- Rating
- 3(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
Built upon a 14B-parameter video Diffusion Transformer (DiT), LiveAnimate adopts a two-stage training pipeline. In the first stage, a pretrained bidirectional model is converted into a block-causal autoregressive generator via Reference-Anchored Teacher-Forcing Adaptation. In the second stage, Block-wise Self-Forcing Distillation compresses the sampling steps down to 3, drastically reducing inference latency. This work proposes Pose-Retrieval Sink Attention (PR-Sink), a fixed-length KV cache composed of a Static Sink, a pose-retrieved Dynamic Sink, and a three-slot Rolling Window, which maintains long-form appearance consistency with constant memory overhead. Combined with optimizations including sequence parallelism and operator fusion, it enables real-time streaming generation. On a three-minute benchmark, this method delivers stable visual quality and identity preservation, with a DiT throughput of 19.63 FPS and an end-to-end throughput of 11.2 FPS.
Critical strengths
- Video demos and evaluation metrics overall demonstrate that this work effectively mitigates cumulative error, and well preserves overall high-frequency details and human identity details over a long time horizon.
- The dynamic sink mechanism designed based on pose similarity is intriguing, and I believe the research community can draw on this idea to optimize pose-conditioned video generation.
- This paper is well organized and easy to understand.
Critical weaknesses
- The methodological innovation is relatively limited. The two-stage training paradigm of teacher forcing + self forcing is a prevailing framework in the field of streaming real-time video generation, and numerous works such as Self-Forcing, Self-Forcing++, LiveAvatar, and StreamAvatar adopt a similar approach. Furthermore, the reference-anchored method introduced in Section 3.2 is actually commonly used in streaming image-to-video tasks, so the innovation of simply integrating the reference as a sink is rather incremental.
- The video demos contain some imperfections. I observe that nearly all videos under the "Shorter rollouts" section exhibit prominent shifts in color and texture details between the first frame and subsequent frames. I cannot confirm whether this indicates actual changes in human identity, as the supplementary materials do not provide reference images of the subjects to assess the magnitude of identity variation.
- Several baselines are missing. Works including SteadyDancer (Zhang et al., 2025) should be included for comparison, since this work also introduces a first-frame preservation mechanism that bears similarities to the attention sink mechanism proposed in this paper.
Questions for authors
- Further elaborate on the actual novelties of this work. Revise the paper structure if necessary, with emphasis on Section 3.4 and condensed presentation of Sections 3.2 and 3.3.
- Explain the first-frame inconsistency issue in the video demos.
- Add more comparative baselines.
Thanks for your response, most of my concerns are well addressed. I decided to raise my score.