Equilibrium Forcing: Adaptive Video Generation Without Noise Conditioning

2026-08-10

AI Summary

This study addresses the lack of adaptive inference in existing video generation models caused by their reliance on fixed noise conditioning. To overcome this limitation, we propose EqF, a noise-condition-free generation framework that introduces a modular architecture to decouple denoising learning from the sampling process. By leveraging Equilibrium Forcing and a closed-loop feedback mechanism, EqF enables data-dependent adaptive inference. Experimental results demonstrate that EqF significantly outperforms traditional noise-conditioned methods on autoregressive video generation benchmarks, yielding substantial improvements in both generation quality and temporal consistency. Consequently, this work establishes a more flexible paradigm for video generation by eliminating the constraints of static noise priors and facilitating dynamic adaptation during inference.

Abstract

Standard autoregressive video generation algorithms based on Diffusion and Flow Matching rely on rigid training objectives and static sampling schedules, limiting inference procedures from adapting to the data. We introduce Equilibrium Forcing (EqF), a simplified framework for video denoising generative models without noise level conditioning. EqF pioneers modular training- and inference-time designs for noise-unconditional generation that decouple learning the denoising field from sampling. This flexibility allows for inference-time algorithms that operate in a closed loop by adapting to feedback from the sample, improving video quality and consistency on challenging autoregressive video generation benchmarks. Extensive analysis elucidates exactly how removing the noise level conditioning enables EqF's data-dependent inference properties to surpass the performance of standard noise level-conditional denoising video methods.

Problem

• Autoregressive Video Generation
• Noise Level Conditioning
• Adaptive Inference
• Diffusion Models
• Flow Matching

Innovation

• Equilibrium Forcing
• Noise-Unconditional Generation
• Modular Design
• Adaptive Inference
• Autoregressive Video Generation