniMaxH3Mod add-on by Luisa (MIT licence), and the H3 model files and turbo LoRA from Comfy-Org. About 6 minutes a clip on our machine.
More MiniMax H3 workflows and the full write-up: free H3 RefMod page
Description
FAQ
Comments (4)
Well afaik, part 1 makes an identity lora, part 2 reads the lora. So not quite "no lora".
Fair question! Part 1 doesn’t train anything, so it isn’t a LoRA. It just reads your pictures and saves them as an encoded reference file, which takes a few seconds. Part 2 hands that file to H3 as a reference. Training never happens and the model itself is never changed, which is why I call it no LoRA. The only LoRA in there is the 4-step speed LoRA. It makes the video faster and has nothing to do with the face.
for me worked perfectly! Thanks for sharing!
Glad it worked perfectly for you, thanks for trying it!