{"id":"06b98257-e6b5-4f94-8652-484eacbf96dc","arxiv_id":"2509.04450","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"VFR generates minute-long virtual try-on videos by auto-regressively chaining short diffusion-generated segments that are kept consistent with a 360-degree anchor video of the user.","lead":"This paper introduces VFR, a video generation model that creates arbitrarily long virtual try-on videos from a single user image and garment photo. It stitches together short generated segments using a 360-degree anchor video of the person to keep clothing appearance consistent across the whole clip.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The anchor video is model-generated and unvalidated, yet every segment is conditioned on it; Table 1 does not show the full model beating its no-anchor baseline on GPT metrics.","rationale":"The reader's CONDITIONAL verdict is justified and I do not think it needs to be changed. The weakest point is indeed the anchor-video assumption: the anchor is the sole appearance prior for the entire long video, but it is generated by the same unverifiable model and is never checked against the true user or garment. I add the observation that this concern is visible in the paper's own numbers: on the 90s and some 50s tasks, removing the anchor (D&D) or prefix (NP) yields higher GPT-Overall scores than the full model, and no variance information is given. That does not prove the method fails, but it means the claimed benefit of anchor conditioning is currently supported only by selected qualitative examples. My proposed multi-view anchor-fidelity test would directly settle whether a self-consistent but inaccurate anchor is being produced. Since the paper is explicitly a technical preview and the central idea is plausible, a conditional verdict remains appropriate; the paper should be accepted only if the authors provide code/data and the above evidence.","tokens_in":12759,"tokens_out":8761,"duration_ms":80380,"concrete_test":"Using a held-out subject with ground-truth multi-view captures, generate the VFR anchor video from a single front-view image and compare every anchor frame to its corresponding ground-truth view with LPIPS, garment-region CLIP similarity, and a face-identity metric; also regenerate the anchor twice from the same input to check whether back/side garment texture is stable. If anchor-to-ground-truth error is large or the two anchors disagree on garment appearance, the anchor-based global consistency claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"VFR's central mechanism (Sec. 3, Fig. 3) is to generate a 5s 360-degree A-pose anchor video from the same single user image and then condition every later segment on that anchor. For the central claim to hold, the anchor must faithfully and completely capture the user's whole-body appearance in the target garment. This is not established. The anchor is produced by the same diffusion model that must solve the underlying try-on problem; no 3D supervision, multi-view training, or external constraint is described. The anchor can be self-consistent while hallucinating back-of-garment texture, body shape, or occluded details, and because every segment is pulled toward it, such an error propagates through the entire generated video. The prefix condition only ensures local smoothness, not correctness. The reported 3D emergence (Fig. 3b) is a NeRF reconstruction of the model's own generated anchor, so it does not validate geometric fidelity. The quantitative experiments do not fill this gap: in Table 1, 90s Hand-Body, the full model's GPT-Overall (82.62) is below the no-anchor D&D variant (86.15); in 4-Medium it is 81.93 vs 83.61; in 4-Hard it is 79.05 vs NP's 79.60. No error bars or multiple seeds are reported, so the benefit of the anchor, and with it the evidence for global consistency, is unsubstantiated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VFR (Virtual Fitting Room), a video-generation framework for virtual try-on from a single user image, a garment image, and a reference motion video. The method generates long videos auto-regressively, segment by segment, using a prefix video condition for local smoothness and an \"anchor video\"—a model-generated 5-second 360-degree A-pose video of the user in the target garment—for global temporal consistency. The authors claim this enables arbitrarily long, minute-scale, high-resolution (720×1152) try-on videos. They introduce a four-part evaluation protocol (360° garment consistency, 360° human+garment consistency, 90s hand-body interaction, and arbitrary poses) and report quantitative comparisons in Table 1 against baselines and ablations. They also claim that implicit 3D consistency emerges, allowing free-viewpoint rendering via NeRF reconstruction of the anchor video.","tokens_in":13103,"tokens_out":4039,"duration_ms":35275,"significance":"If validated, the paper would address an important gap: prior virtual try-on methods are limited to images or short (5–10s) videos, whereas VFR targets arbitrarily long videos with both local smoothness and global appearance consistency. The anchor-video idea is conceptually simple and potentially influential for long video generation beyond try-on. The proposed evaluation protocol is a useful step toward standardizing long try-on video assessment. The paper is clearly written and includes compelling qualitative examples. However, the central empirical claim is not supported by the reported numbers: the full model does not consistently beat its own ablations on the GPT metrics, and no uncertainty estimates are provided. The anchor video, which is load-bearing for the method, is generated by the same model and is not independently validated. The dependence on an anonymous, unpublished base model (Dress&Dance) further undermines reproducibility. These issues currently prevent acceptance.","major_comments":[{"comment":"The statement in §4.2 that \"our full VFR consistently outperforms all the baselines and variants\" is contradicted by Table 1 on multiple GPT Overall scores. In the 90s Hand-Body task, the No-Prefix (NP) variant scores 84.62 and the Dress&Dance (D&D) variant scores 86.15, both higher than full VFR's 82.62. In the 4-Medium task, D&D scores 83.61 versus VFR's 81.93; in the 4-Hard task, D&D scores 80.39 and NP scores 79.60, both above VFR's 79.05. The claim of consistent superiority also conflates VBench metrics, where VFR does lead, with GPT metrics, where it does not. Moreover, no error bars, number of seeds, or statistical significance tests are reported, so even the VBench differences may be within noise. This is a load-bearing issue because the paper's central claim is empirical: the anchor and prefix conditioning are supposed to improve global consistency and try-on quality. The authors should either revise the claim, report multiple runs with variances, and/or provide a metric-by-metric statistical analysis that accounts for the observed contradictions.","section":"§4.2 and Table 1"},{"comment":"The anchor video is the core mechanism for global consistency, but the paper provides no independent validation that a model-generated 5-second 360-degree A-pose video faithfully captures the user's whole-body appearance. The anchor is produced by the same diffusion model that must solve the underlying try-on problem, so it can be self-consistent while hallucinating back-of-garment texture, body shape, or occluded regions. The NeRF reconstruction in Figure 3(b) is built from the model's own generated anchor, so it cannot validate geometric fidelity. To support the central claim, the authors should measure anchor fidelity against ground-truth multi-view data (e.g., real 360° captures of subjects), or at minimum show that the final video quality is sensitive to anchor quality, for instance by corrupting the anchor and demonstrating degradation. Without such evidence, the anchor-based consistency mechanism remains a plausible but unsubstantiated assumption.","section":"§3 and Figure 3"},{"comment":"The method is built on Dress&Dance [8], which is cited as \"Anonymous Author(s)\" and described as \"Under Review, January 2025.\" According to §4.1, \"Our VFR model is built on Dress&Dance [8] with the addition of 'prefix video' and 'anchor video' CondNets,\" and the baselines also rely on Dress&Dance for initial frames and try-on quality. This makes the contribution non-self-contained: the core try-on capability is inherited from an unpublished, unavailable model, so a reader cannot reproduce the results or separate VFR's contribution from that of Dress&Dance. The authors should disclose the relation between VFR and Dress&Dance (including whether Dress&Dance is their own prior work), provide sufficient architecture and training details for the base model, or release the checkpoint so that the incremental contribution is testable.","section":"§4.1 and Reference [8]"}],"minor_comments":[{"comment":"The 360° Garment Consistency (5s) row shows only the full VFR, with no baseline or ablation comparison. Adding at least the D&D and NP variants there would make the table self-consistent with the other tasks and would help assess the anchor-consistency claim.","section":"Table 1, row 1"},{"comment":"The abstract states the output is \"720×1152 resolution at 8 FPS and can be further refined to 24 FPS,\" while §1 says \"720×1152 resolution at 24FPS\" and the contribution says \"720×1152 resolution at 24FPS.\" Please harmonize these numbers and state clearly which resolutions/framerates were used in the experiments.","section":"Abstract and §1"},{"comment":"Training details are sparse: the model is trained for 10,000 iterations, but no batch size, dataset size, number of GPUs, or learning rate is reported. For a method that depends on training a video diffusion model, these details are important for reproducibility.","section":"§4.1"},{"comment":"The subsection heading \"Hand-Body Interaction Faithfullness\" contains a typo; it should be \"Faithfulness.\"","section":"§4.2"},{"comment":"The Discussion states that generating a 30s video takes 1–2 hours, which is substantial and may undercut the practical framing of \"arbitrarily long\" videos in the introduction. Adding inference-time analysis and a discussion of scalability would strengthen the paper.","section":"Discussions"}],"recommendation":"major_revision","confidential_remarks":"The citation to Dress&Dance [8] as anonymous under-review work is a likely self-citation in disguise, which is a novelty-disclosure concern for a double-blind venue. The paper is labeled a \"Technical Preview,\" and the experiments are preliminary: Table 1 contradicts the consistency claim, and there are no statistical tests. I recommend major revision rather than reject because the core idea is plausible and the qualitative results are promising, but the authors must supply stronger evidence and a revised quantitative narrative."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me save you a full read: this paper is a plausible engineering advance that likely does what it says—generates minute-scale virtual try-on videos from a single image—but the quantitative section is not as clean as the text claims. In Table 1, the full VFR model loses to its own No-Prefix and Dress&Dance variants on GPT Overall in the 90s Hand-Body and Arbitrary-Poses tasks (e.g., 82.62 vs 84.62 and 86.15; 79.05 vs 79.60 and 80.39). The paper's phrase \"consistently outperforms\" is not supported by its own headline metric.\n\nWhat is genuinely new: the anchor-video idea. Generating a short 360-degree A-pose video of the user in the target garment, then conditioning every later segment on that anchor, is a sensible way to enforce global appearance consistency without training on long videos. The evaluation protocol—four difficulty levels from a stationary A-pose to 90-second hand-body interaction—is a useful contribution in itself, and the qualitative examples do show stable garment textures and preserved accessories over durations that prior try-on methods (Fashion-VDM, Dress&Dance at ~5-10s) did not reach.\n\nThe soft spots are real but not fatal. The anchor is itself generated by the same diffusion model, so it can hallucinate back-of-garment or occluded details, and the NeRF reconstruction in Fig. 3 does not validate geometric fidelity—it only shows that the anchor is self-consistent. The quantitative evaluation has no error bars, no multiple seeds, and no released code; the base model is cited as anonymous and under review. These are standard problems for a technical preview, but they matter because the paper's central claims of \"consistently\" beating variants and \"3D emerging for free\" rest on that evidence.\n\nWho this is for: anyone working on video try-on, long video generation, or human appearance consistency will want to read it and probably cite it. The method deserves a serious referee: the core feasibility result is likely correct, and the anchor mechanism is worth testing independently. What it needs in review is a corrected narrative around the ablation table, error bars or at least a clear statement of variance, and validation of the anchor (e.g., against multi-view captures or a real spinning video). I'd accept it for review with the expectation of major revisions, not desk-reject.","headline":"Plausible engineering advance in long video try-on, but the paper's own ablation table contradicts its 'consistently outperforms' claim; worth serious review with expected major revisions.","tokens_in":13621,"tokens_out":3690,"would_cite":true,"duration_ms":29437,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"VFR turns a single photo and one garment into a minute-long try-on video, sustained by a 360-degree anchor video that keeps the garment's appearance consistent across all generated segments.","keywords":["virtual try-on","long video generation","autoregressive generation","temporal consistency","anchor video conditioning","video diffusion models","free-viewpoint rendering","single-image animation"],"falsifier":"Generate a long try-on video in which the person rotates fully, so the back of the garment is visible both early and late in the video; if the back pattern, logo, or texture differs between distant timestamps despite the anchor conditioning, the central claim of global temporal consistency fails. This could be quantified by measuring the similarity of back-of-garment crops across segments with a vision-language or pixel-level consistency metric.","tokens_in":12569,"feed_emoji":"👗","tokens_out":4107,"duration_ms":39024,"temperature":0.7,"pith_summary":"The paper presents VFR, a video generative model that takes a single image of a user, an image of a garment, and a reference motion video, and produces a try-on video of arbitrary length. It splits generation into short auto-regressive segments, using a prefix video condition to keep adjacent segments smooth and a generated 360-degree 'anchor video' of the user in the garment to keep global appearance consistent. If correct, this would be the first method to produce minute-scale, temporally coherent virtual try-on videos from a single photo, letting shoppers see a garment through long, varied motions. The paper also reports that this temporal consistency implicitly yields 3D consistency, enabling free-viewpoint rendering of the draped user.","feed_headline":"One photo plus one garment becomes a minute-long try-on video","feed_subtitle":"VFR chains short clips around a 360-degree anchor clip, keeping the garment's look stable across all frames.","key_machinery":"The central mechanism is the anchor video combined with prefix conditioning. The anchor video is a short 360-degree clip of the user in a simple 'A' pose wearing the target garment, generated by the same model; it supplies a stable, whole-body appearance reference for every later segment. The prefix video is the overlapping tail of the previously generated segment, passed as a condition to ensure the next segment continues smoothly without sudden jumps or morphing. These two conditions are implemented as additional CondNets on top of the base Dress&Dance [8] model, and an immediate refiner is trained to clean up each segment's output. Together they are what let the generator produce arbitrarily long videos without training on long videos.","core_discovery":"The paper claims that long virtual try-on generation can be decomposed into a two-stage process: first generate a short 360-degree video of the user in an 'A' pose wearing the target garment, then generate each video segment conditioned on both the overlapping tail of the previous segment and this anchor video. The anchor acts as a global memory of the user's whole-body appearance, so segments consistent with the anchor are also consistent with each other. Experiments show 30-second and 90-second try-on videos at 720x1152 resolution, with base 8 FPS output refined to 24 FPS, exhibiting both local smoothness and long-range garment and identity preservation. The paper further observes that the same temporally consistent outputs can be reconstructed into a 3D human mesh, indicating that 3D structure emerges without explicit 3D supervision.","pith_inferences":["Editorial extension: the anchor-video idea is not obviously limited to clothing; the same pattern of generating a 360-degree identity anchor could be tested for long-form face animation, character consistency in narrative video, or other tasks where a single identity must persist across minutes.","Editorial extension: one can probe the limits of the anchor by varying its length and pose coverage; a natural test is whether a shorter anchor or one without full 360-degree coverage causes garment details on the back to drift in longer videos.","Editorial extension: the reported implicit 3D consistency suggests a testable hypothesis that temporal-consistency training alone can substitute for explicit 3D supervision in other single-image-to-video scenarios, which could be examined by ablating the anchor and measuring 3D reconstruction quality.","Editorial extension: the 8 FPS base plus a 24 FPS refiner hints at a two-stage efficiency design; a direct extension would be applying the same refine step to increase spatial resolution or to extend segment length, trading quality against the stated 1-2 hour generation time."],"forward_implications":["A shopper could see a garment on themselves across motions lasting a minute or more, not just a few seconds, which is closer to real fitting-room exploration.","Long videos can be produced with only short-video training data, since the model learns to generate consistent segment-by-segment rather than needing long sequences as ground truth.","The reported emergent 3D consistency would allow free-viewpoint rendering of the user in the target garment, effectively turning a single photo into a navigable 3D try-on.","The paper's four-tier evaluation protocol provides a concrete way to measure long video try-on quality, covering garment fidelity, human appearance, hand-body interaction, and arbitrary poses.","If the approach scales, it could shift online clothing retail from static images to interactive long-form video previews, though the paper notes current generation takes 1-2 hours for a 30-second video."],"supporting_citations":[{"why":"Base virtual try-on video model that VFR extends with anchor and prefix CondNets, and the source of training data and initial checkpoints.","marker":"[8]"},{"why":"FramePack, the concurrent long-video generation baseline that VFR compares against and whose memory-based conditioning is shown to be insufficient for try-on consistency.","marker":"[14]"},{"why":"Kling Video 2.0, used as a strong image-to-video baseline in the repeated image-to-video comparison.","marker":"[81]"},{"why":"VBench supplies the subject consistency, background consistency, and motion smoothness metrics used to quantify temporal consistency and smoothness.","marker":"[84]"},{"why":"GPT-based evaluation scores are used to assess overall try-on quality, user appearance, motion, and visual quality.","marker":"[83]"},{"why":"DiffEdit is used in the 'No Prefix' ablation to outpaint video with the prefix, isolating the effect of prefix conditioning.","marker":"[82]"},{"why":"NeRFStudio viewer is used to reconstruct and display the 3D mesh that emerges from the generated anchor video.","marker":"[15]"},{"why":"Fashion-VDM is cited as a prior video try-on method limited to short videos, motivating the need for arbitrarily long generation.","marker":"[7]"}],"fun_headline_variants":["Turn one photo into an endless try-on video","Single image to arbitrarily long try-on video","From one snapshot to a video of any length","Try-on video: generate beyond a minute from one image"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a short 360-degree 'A-pose' video of the user, generated by the same model, captures enough of the person's whole-body appearance that conditioning on it keeps the garment and identity stable even when later poses deviate far from that anchor pose.","fun_headline_variants_meta":{"raw":{"variants":["Turn one photo into an endless try-on video","Single image to arbitrarily long try-on video","From one snapshot to a video of any length","Try-on video: generate beyond a minute from one image"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000169,"raw_usage":{"total_tokens":1225,"prompt_tokens":867,"completion_tokens":358,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":298}},"tokens_in":483,"tokens_out":358,"duration_ms":3988,"temperature":1.0,"reasoning_tokens":298,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:28:30.849993+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a long try-on video in which the person rotates fully, so the back of the garment is visible both early and late in the video; if the back pattern, logo, or texture differs between distant timestamps despite the anchor conditioning, the central claim of global temporal consistency fails. This could be quantified by measuring the similarity of back-of-garment crops across segments with a vision-language or pixel-level consistency metric.","supporting_citations":[{"cited_title":"Packing input frame contexts in next-frame prediction models for video generation","cited_arxiv_id":null,"evidence_quote":"FramePack, the concurrent long-video generation baseline that VFR compares against and whose memory-based conditioning is shown to be insufficient for try-on consistency."},{"cited_title":"Kling ai: Next-generation ai creative studio, 2024","cited_arxiv_id":null,"evidence_quote":"Kling Video 2.0, used as a strong image-to-video baseline in the repeated image-to-video comparison."},{"cited_title":"VBench: Comprehensive benchmark suite for video generative models","cited_arxiv_id":null,"evidence_quote":"VBench supplies the subject consistency, background consistency, and motion smoothness metrics used to quantify temporal consistency and smoothness."},{"cited_title":"Nerfstudio: A modular framework for neural radiance field development","cited_arxiv_id":null,"evidence_quote":"NeRFStudio viewer is used to reconstruct and display the 3D mesh that emerges from the generated anchor video."}],"review_version":2}