- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper presents ProxyUp, a training-free video generation method that uses motion proxy videos as input to generate realistic scenes with coherent interactions. The method mainly combines two existing techniques: (1) region-wise latent noising (RLN), which applies SDEdit-style latent initialization to the masked foreground of a source video while replacing the background with noise, and (2) stochastic flow relaxation (SFR), which directly adopts the Predict-and-Perturb (P&P) strategy from Self-Refining Video Sampling. The resulting latent is then denoised using a pretrained Wan2.2 model. Experiments show improvements over several video editing and motion transfer baselines.
Critical strengths
- The idea of using incomplete motion proxies to generate realistic video interactions is interesting.
- The qualitative results are generally promising, especially for examples involving object interactions.
Critical weaknesses
- Limited technical novelty. This is my main concern. The proposed framework seems to be a straightforward combination of existing video editing and sampling techniques, with the main difference being the use of motion proxies as source videos. RLN is essentially masked inversion and noise initialization, while SFR directly uses the existing P&P strategy. I do not see a substantial technical contribution beyond adapting these methods to a different type of input video.
- Insufficient experimental comparisons. The paper compares against several existing video editing methods, but does not include some natural baselines, such as SDEdit combined with P&P refinement. Such comparisons are important because the proposed pipeline is largely built from these existing components. The current ablations only show that RLN and SFR are useful, but do not demonstrate that their combination is better than simpler alternatives under the same computational budget.
- Unclear practical applicability. The method assumes that a suitable motion proxy is already available. The evaluation uses simulated videos and existing real-world recordings, but it is unclear how users would obtain these proxies in practical applications. In some cases, creating a proxy with the desired motion may itself require considerable effort.
- High inference cost. Despite being training-free, the method requires 32 inversion steps, 15 refinement iterations, and 32 sampling steps. This is considerably more expensive than standard video generation in terms of sampling operations. However, the paper does not report inference time or computational overhead, making it difficult to assess the practicality of the approach.
- Limited evaluation of motion preservation. The primary motion-related metric (MR) evaluates whether the generated video completes the events described in the prompt, rather than measuring how closely its motion follows the input proxy. Although the paper reports DPS in the appendix, it does not compare this metric against the baselines. As a result, it remains unclear how effectively the method actually preserves proxy motion.
Questions for authors
- What is the key methodological contribution beyond applying masked latent initialization and existing P&P refinement to motion proxy videos?
- Could the authors compare against SDEdit with P&P refinement and other straightforward combinations of existing techniques under matched inference budgets?
- How are motion proxies expected to be obtained in practical applications, especially when no suitable simulation or reference video is available?
- What is the actual inference time of ProxyUp compared with Wan2.2 and the other baselines?
- Could the authors provide direct quantitative comparisons of motion preservation using DPS or trajectory-based metrics?
Dear Reviewers,
Thank you for your time and thoughtful reviews. Please be reminded that the author-reviewer discussion period ends on November 18.
If you haven’t already, please review the authors’ rebuttals and actively participate in the discussion. If you have any remaining questions or concerns, please raise them before the deadline so the authors have sufficient time to respond.
Thank you again for your valuable contributions to ICLR 2027.
Best regards, Area Chair