{"id":"a63980bb-7cf8-47fa-ac16-c3b7822f66b9","arxiv_id":"2501.11776","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"EfficientVITON fine-tunes Stable Diffusion with zero cross-attention and non-uniform timesteps to produce virtual try-on images faster, reporting FID 8.433 and LPIPS 0.0762 on VITON-HD.","lead":"EfficientVITON combines Stable Diffusion with a spatial encoder and zero cross-attention blocks to perform virtual try-on, using a non-uniform timestep schedule to cut inference time. It reports competitive FID and LPIPS scores on VITON-HD, but the closest diffusion baseline, StableVITON, is not compared.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported SOTA numbers come from a RePaint configuration whose inference time is not reported, while the 72.4% speedup applies to the base model; no single configuration is shown to achieve both.","rationale":"The reader correctly identified the missing uniform-timestep ablation as a load-bearing gap for the efficiency claim. I agree that such a comparison is necessary to attribute the speedup to the proposed schedule. However, the more decisive issue is that the paper's two headline numbers come from different configurations: the SOTA-quality numbers are obtained with RePaint, whose inference cost is unreported, while the 72.4% speedup is reported only for the base model without RePaint. This makes the joint claim of 'state-of-the-art results' and '72.4% faster without quality loss' internally unsupported, because no single configuration is demonstrated to have both properties. The efficient base configuration is not clearly SOTA on both metrics (LPIPS 0.0842 vs. DCI-VTON 0.0804, with no error bars), and the RePaint configuration may be far slower than the reported 16s. The paper contains no fundamental mathematical error and does present a plausible incremental combination of StableVITON and Fast-DDPM, but the central quantitative claim requires the timing measurement and an ablation before acceptance. A conditional verdict remains appropriate, so I recommend no change to the reader's verdict.","tokens_in":7851,"tokens_out":5637,"duration_ms":57818,"concrete_test":"Measure end-to-end inference time and number of denoising steps for the exact 'Ours (RePaint)' pipeline used for Table I, and compare with the 16s and 58s entries in Table II. If the RePaint runtime substantially exceeds 16s, then no single configuration delivers both the SOTA metrics and the claimed 72.4% speedup, and the central claim must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim bundles two results that are never demonstrated in the same configuration. Table I's best metrics (FID 8.433, LPIPS 0.0762) are for 'Ours (RePaint)', which uses the RePaint refinement described in Sec. III-D. RePaint is a resampling-based inpainting procedure that typically requires many additional denoising rounds, so its wall-clock cost is not captured by the 16s figure in Table II. The paper reports no inference time, step count, or compute for the RePaint variant. Meanwhile, the efficient 'Ours' row (the 16s configuration) has LPIPS 0.0842, which is worse than DCI-VTON's 0.0804, and no error bars are given, so it is not clearly state-of-the-art. Additionally, the paper offers no ablation comparing the non-uniform timestep schedule against a uniform schedule at the same reduced step count, so the 72.4% inference reduction cannot be attributed to the proposed schedule rather than to a simple reduction in the number of denoising steps. Together these gaps mean the paper's headline—state-of-the-art quality plus a large speedup with no quality loss—is currently unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"EfficientVITON adapts Stable Diffusion for image-based virtual try-on by combining a spatial encoder and zero cross-attention blocks with a pre-trained VAE/U-Net, using two-stage fine-tuning with an attention total variation loss and a non-uniform timestep sampling schedule borrowed from Fast-DDPM. On VITON-HD, the authors report FID 8.703/8.433 and LPIPS 0.0842/0.0762 for the base and RePaint-augmented variants, and claim 72.4% inference and 45.3% training time reductions. The central claim is that these gains are achieved without sacrificing visual quality, yielding state-of-the-art results. The manuscript provides qualitative examples and compares against several published baselines, but the experimental evidence does not currently support the headline claim.","tokens_in":8104,"tokens_out":6080,"duration_ms":59587,"significance":"If fully supported, a diffusion-based virtual try-on model that is simultaneously faster and better than existing baselines would be practically valuable, particularly for e-commerce and real-time applications, and the VITON-HD benchmark makes the quantitative claim measurable. The paper is also transparent about the provenance of its main components: it attributes the zero cross-attention design to StableVITON [3] and the non-uniform timestep schedule to Fast-DDPM [23]. However, the most relevant baseline is omitted from the main table, the best quality numbers are obtained in a configuration different from the one used for the speedup numbers, and no ablation isolates the proposed schedule or loss. These gaps prevent the paper from establishing its contribution as stated. The authors also do not provide code, checkpoints, or a full description of the timestep schedule, which further limits reproducibility.","major_comments":[{"comment":"The efficiency and quality claims are never demonstrated in a single configuration. The 16 s inference time and 72.4% reduction apply to the base 'Ours' model, while the best FID/LPIPS values in Table I belong to 'Ours (RePaint)', a resampling-based refinement whose wall-clock time and step count are not reported. RePaint is not a free lunch; its computational cost must be quantified before the paper can claim simultaneous state-of-the-art quality and near-real-time speed.","section":"§IV-C, Table II vs. §III-D"},{"comment":"The state-of-the-art claim is not assessable because StableVITON [3], the closest baseline and the source of the architecture's spatial encoder and zero cross-attention blocks, is absent from Table I. In addition, no error bars, evaluation protocol, or number of FID/LPIPS samples are reported, so the small differences between Ours and DCI-VTON are not shown to be significant.","section":"§IV-B, Table I"},{"comment":"No ablation isolates the non-uniform timestep schedule from the reduction in total step count. Since the schedule is adopted from Fast-DDPM [23], the paper should compare, at the same number of denoising steps, the non-uniform schedule against the uniform schedule, and should also report the quality metrics for the 'Before' configuration whose timings appear in Table II. Without this, the 72.4% inference improvement cannot be attributed to the proposed schedule rather than to a simple decrease in the number of steps.","section":"§III-D"},{"comment":"The 'Ours (RePaint)' row applies an additional refinement procedure only for the proposed method; the baseline rows are not given this paired-evaluation treatment. This is an asymmetric comparison, and for LPIPS it is material: the base 'Ours' value (0.0842) is worse than DCI-VTON (0.0804), so the reported LPIPS advantage comes from RePaint rather than from the proposed architecture.","section":"§IV-B, Table I"},{"comment":"The contribution of the attention total variation loss and the two-stage fine-tuning is not tested by any ablation. Since LATV is an added term with a free weighting coefficient λATV, the paper should quantify how FID/LPIPS and training time change when this loss and the second training stage are removed; otherwise these components cannot be credited as contributions.","section":"§III-D, Eq. (3)"},{"comment":"The non-uniform timestep schedule is not specified beyond a conceptual description. No formula, list of timesteps, or step-count hyperparameter is given, and the PLMS/RePaint variants are not tied to the reported timings. This prevents independent reproduction of the efficiency result and should be corrected.","section":"§III-D"}],"minor_comments":[{"comment":"The symbols F and M in L_ATV = ||∇(F M)||_1 are undefined; the reader cannot tell which tensors are being regularized.","section":"§III-D, Eq. (2)"},{"comment":"There is a broken sentence in the diffusion-models paragraph: 'They have proven highly effective at managing complex textur [9] and Stable-VITON [3] are diffusion-based models...' This should be rewritten.","section":"§II, Related Work"},{"comment":"Table II is ambiguous: the rows and columns do not clearly distinguish the training-time values from the inference-time values for the 'Before' and 'After' configurations, so the reader has to guess which number is which.","section":"§IV-C, Table II"},{"comment":"The cited papers do not match the described systems: reference [15] is titled 'Cloth Interactive Transformer for Virtual Try-On', not 'Text2Cloth', and reference [16] is titled 'Towards Scalable Unpaired Virtual Try-On via Patch-Routed Spatially-Adaptive GAN', not 'Sketch2TryOn'.","section":"References [15] and [16]"},{"comment":"The qualitative results in Fig. 13 do not state which inference configuration (base PLMS or base PLMS plus RePaint) produced the images; this should be reported for reproducibility and fair comparison.","section":"§IV-A, Fig. 13"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is best viewed as a workshop-style report. Its core ideas are largely assembled from StableVITON [3] and Fast-DDPM [23], and the experiments as presented do not isolate or verify the claimed novelty. The main fixable issues are the missing StableVITON baseline, the missing same-configuration quality/efficiency comparison, the missing schedule ablation, and the asymmetric use of RePaint. If those experiments are added and the claims are appropriately narrowed, the paper could become a solid incremental contribution; in its current form, the headline claim is unsupported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a StableVITON-style baseline with Fast-DDPM's non-uniform timestep schedule, plus an attention total-variation loss. Credit where it's due: the authors cite their sources honestly, the preprocessing pipeline is standard, and the quantitative table includes direct FID/LPIPS measurements on VITON-HD. That is more than many drafts do.\n\nThe problem is that the central claim bundles two results that never appear in the same configuration. Table I's best numbers (FID 8.433, LPIPS 0.0762) are for \"Ours (RePaint)\". RePaint is an inpainting-style refinement that typically costs many additional denoising rounds, and the paper reports no inference time, step count, or compute for it. The 16s figure and the 72.4% speedup belong to the \"Ours\" row, whose LPIPS 0.0842 is actually worse than DCI-VTON's 0.0804. So no single configuration is shown to deliver both SOTA-quality and the speedup. That gap is load-bearing, not cosmetic.\n\nThere are also standard reproducibility issues: StableVITON is missing from the comparison table, there are no error bars, and there is no ablation of the non-uniform schedule against a uniform schedule at the same reduced step count. The attention TV loss and the two-stage training are described in a few sentences, with no ablations, so their contribution is hard to assess. Table II compares \"before/after\" on what looks like a single run, with no settings given.\n\nNone of this suggests fabrication or a corrupt method. The efficiency idea is plausible and worth testing. But the paper overclaims: \"state-of-the-art results without quality loss\" is simply not demonstrated by the evidence presented.\n\nWho is this for? Someone tracking efficiency tricks for diffusion-based try-on might skim it, but it is not a reliable reference for the speedup claim. If the authors add a proper comparison against StableVITON, report the RePaint cost, and run the uniform-vs-non-uniform ablation, it could become a solid workshop paper. As is, I would not send it to a strong venue, and the missing comparisons are significant enough that I would not encourage a serious editor to spend referee time on the current version.","headline":"The headline claim—SOTA quality plus a 72.4% speedup—is unsupported because the best metrics come from a RePaint configuration whose inference time is never reported, while the speedup applies only to the base model.","tokens_in":8634,"tokens_out":2694,"would_cite":false,"duration_ms":26313,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"EfficientVITON claims a diffusion-based virtual try-on system that beats prior methods on VITON-HD while cutting inference time by 72.4% and training time by 45.3%.","keywords":["virtual try-on","diffusion models","Stable Diffusion","non-uniform timestep sampling","zero cross-attention","spatial encoder","VITON-HD","image-to-image translation"],"falsifier":"Run EfficientVITON at the same reduced step count with a uniform timestep schedule, keeping all other components fixed, and measure FID and LPIPS on VITON-HD; if uniform sampling matches or beats the non-uniform row, the schedule is not the mechanism behind the efficiency claim. Also rerun the non-uniform model without the RePaint refinement pass to check whether the headline FID of 8.433 is carried by the extra refinement rather than by the model itself.","tokens_in":7688,"feed_emoji":"👗","tokens_out":9309,"duration_ms":77540,"temperature":0.7,"pith_summary":"EfficientVITON is a virtual try-on system built on a pre-trained Stable Diffusion model. The paper claims that a spatial encoder plus zero cross-attention blocks preserve garment details, while a non-uniform timestep schedule cuts inference time by 72.4% and training time by 45.3% without sacrificing output quality. On VITON-HD it reports the best FID in its comparison table at 8.433 with a RePaint refinement pass, along with an LPIPS of 0.0762. The practical stake is that diffusion-based try-on can become photorealistic and fast enough for e-commerce deployment.","feed_headline":"Faster diffusion makes virtual try-on 72.4% quicker at same quality","feed_subtitle":"EfficientVITON posts top FID on VITON-HD, with a non-uniform denoising schedule that also trims 45.3% off training time.","key_machinery":"The load-bearing mechanism is the non-uniform timestep schedule, which replaces a long chain of uniformly spaced denoising steps with a small set of strategically selected steps and is what produces the claimed speedups. Around it sits an architecture built for detail transfer: a spatial encoder initialized with U-Net weights extracts multi-resolution clothing features, and zero cross-attention blocks (cross-attention layers whose linear projection is zero-initialized) let the U-Net decoder query those clothing features, performing warping directly in latent space. The training objective is $L_{total}=L_{LDM}+\\lambda_{ATV}L_{ATV}$, where $L_{ATV}=\\|\\nabla(FM)\\|_1$ is an attention total variation loss that is meant to suppress noisy attention maps.","core_discovery":"On the paper's own terms, the central discovery is that a diffusion-based virtual try-on pipeline can deliver sharper measured realism than the GAN and diffusion baselines it compares against while costing much less to train and run. EfficientVITON fine-tunes Stable Diffusion with a spatial encoder initialized from U-Net weights and zero cross-attention blocks, so clothing features act as keys and values for patch-wise warping in latent space. Training runs in two stages and combines the latent diffusion loss with an attention total variation loss. With non-uniform timestep sampling, the paper reports FID 8.703 and LPIPS 0.0842 on VITON-HD, improving to 8.433 and 0.0762 when a RePaint refinement pass is added, with inference time falling from 58 seconds to 16 seconds and training time from 1570 hours to 859 hours.","pith_inferences":["If the non-uniform schedule is the true source of the speed gain, it should be ablatable: the same model at the same step count with uniform spacing should look worse, and testing that directly would measure the contribution independently.","The gap between 'Ours' (8.703 FID) and 'Ours (RePaint)' (8.433 FID) suggests the headline quality number depends partly on a refinement pass, so a reader should separate the model's own output from the model-plus-refinement output.","A natural next test is to apply the same spatial-encoder and timestep-schedule combination to unpaired or video try-on, which the paper lists as an unresolved limitation of the field."],"forward_implications":["On the paper's reported hardware, inference drops from 58 seconds to 16 seconds per image, which moves diffusion-based try-on into the range of interactive use.","The non-uniform timestep recipe is independent of the try-on-specific encoders, so the paper's conclusion states it can be carried over to other image synthesis tasks built on latent diffusion.","The two-stage fine-tuning and attention total variation loss are claimed to keep attention maps clean enough to preserve sharp garment logos, text, and folds across poses and body types.","A model with this speed and this FID on VITON-HD becomes a plausible drop-in for e-commerce try-on features, the application context the paper motivates."],"supporting_citations":[{"why":"Supplies the non-uniform timestep sampling method that the paper credits for the 72.4% inference-time and 45.3% training-time reductions.","marker":"[23]"},{"why":"Provides the pre-trained Stable Diffusion latent diffusion model that EfficientVITON fine-tunes for try-on.","marker":"[5]"},{"why":"Introduces the zero cross-attention and semantic-correspondence design that EfficientVITON adapts for latent-space warping.","marker":"[3]"},{"why":"Defines the VITON-HD benchmark dataset and preprocessing targets used for training and evaluation.","marker":"[1]"},{"why":"RePaint denoising refinement is used in the paper's best paired-evaluation results.","marker":"[24]"},{"why":"FID is the primary realism metric on which the paper reports its best comparison-table score.","marker":"[25]"},{"why":"LPIPS is the perceptual similarity metric used alongside FID in the comparison table.","marker":"[26]"}],"fun_headline_variants":["Virtual try-on gets 72% faster, sharper with EfficientVITON","Diffusion try-on: 72.4% speedup, SOTA FID on VITON-HD","EfficientVITON: 16-second try-ons, top FID on VITON-HD","Sharp virtual try-ons in 16s: EfficientVITON beats GANs","Faster diffusion: 72% cut, better realism, SOTA on VITON-HD"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that speed can be gained without quality loss rests on the untested assumption that non-uniform timestep sampling at the reduced step count preserves generation quality in this try-on setting, since the paper never compares its own model against a uniform schedule at the same step count.","fun_headline_variants_meta":{"raw":{"variants":["Virtual try-on gets 72% faster, sharper with EfficientVITON","Diffusion try-on: 72.4% speedup, SOTA FID on VITON-HD","EfficientVITON: 16-second try-ons, top FID on VITON-HD","Sharp virtual try-ons in 16s: EfficientVITON beats GANs","Faster diffusion: 72% cut, better realism, SOTA on VITON-HD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000233,"raw_usage":{"total_tokens":1531,"prompt_tokens":1021,"completion_tokens":510,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":637,"completion_tokens_details":{"reasoning_tokens":386}},"tokens_in":637,"tokens_out":510,"duration_ms":5304,"temperature":1.0,"reasoning_tokens":386,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:52:42.175771+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run EfficientVITON at the same reduced step count with a uniform timestep schedule, keeping all other components fixed, and measure FID and LPIPS on VITON-HD; if uniform sampling matches or beats the non-uniform row, the schedule is not the mechanism behind the efficiency claim. Also rerun the non-uniform model without the RePaint refinement pass to check whether the headline FID of 8.433 is carried by the extra refinement rather than by the model itself.","supporting_citations":[{"cited_title":"High-resolution image synthesis with latent diffusion models,","cited_arxiv_id":null,"evidence_quote":"Provides the pre-trained Stable Diffusion latent diffusion model that EfficientVITON fine-tunes for try-on."},{"cited_title":"Stableviton: Learning semantic correspondence with latent diffusion model for virtual try-on,","cited_arxiv_id":null,"evidence_quote":"Introduces the zero cross-attention and semantic-correspondence design that EfficientVITON adapts for latent-space warping."},{"cited_title":"Viton-hd: High-resolution virtual try-on via misalignment-aware normalization,","cited_arxiv_id":null,"evidence_quote":"Defines the VITON-HD benchmark dataset and preprocessing targets used for training and evaluation."},{"cited_title":"Repaint: Inpainting using denoising diffusion probabilistic models,","cited_arxiv_id":null,"evidence_quote":"RePaint denoising refinement is used in the paper's best paired-evaluation results."},{"cited_title":"The unreasonable effectiveness of deep features as a perceptual metric,","cited_arxiv_id":null,"evidence_quote":"LPIPS is the perceptual similarity metric used alongside FID in the comparison table."}],"review_version":1}