One Character, Twenty Worlds
Filip Ronaj
Every frame scored: face match to the one reference image with OpenCV SFace. Average 0.69, lowest 0.45, same-person line 0.36. Three different characters score 0.10 or less.
Self-initiated study, not client work.
The question: can one AI character stay the same person across very different scenes when the only identity input is a single reference image?
One original character, one reference image, twenty scenes with new light, angle, outfit and expression, built in ComfyUI without training a character LoRA. Every scene was scored against the reference with OpenCV SFace face matching: average 0.69, lowest 0.45, where 0.36 is the same-person line. Three different characters score 0.10 or less.
Two scenes were then animated into short image-to-video clips; the flower-market clip is below.
The flower-market scene animated into a 5-second clip (image-to-video, MiniMax H3). All 21 sampled frames stay above the same-person line.
Like this project
Posted Oct 15, 2026
Self-initiated study: one AI character in twenty scenes from one reference image. Face match (OpenCV SFace): average 0.69, lowest 0.45; same-person line 0.36.
Likes
0
Views
0
Timeline
Oct 8, 2026 - Oct 13, 2026
