ElasticTTT: Preventing Prior Collapse in Diffusion-Based Video Editing

ElasticTTT mitigates Prior Collapse in test-time tuning by using Target Distribution Regularization and Contrastive CFG to prevent the model from discarding text conditions. The framework preserves generative elasticity, ensuring that video edits follow prompts without collapsing into the source video's latent distribution.