Goal-image generation for Flex-π zero-shot survey + small fine-tune · results browser
How to read the badgesΔLPIPS→end −0.03 LPIPS to the end-of-episode frame minus the same for the current frame (green: the image is closer to the end frame than the current frame; red: more than +0.10 further).
ΔLPIPS→+4s the same against the +4 s frame.
dir 0.31 DINO goal direction: cosine between the feature change (generation − current) and (end frame − current); +1 = moved exactly toward the end frame.
Δinput 0.18 LPIPS between the generation and its input (0 = unchanged copy).
12 s generation time on a shared A100.
On the current frame: id→end 0.22 / id→+4s 0.29 = LPIPS of the current frame to those futures (the identity baseline a model has to beat).
Click an image for full size, its prompt and (for think-mode runs) the model's written plan; ← → step through the row, Esc closes.