Speculative Forcing for Efficient Autoregressive Video Generation

Short video generation

On 5-second videos with LongLive, both verifiers keep the VBench total of the four-step generator while using fewer than 1.7 denoising steps per token on average.

Long video generation

On 30-second videos, the learned verifier keeps the quality and consistency of the four-step generator at about 1.7 steps per token. Drift, the variation of imaging quality over the rollout, is also lower than with four-step sampling.

World generation

On Lingbot-World 2.0, camera control keeps revealing new content, so more tokens need the full schedule. Speculative Forcing still keeps the quality of the four-step generator with about 2.5 steps per token, with lower Drift and comparable camera-pose error.

Denoise, verify, commit

Few-step autoregressive generators run every token through the full denoising schedule. Many tokens already match their full-step output after the first steps, while others keep changing until the end. Speculative Forcing lets each token stop at its own step, without changing the generator’s weights.

• Draft

  Each denoising step already predicts a clean estimate for every token. We treat this intermediate estimate as a draft of the token’s final output.

• Verify

  A verifier predicts how far each draft is from the full-step output. We study a training-free verifier and a learned one, a small 3D convolutional head on the frozen generator’s features.

• Commit

  Accepted tokens are fixed as final outputs and skip the remaining steps. Their cached keys and values stay available, so the rejected tokens keep refining with the full context.

Ablation study

All ablations use the world generation benchmark on Lingbot-World 2.0, where intermediate distortions are most pronounced.

Verification tolerance

One test-time tolerance ε trades speed against quality without retraining. With ε between 0.20 and 0.24 the sampler keeps the quality of four-step sampling and reaches the lowest Drift.

Quantitative results

Mean steps is the measured average number of denoising passes per token. Throughput is measured on a single NVIDIA H100. Bold marks the best value in each column.

Routing strategy

Lingbot-World 2.0, every method calibrated at ε = 0.24 to a similar mean number of steps. Channel variance and temporal difference are selection signals adapted from prior work. Chunk-wise acceptance uses our learned verifier but commits whole chunks. The oracle runs every step to measure the true gap and reports no throughput.

Abstract

Few-step autoregressive video generators enable streaming generation, yet they run every token through the final denoising step, regardless of how much refinement it needs. We find this uniform computation is often unnecessary: many tokens already closely match their eventual full-step outputs in the initial steps, while others require refinement until the end. We introduce Speculative Forcing, a framework that treats intermediate clean estimates as drafts and predicts when individual tokens can be safely accepted early. Once accepted, a token is frozen as its final output and undergoes no further denoising, while its cached keys and values remain available as context for the remaining tokens. We explore two alternative verifier designs: a training-free one based on similarity to clean content and stability across steps, and a learned one that reads the frozen generator’s intermediate features to estimate each token’s remaining gap to its full-step output. Across short-video, long-video, and world generation, Speculative Forcing preserves the quality of full-step sampling while substantially reducing computation. On LongLive, it matches the four-step generator with ∼1.7 denoising steps per token, running over 1.5× faster at 41.7 and 37.8 fps. On Lingbot-World 2.0, where camera motion continually changes the visible content, it maintains quality with ∼2.5 denoising steps per token and runs 1.35× faster at 24.5 fps.

Citation

@article{choi2026speculative,
  title   = {Speculative Forcing for Efficient Autoregressive Video Generation},
  author  = {Choi, Hyunwook and Kim, Seungryong and Han, Dongyoon and Yun, Sangdoo},
  journal = {arXiv preprint arXiv:2610.22347},
  year    = {2026}
}