Skip to content
Bud's Wiki

Equilibrium Forcing (EqF) — extending Equilibrium Matching to budget-adaptive long-horizon video generation

TL;DR

Section titled “TL;DR”

Yilun Du announces Equilibrium Forcing (EqF), an extension of Equilibrium Matching (Equilibrium Matching: Generative Modeling with Implicit Energy-Based Models) to video generation. The announcement claims two ingredients on top of EqM: (a) the video generator adapts its inference procedure to the sample it is currently generating (per-sample budget-adaptive sampling), and (b) this improves long-horizon video generation. Beyond the tweet itself there is no linked paper, project page, or code at filing time — this page exists as a pointer for later paper-page ingestion once the underlying artifact appears.

Key claims

Section titled “Key claims”
  • EqF extends EqM (single time-invariant equilibrium gradient / implicit energy landscape) into a video generator [tweet body].
  • EqF enables the generator to adapt its inference procedure to the sample it is generating, framed as budget-adaptive sampling [tweet body].
  • EqF improves long-horizon video generation [tweet body].

Method

Section titled “Method”

Not disclosed in the tweet. The natural read, given the EqM foundation, is that the autonomous (time-invariant) gradient field of EqM lets the sampler choose step count, step size, and optimizer per-sample rather than following a fixed integration horizon — and EqF is the video-domain instantiation of that property, where “adaptive to the sample” means more sampling steps on frames or clips that are harder to bring back to the data manifold. This is inference from the EqM abstract’s “adjustable step sizes, adaptive optimizers, and adaptive compute” property carried into video; the tweet does not confirm the mechanism.

Results

Section titled “Results”

No quantitative results in the tweet. No linked paper, code, or benchmark at filing time.

Why it’s interesting

Section titled “Why it’s interesting”

The Slack framing (@amit: “chip-by-chip the error accumulation issue in autoregressive generations will be solved. this seems like a right step”) is worth interrogating against the wiki. The Autoregressive Video Generation page has been tracking a forgetting–drifting dilemma: streaming-AR video models trained on 5s clips diverge from the real-video manifold when rolled out for 20–60s. The filed fixes cluster into four training-time injection sites for a “make memory look drifted at training time” idea (MALT’s noise-augmented memory, Context Forcing’s Error-Recycling, LongVie 2’s degradation-aware training, FlowAct-R1’s generated-GT memory sampling) plus inference-time correctors (WMReward’s VJEPA-2 surprise BoN on frozen MAGI-1; FlowAct-R1’s Memory Refinement re-noise+re-denoise pass). EqF, if the tweet’s characterization holds, would be a fifth class: an architectural-parameterization fix that makes the sampling procedure itself adaptive to how far the current sample has drifted, without needing a separate reward model or a memory-perturbation training curriculum. That would sit closer to Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation / Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation on the parameterization axis than to the memory-cache-management axis, and would inherit EqM’s “no timestep conditioning” property which is unusual for video diffusion.

The tweet is thin — one sentence, no numbers, no arxiv link — so this page should be treated as a watchlist entry. When a paper / project page / code appears, it will get its own arxiv-keyed page and this tweet page will become a pointer.

See also

Section titled “See also”