A consistent character from shot to shot
Train a character LoRA, then use it as the first frame of a video, to keep the same face across every shot of an ad or a series.
- Type
- ComfyUI workflow
- Status
- In production
- Updated
- September 15, 2026
- Stack
- ComfyUILoRAFlux.2MiniMax H3Comfy Cloud
The problem
The first wall AI video hits in production: the character's face changes from one shot to the next. You can't tell a story or hold a brand together under those conditions.
The pipeline
- Training a character LoRA from a set of reference images
- Starting portrait generated with Flux.2 + the LoRA, used as the first frame
- Image-to-video animation with MiniMax H3, with a prompt timed second by second to drive the expressions (down to tears rolling over 10 seconds)
- Upscale to 4K with SeedVR2