Character Consistency · 2026
Melinda
A LoRA-trained face carried through wardrobe, light and motion.
Final video · Jersey
- Task
- Prove that a custom-trained character can survive a full production pipeline: text-to-image references, an identity pass on generated frames, and image-to-video — without the face changing from shot to shot.
- My role
- Dataset curation
- LoRA training
- Identity pass
- Video generation
- QC & selection
- Pipeline
Process
4 steps- 01
References
The character is defined by a trained LoRA. These references come straight out of text-to-image with the LoRA applied — no retouching.
Z-Image LoRA · reference 01 
Z-Image LoRA · reference 02 - 02
Generated frames
Every scene frame runs through an identity pass and is compared side by side (A / B). Frames are selected on likeness first, aesthetics second.
Frame 01 · pass A 
Frame 01 · pass B 
Frame 04 · pass A 
Frame 04 · pass B - 03
Video generation
Selected frames are animated with image-to-video. Kit, environment and camera change; the face does not.
Kit change Car - 04
Final shot
Delivered vertical, native audio, no cleanup passes on the face.
Jersey
Generated frames
8 stills
