ElasticTTT: Prior-Preserving Test-Time Tuning for Video Editing
ElasticTTT is pitched as a fix for “Prior Collapse” in one-shot diffusion video editing.
The paper says standard test-time tuning can push a pretrained diffusion model into memorizing the source video, ignoring text conditions or tangling separate regions. ElasticTTT adds target distribution regularization, contrastive CFG, and an asynchronous noise schedule to keep the model’s prior behavior usable during edits. The authors report state-of-the-art one-shot video editing results, backed by evaluations and theoretical analysis. ArXiv · AI/CL/LG's note
The paper says standard test-time tuning can push a pretrained diffusion model into memorizing the source video, ignoring text conditions or tangling separate regions. ElasticTTT adds target distribution regularization, contrastive CFG, and an asynchronous noise schedule to keep the model’s prior behavior usable during edits. The authors report state-of-the-art one-shot video editing results, backed by evaluations and theoretical analysis. ArXiv · AI/CL/LG's note
score 5