Any Forcing

Any-Step Generality and One-Step Stability
for Autoregressive Video Generation

Generated by Any ForcingShort & long videos · 480p

TL;DR

Any Forcing achieves any-step generality and one-step stability in autoregressive video generation with a unified adversarial discriminator and an adaptive long-term memory selection mechanism.

85.095s VBench · uniform 4 steps

85.065s VBench · first 4, rest 1

84.795s VBench · first 2, rest 1

84.5630s VBench · first 4, rest 1

84.5330s VBench · first 2, rest 1

See the difference

One-step autoregressive video generation

Short-video high-quality.
Long-video stability.

Text prompt

The framework

Learn quality. Remember what matters.

Training-time alignment.
Inference-time memory.

Any Forcing overview: unified adversarial discriminator with RA, CA and FE tasks, and adaptive memory selection with global relevance, local value contribution and positional remapping — **Overview of Any Forcing.** Unified adversarial supervision during training; adaptive historical memory during long-video inference.

Adaptive long-term memory selection

One-Step Long-Term Memory

Global sink012

Memory pool4080

Candidate100

Recent context101

Current frame102

Memory entry A

Global relevances_C > s_A

Local contributionU_C > U_A

Candidate C

Generated iceberg frame 102 — Current frame **102**

Temporal position

Quantitative evaluation

Flexible steps. Consistent performance.

Visual comparison figure

Any Forcing


Method	NFE	Total ↑	Quality ↑	Semantic ↑

480p · 946 prompts · 5 samples per prompt. Best scores are bold; second-best scores are underlined.

1^* and 1^†: four and two first-frame steps, respectively, followed by one step per frame.

Generated with Any Forcing