kitsune
A lab for turning a recorded clip of myself into an AI character, with open models running locally.
- ComfyUI
- FLUX.1 Kontext
- Wan 2.1 VACE
- PyTorch
What it is
A learning project in AI video. I restyle one frame of a short clip of myself into a different character, then carry that look across the whole clip while keeping my movement. Everything runs locally in ComfyUI on a laptop. Recorded clips only, never live, and anything I post is labelled as AI.
What’s interesting
- Two models, two jobs. FLUX.1 Kontext edits one frame (prompt the change, not the scene). Wan 2.1 VACE takes that frame as the appearance and a pose or depth map of the original clip as the motion.
- Pose vs depth. Pose keeps the body but loses the hair and clothes outline; depth keeps the outline but carries some of my shape over. The interesting part is choosing per clip.
- Small model first. Quantized models on a Mac to learn the workflow, then the same workflows with the large models on a rented GPU.