acceptodds
Withdrawn from ICLR 2027

Commit-One: Decoupling Generation and Commitment for Interactive Video Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Interactive video generation requires new conditions to influence ongoing playback promptly. A natural route is fine-grained autoregressive generation, but reducing chunk size can restrict joint temporal modeling and parallel computation. *Must fine-grained interaction therefore come at the cost of a large chunk size?* We introduce Commit-One, an algorithm-system co-design built on a different principle: **low-latency interaction requires fine-grained output commitment, not necessarily fine-grained generation.** By decoupling chunk size from output commitment, Commit-One preserves large-chunk generation while allowing future predictions to remain revocable. This enables chunk-wise autoregressive models to deliver fine-grained video streams without reducing their chunk size. The optimization target therefore shifts from making chunks smaller to making joint denoising faster: chunk size becomes a design choice balancing modeling capacity and computational efficiency, rather than a hard constraint on interaction granularity. The same principle opens the possibility of using even bidirectional teacher models directly for interactive streaming. Commit-One thus offers an alternative route to interactive video generation, centered on accelerating high-quality joint generation rather than shrinking chunk size.

Primary area:
generative models
Keywords:
Interactive Video Generation, Autoregressive Video Generation, Video Diffusion

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Uo5EOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper examines interactive video generation through the lens of a serving-system problem. By decoupling chunk size from output commitment, Commit-One enables the generation of large chunks while ensuring that future predictions remain revocable. Experimental results demonstrate that this approach improves visual quality across various autoregressive video generation models.

Critical strengths

Well-motivated and practically relevant problem. The quality–latency tension in interactive video generation is real: small chunks improve responsiveness but weaken joint temporal modeling, while large chunks improve quality but impose coarse update boundaries. The idea of decoupling generation chunk size from output-commitment granularity is simple, intuitive, and useful. The interaction system proposed by this method is rigorous and clear; it encompasses stages such as conditional arrival, generation timing, and playback waiting, and incorporates precise scheduling logic along with a mechanism for re-planning. This paper presents a comprehensive set of experiments; evaluations conducted across various model backbones demonstrate the generalization capability and effectiveness of the proposed method. The results obtained in this study align with expectations: the bidirectional model outperforms the "large chunk" model, which in turn outperforms the "small chunk" model—all while maintaining nearly the same level of real-time interactivity as the "small chunk" model.

Critical weaknesses

When user input changes frequently (high-frequency interaction)—for instance, when integrated into a world model to track mouse and keyboard inputs—the latter portion of a large chunk becomes virtually unreusable. This leads to significant computational waste and increased interaction latency. When using a large chunk size, could frequent changes in user instructions cause issues with caching the latents from earlier chunks? For example, if bidirectional attention is still used within each chunk, could introducing subsequent latents that are temporally discontinuous lead to abrupt transitions in the generated video or artifacts caused by discontinuities during VAE decoding? The paper mentions that the H200 GPU can support parallelization with large chunk sizes. However, smaller chunk sizes may actually offer greater advantages in terms of deployment and execution on consumer-grade GPUs. Therefore, I believe this argument does not fully justify the motivation behind the proposed method. The authors have not adequately considered the responsiveness of the proposed method in certain application scenarios. A more thorough discussion of this issue is needed.

Questions for authors

See Weaknesses.

Reviewer_ULAPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Commit-One for interactive video generation. It generates multiple latent frames together, but delivers them incrementally. This allows future predictions to be replaced when the user want to change the condition.

Critical strengths
  1. It is useful to distinguish between generation and output commitment. It provides a practical way to improve interaction experience without reducing the generation chunk size.
  2. Table 3 demonstrates consistent, large improvements on latency across 3 video models.
Critical weaknesses
  1. The reuse of KV cache has a potential problem that is acknowledged, but it is not evaluated. Since the attention is bidirectional within a chunk, the remained KV may already contain information from the future frames. It might hinder to generate the video with huge scene change or with multiple-events, especially the new prompt is totally differ from the previous one.
Questions for authors

Please see Weaknesses.

Reviewer_z2xiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper argues that low-latency interactive video generation needs fine-grained output commitment, not fine-grained generation. Its system, Commit-One, keeps large-chunk joint diffusion sampling but commits output one latent at a time, so cached future predictions stay revocable. The authors build a serving-latency model showing that chunk-boundary commitment adds roughly half a chunk period of wait, and a chunk-size study suggesting larger chunks often improve quality at modest extra cost. They also propose a two-state finite-state-machine controller that reuses cached predictions and replans only the replaceable suffix when conditions change. Serving experiments on Wan-1.3B, LTX-2.3 and Sol-H3 show substantially lower response latency for large-chunk models.

Critical strengths
  1. By separating chunk size from output commitment, and showing analytically that boundary-only commitment adds an unavoidable ~T/2 wait whatever the generation speed, the paper turns "make chunks smaller" into "make joint generation faster." That reframing could redirect how the community designs interactive video systems.
  2. Commit-One cuts large-chunk response latency severalfold on Wan-1.3B, LTX-2.3 and an accelerated bidirectional model (Sol-H3), without retraining and at similar GPU activity and stall rates. That consistency is the main evidence that the idea works as a general serving layer rather than a model-specific trick.
Critical weaknesses
  1. The output video itself is never evaluated under Commit-One. The serving study reports only latency, GPU activity and stalls. It does not measure visual quality, temporal continuity at splice points, how quickly the new condition visibly takes effect, or how much computation is actually discarded. The authors also admit that retained bidirectional KV may still carry information from the discarded suffix. So it is unclear whether the latency gains come at a hidden cost to quality or compute.
  2. The novelty over prior work is unclear. The serving results cover only 480p, with no variance or confidence intervals, and each quality configuration has a single training run. Meanwhile, related ideas already exist: speculative multi-frame sampling (NFD), replanning of denoised futures (Equilibrium Forcing), and rolling-window joint denoising. The paper neither compares against these directly nor separates its contribution from them, so the core idea reads more as a careful systems reframing than a demonstrated advance.
Questions for authors
  1. Does Commit-One reduce visual quality or temporal continuity at replacement boundaries compared with chunk-commitment baselines, measured for example with VBench or a user study on the served streams?
  2. What fraction of sampled latents was actually discarded in the high-frequency traces, and how much extra DiT compute did that cost?
  3. How variable is the latency across runs, and do the gains still hold at 720p and with more frequent or continuous updates?
  4. How does Commit-One compare directly with related speculative or replanning schemes such as NFD [1] and Equilibrium Forcing [2], and with rolling-window joint denoising [3–5], under the same traces?

[1] Cheng, Xinle, et al. "Playing with transformer at 30+ fps via next-frame diffusion." arXiv preprint arXiv:2506.01380 (2025).
[2] Lillemark, Hansen Jin, et al. "Equilibrium Forcing: Adaptive Video Generation Without Noise Conditioning." arXiv preprint arXiv:2608.14706 (2026).
[3] Kim, Jihwan, et al. "Fifo-diffusion: Generating infinite videos from text without training." Advances in Neural Information Processing Systems 37 (2024): 89834-89868.
[4] Kodaira, Akio, et al. "Streamdit: Real-time streaming text-to-video generation." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026.
[5] Liu, Kunhao, et al. "Rolling forcing: Autoregressive long video diffusion in real time." International Conference on Learning Representations. Vol. 2026. 2026.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.