Research map / Video / Video generation
Equilibrium Forcing: Adaptive Video Generation Without Noise Conditioning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Equilibrium Forcing (EqF), a noise-unconditional video denoising objective whose inference-time warp and closed-loop noise-level readout enable adaptive, gradient-based (NAG) autoregressive sampling that outperforms noise-conditional FM and diffusion baselines.
Problem
Noise-conditional autoregressive video denoisers follow open-loop noise schedules, so the scheduled noise level diverges from the sample's true noise level, causing compounding error over long rollouts.
Approach
Train a single velocity field without noise-level conditioning (unmodulated velocity target), decouple sampling dynamics via an inference-time eta warp, and use a learned framewise noise-level readout from internal activations to close the loop, enabling attractor-style NAG sampling and budget-adaptive step allocation.
Key results
- On Minecraft 300-frame rollouts, EqF with NAG (mu=0.3, c-function warp) reaches FVD 64.34 ± 1.78 vs 104.13 for FM Euler and 106.00 for Diffusion DDIM.
- On RealEstate10K finetuned from Wan 2.1 1.3B, EqF-FT with adaptive NAG achieves FVD 139.5 vs 174.3 for FM-FT.
- On ImageNet, EqF with NAG reaches FID 15.15 vs 18.91 for FM Euler and 15.74 for EqM NAG.
Surprising findings
- Standard FM inference terminates with nonzero residual noise, and conditioning on predicted noise levels only partly mitigates compounding error.
- EqF's learned field matches FM's denoiser closely (Pearson correlation >0.98 across datasets), so gains are attributed to inference rather than a different learned field.
Limitations
- Only video and ImageNet evaluated; no large-scale general video pretraining.
- Noise-level readout error could introduce source of error if partially noised samples are off-distribution.
- Datasets are not general video domain.
Experimental setup
Datasets: Minecraft RealEstate10K Droid ImageNet
Benchmarks: Minecraft autoregressive rollouts RealEstate10K autoregressive rollouts Droid prompt-conditioned generation ImageNet
Metrics: FVD VBench FID
Baselines: Flow Matching Diffusion (DDPM/DDIM) Equilibrium Matching NU-FM FM-FT
Builds on: Flow Matching Diffusion Forcing Equilibrium Matching Noise-Unconditional Flow Matching Nesterov Accelerated Gradient
Closest papers by meaning
- Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion (NeurIPS 2025 · similarity 0.89)
- FIFO-Diffusion: Generating Infinite Videos from Text without Training (NeurIPS 2024 · similarity 0.88)
- Equilibrium Matching: Generative Modeling with Implicit Energy-Based Models (NeurIPS 2026 · similarity 0.88)
- Spectral Progressive Diffusion for Efficient Image and Video Generation (NeurIPS 2026 · similarity 0.87)
- SF-V: Single Forward Video Generation Model (NeurIPS 2024 · similarity 0.87)
- AR-Diffusion: Asynchronous Video Generation with Auto-Regressive Diffusion (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.