ElasticTTT: Prior-Preserving Test-Time Tuning for Video Editing

2026-07-23

AI Summary

This work identifies and formally defines “prior collapse” in standard Test-Time Training (TTT) for video editing—a failure mode wherein the model disregards textual conditions and spatial latent variables, degenerating into mere reproduction of the source video or mixing regional features. To address this, the authors propose ElasticTTT, a novel framework that mitigates the fundamental mismatch between distribution mapping and single-point optimization inherent in TTT. ElasticTTT integrates target-distribution regularization to prevent overfitting, contrastive classifier-free guidance (CFG) to debias inference, and asynchronous noise scheduling to preserve unedited regions. Evaluated on single-example video editing tasks, the method achieves state-of-the-art performance, with both theoretical analysis and empirical results demonstrating its superior ability to maintain generative priors and significantly enhance editing fidelity.

Abstract

Test-Time Tuning (TTT) on pretrained diffusion models has emerged as a powerful paradigm for video editing. However, there exists a foundational mismatch between the distribution-mapping nature of generative models and the single-point optimization of standard TTT. In this paper, we demonstrate that this mismatch triggers \textit{Prior Collapse}, a degenerate state where the model discards the text conditions and spatial latents, collapsing generations to the source video, or entangling the features of distinct regions. To resolve this, we propose \textbf{ElasticTTT}, a novel framework that preserves the prior generative distribution and rescues generative elasticity. Specifically, we propose \textit{Target Distribution Regularization} to prevent sharp memorization minima, \textit{Contrastive CFG} to guide inference away from source biases, and \textit{Asynchronous Noise Schedule} to preserve unedited regions. Extensive evaluations, supported by theoretical analysis, demonstrate that ElasticTTT successfully preserves the generative prior of the base model, achieving state-of-the-art performance on one-shot video editing.

Problem

• Test-Time Tuning
• Prior Collapse
• Video Editing
• Diffusion Models
• Generative Prior

Innovation

• Test-Time Tuning
• Prior Collapse
• Target Distribution Regularization
• Contrastive CFG
• Asynchronous Noise Schedule