Virtual Fitting Room: Generating Arbitrarily Long Videos of Virtual Try-On from a Single Image

The technical novelty of the proposed method appears limited, as it primarily combines existing techniques.

• We respectfully disagree with this assessment.
  • Our VFR is, to our knowledge, the first framework designed specifically for long video virtual try-on. Prior work on video try-on (L83, L90; e.g., D&D, ViViD, Fashion-VDM) focuses on short clips, while prior work on long video generation addresses non-try-on domains (L99), such as text/image-to-video generation (e.g., FramePack) or static scene rendering (e.g., StarGen, CVPR’25).
  • Our core contributions include the anchor video mechanism and the training scheme for temporal consistency over long durations with short video data—both of which, to our knowledge, are novel in the context of both try-on and (long) video generation.
  • Notably, prior long-video methods (e.g., FramePack, StarGen, LCT) emphasize local smoothness but do not address global temporal consistency (L104), which is explicitly tackled by our anchor video design, tailored for the try-on setting.
• We further highlight that all existing SOTA methods, including strong commercial systems trained on extensive internal data (e.g., Kling), fail to produce high-quality long try-on videos without inconsistency or degradation. Our method is, to our knowledge, the first to satisfy all key requirements in this space.
• We would welcome clarification on which “existing techniques” the reviewer believes are merely being combined, so we may better address the concern.

I don't think it is appropriate to claim "implicit 3D-aware" generation, which also appears confusing in Figure 4. Any well-trained video generative model inherently exhibits such properties to some degree.

• We appreciate the opportunity to clarify this point. The term “implicit 3D-aware generation” in our paper (Fig. 4(b), the “cloud” portion) does not refer to the general notion of “3D-consistent video generation” or “3D world modeling” that some video generative models may exhibit to a limited extent.
• Specifically, in our context, “implicit 3D-aware generation” refers to the following:
  • Our anchor video, as shown in Fig. 3(b), can be reconstructed into a 3D form and, thus, can be regarded as a pseudo-3D representation (L162).
  • Our generation process is explicitly conditioned on this pseudo-3D anchor video, enabling our model to preserve consistent user and garment appearance across time—what we refer to as implicit 3D-awareness (L164).
  • As other video generative models do not condition on such anchor videos, they do not support the “implicit 3D-aware generation” described in our work.
• More broadly, even leading commercial models do not reliably achieve 3D or temporal consistency, as illustrated by Kling’s performance in Fig. 6. In contrast, our method is the only one, to our knowledge, that enforces such consistency through anchor-based conditioning.

Thanks to the authors for their response. My concerns about novelty have been addressed. The proposed techniques, e.g. anchor video mechanism, demonstrate clear improvements in long video virtual try-on quality by enhancing global temporal consistency, where general approaches like FramePack fall short.

However, I remain unconvinced about the terminology "implicit 3D-aware generation." As it creates confusion with established implicit 3D representations (e.g. Signed Distance Functions (SDFs), Occupancy Functions, and Neural Radiance Fields (NeRFs)) in computer graphics and vision. Using video conditioning as a proxy for 3D information does not constitute true implicit 3D representation. I would suggest reconsidering this terminology to avoid misleading readers who might expect actual implicit 3D modeling, when the method primarily relies on video-based conditioning for spatial awareness.

The motion of garments appears overly constrained by the underlying 3D human model in the anchor videos （might be SMPL). For instance, loose clothing exhibits minimal dynamics or moves as a rigid body, failing to realistically simulate cloth-specific behaviors such as the flow of skirts.

• We clarify that we do not use any underlying 3D human model such as SMPL. The anchor video is a generated virtual try-on sequence at a fixed A-pose and under a rotating camera trajectory (L158), rather than being based on any explicit 3D mesh or rigged model.
• Although the anchor video itself is relatively static, our final generated videos are capable of producing reasonable garment-specific dynamics. Examples include the flow of the black skirt and the checkered skirt (SV3:25), as well as the movement of the yellow trousers (SV4:14).

Several implementation details are unclear. It appears that the anchor video is generated by the proposed VFR model conditioned on a given A-pose. However, Section B.1 of the supplementary material only mentions conditioning on the prefix video and anchor video for DIT. It remains ambiguous how the initial anchor video is produced.

• During inference, the anchor video is generated using our VFR model, conditioned on the garment image, user image, motion prompt, and A-pose template. Since the anchor video is a short (5s), low-FPS video, it is produced via a non-autoregressive diffusion process. This avoids the inconsistency often observed in AR generation and does not require conditioning on another prefix or anchor video.

The paper does not describe the process of fusing a target user with a specified garment to produce the initial try-on appearance, which is crucial for practical applicability.

• We refer the reviewer to our prior work Dress&Dance [8] (Sec. 3) for details on how a target user and a specified garment are fused. This is achieved using CondNets (Sec. 3.1) for multi-modal conditioning, and a garment warm-up training phase (Sec. 3.2) to enhance garment-user fusion capability. Our method adopts this architecture with minimal modifications.
• When accessing the corresponding repository for [8], please use the “Download Repository” button to obtain the full codebase, instead of downloading individual files, to avoid the bugs of anonymous GitHub.

The inference pipeline lacks clarity regarding the generation of the first frame and the initial prefix video. It is important to clarify whether these are produced by a separate model or integrated into the VFR pipeline, and how much they affect overall video quality.

• The first frame does not require separate generation. The initial prefix video is generated using the same VFR model without providing a prefix condition (L135). There is no separate model involved—our entire generation pipeline, including the generation of the first segment and the anchor video, uses a single VFR model with variable conditioning inputs.

The necessity and efficiency of the repeated refinement steps are not well justified. It would be helpful to report how the number of refinement iterations is determined, the computational overhead introduced, and the potential performance impact if the refiner is applied to other datasets such as Dress&Dance.
During the generation process, how much time is spent on the VFR module versus the refiner?

• The necessity of immediate refinement is demonstrated in our ablation study (Suppl. L77, Suppl. Tab. C.1 “Ours NR”, SV4:56), where removing refinement leads to notable degradation over long videos.
• Our refiner is fine-tuned from a partially trained VFR and takes both the prefix and anchor video as input. Therefore, it would not be fair to apply the same refiner to baselines, which lack these conditioning signals. Without these two conditions, the refinement could mitigate some smoothness issues in the baselines, but the fundamental inconsistency would likely persist.
• Our generation uses a fixed 2-step refinement strategy. Since refinement operates on an already generated video, it requires fewer denoising steps than the initial generation. On a single GPU, refinement takes approximately 30–40 minutes, which is about one-third the time required for the main video generation. Altogether, refinement accounts for roughly one-quarter of the total generation time (~2 hours; see L290).

The paper does not contain any quantitative comparisons against baseline methods in the main text. The only quantitative results in the supp. mat. is actually an ablation study. So it is questionable whether this paper is suitable as a submission to NeurIPS? Maybe it will better to submit it to Siggraph or other conferences.

• We clarify that Table C.1 in our supplementary material presents quantitative comparisons against two baselines: D&D+FramePack and D&D+Kling.
• As the benchmark is large, the evaluation of these baselines is resource-intensive. Thus, we focused on the most representative and competitive baselines (e.g., Kling, a commercial product trained on extensive internal data).
• As there are currently no existing methods for long video virtual try-on, these composed baselines and our ablation variants represent the only viable points of comparison.
• We will move Table C.1 to the main paper and include additional baselines in our revision.

The model is trained on a dataset named "Dress&Dance" which is cited as being "Under Review" and linked to an anonymous URL. This makes it unclear whether the strong qualitative results stem primarily from the proposed VFR method or from the quality of the specific, inaccessible training data. Since the dataset is cited separately, the reviewer will assume that this paper has no dataset contribution claim. In this context, it is less persuasive that the proposed method contributes to the final performance rather than the private data. Besides, the anonymous URL is broken so the reviewer cannot download pdf files containing descriptions of the dataset.

• We kindly remind the reviewer that our most powerful baselines—D&D+FramePack and D&D+Kling—use the same Dress&Dance (D&D) dataset as our VFR model. This demonstrates that the private data alone is not sufficient to produce high-quality long try-on videos.
• Public video virtual try-on datasets, such as VVT (256×192) and ViViD (832×624, highly compressed), lack the resolution and fidelity needed for generating minute-scale outputs with detailed garment appearance. As such, the D&D dataset (≥1080p) is the only viable option for our high-resolution task.
• The inaccessibility of the anonymous GitHub link is caused by a known bug in the tdurieux/anonymous_github repository (see GitHub issues #395, #404), and is unrelated to our submission.
  • As a reliable workaround, we kindly ask the reviewer to use the “Download Repository” button on D&D’s anonymous GitHub page to bypass the bug and access the full codebase, paper, and videos.

Since the anchor video in this method is vital, it is also unclear how the proposed method could ensure or improve 3D consistency in its training. This weakness is also related to the second one, where the authors fail to elaborate the details of the training set, therefore it is unclear how they promote the video quality and 3D consistency when they train the model to generate the anchor video.

• Since the anchor video is a short, low-FPS sequence, its generation does not require conditioning on another anchor video or a prefix. Instead, during inference, the anchor video is produced using our VFR via a single-step, non-autoregressive diffusion process. This avoids temporal inconsistency (which only happens in autoregressive generation) and does not rely on prefix conditioning.
• To support anchor video generation during training, we randomly drop both the anchor video and prefix conditions with a certain probability. This simulates the setting of anchor video generation. The dropout is handled by the CondNet conditioning mechanism (D&D Sec. 3.1), which passes an empty sequence of tokens to the transformer blocks when conditions are absent.
  • Please refer to the D&D paper, included in the repository accessible via the “Download Repository” button.

I think this paper's scope is not suitable for NeurIPS, and it will also be better to conduct the experiments on public benchmarks instead of private dataset that is not publicly available, given that this paper has no dataset contribution claim.

• We emphasize that our performance is not attributable to the dataset itself, as both D&D+FramePack and D&D+Kling—trained on the same data—fall short in video quality and consistency. Our ablation studies further highlight the contributions of our novel designs.

The authors did not address the limitations and potential negative societal impact of their work.

• We have outlined the limitations and potential societal impacts of our method in L288–L299, which have also been acknowledged by other reviewers.

Also, FYI, the evaluation of the effectiveness of a method can be diversed. There is no written rule that a NeurIPS paper has to have a metric evaluation.

Therefore, no numerical evaluation itself cannot be justifiable argument for not "suitable as a submission to NeurIPS".

--AC

Thank you to the authors for their detailed rebuttal and to the Area Chair for the follow-up and clarification. I have carefully reconsidered my review in light of the discussion.

My primary concern remains the paper's lack of rigorous and verifiable evaluation. While I acknowledge the AC's point that quantitative metrics are not a strict requirement for a NeurIPS paper, their absence becomes a critical flaw when combined with the use of a private, inaccessible dataset. For example, in 3D asset generation, it is also hard to quantitatively compare results from different methods, but most of them could clearly state their pretrained model (e.g., Stable Diffusion 1.5) and/or the dataset (e.g., Objaverse). It is easier for researchers to compare specific cases qualitatively, where no quantitative metrics in a submission is acceptable.

The authors' argument that their baselines also used the "Dress&Dance" dataset does not fully resolve this. It demonstrates the method's superiority within a closed ecosystem of baselines (D&D+FramePack and D&D+Kling) but prevents the broader research community from independently verifying the results or comparing them against other present or future methods. This fundamentally limits the paper's contribution and impact. The reliance on a private dataset for a newly-defined task, without providing any results on established public benchmarks (even if acknowledged as lower-resolution), makes the claims difficult to situate within the existing literature. For this submission, the result is convincing only in two situations: 1) after the private dataset is officially public as a benchmark for the community, compare this submission with other baseline methods in this benchmark or 2) the authors could show performance comparison with other baseline methods in existing low-resolution benchmarks, e.g., VVT and ViViD.

For these reasons, the work lacks the standard of reproducibility and verifiable evidence expected for a publication at NeurIPS. Therefore, I will maintain my original rating of reject.

Although this paper has made some progress in the try-on field, the overall technical contribution is still limited. This AR video generation way has also been used in other tasks using video generation model, like StarGen [1]. Regarding the video AR generation, the article can also discuss and compare the AR video generation solution in LCT [2], which can use some KV-cache to speed up.

• First, our overall technical contribution is not merely an “AR video generation way.” VFR takes a single image of the user and the garment to produce a high-resolution, smooth, and temporally consistent long try-on video where the user performs the desired motion. We devise novel training schemes that allow our model to be trained using only short video data.

  • This is fundamentally different from StarGen, which focuses on generating videos of a static scene.

• LCT extends full attention mechanisms from individual shots to the scene level by incorporating interleaved 3D positional embeddings and an asynchronous noise strategy.

  • In contrast, our VFR adopts a standard and streamlined conditional diffusion framework with multiple conditioning signals. While VFR currently does not use KV-caching, it could potentially benefit from a similar acceleration strategy (e.g., enabling KV-cache for prefix videos), which we consider a promising direction for future work.
  • We also note that LCT is a concurrent work and has not been published in any peer-reviewed venue.

• Lastly, neither StarGen nor LCT addresses key challenges such as over-time performance degradation, temporal inconsistency, and local smoothness, as can be observed from the videos available on their respective websites. In contrast, our VFR is specifically designed to mitigate these issues, highlighting its distinct and effective technical contributions.

• We will clarify these points further in the final revision.

I am quite confused about how to get the anchor video during inference, which is a 360 video in the paper. But during training, two clips are sampled from one video. Will there be a gap between the anchor video during training and testing if they are too different.

• During inference, the anchor video is generated as a short (5s), low-FPS video try-on task using our own method (L158), performed via a single-step, non-AR diffusion generation. This avoids the temporal inconsistency that only arises in autoregressive generation.
• In training, the model is explicitly optimized to “produce videos that are consistent with the (arbitrarily) given anchor video” (L197). More specifically, because the two sampled clips in training may differ substantially due to occlusions and motion dynamics, the model naturally learns to handle diverse anchor videos, including the A-pose anchor used at inference time.

The refinement process is not elegant and needs to be performed multiple times, which consumes computing power. Is there a more elegant and end-to-end approach?

• The refinement process plays a key role in preventing performance degradation, as demonstrated in our ablation study (Suppl. L77, Suppl. Tab.C.1 “Ours NR”, SV4:56).
• As the refinement process operates on an already generated video and does not need to generate from scratch, it is substantially easier than the original generation. This allows us to distill it into an efficient one-step denoising model with comparable performance. We leave this optimization and the end-to-end approach as interesting directions for future work.

Dear Reviewer VQN7:

Can you share your feedback to authors' responses?

Best, AC

Thank you for your feedback. We respectfully disagree with your assessment for two reasons:

1. Our work is specifically focused on virtual try-on, a widely accepted topic at NeurIPS, as evidenced by recent papers like PM-Jewelry (NeurIPS'24 Workshop), AnyFit (NeurIPS'24), Greatness in Simplicity (NeurIPS'23), etc. Our VFR is the first work to demonstrate a high-resolution, smooth, and temporally consistent long try-on video where the user performs the desired motion. We compare our approach to state-of-the-art methods, including commercial products like Kling 2.0 (trained on massive internal data) and concurrent academic works like FramePack, to highlight its significant advantages.

2. While the suggested references, such as StarGen and LCT, are recent and important works, their capabilities are more limited compared to ours, as we've detailed in our rebuttal.

We believe that discovering a simple solution and demonstrating its effectiveness on a novel and practical use case is a substantial contribution to the NeurIPS community.