{"id":"638407fe-e7d3-4121-aa13-b4b5959d2f93","arxiv_id":"2607.23189","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A 3D garment generation framework that fuses sketch and texture conditions via a diffusion transformer, outputting simulation-capable 3D Gaussians and meshes.","lead":"This paper presents Fashion-3DLR, a system that turns paired 2D fashion elements—sketches, textures, brush areas—into 3D garment assets, then simulates cloth motion and supports virtual try-on. It targets a real industry workflow, but the reported performance gains are not verifiable as submitted and the comparison tables contain internal inconsistencies.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2 reuses Table 1 baseline numbers verbatim, invalidating the quantitative SOTA comparison.","rationale":"The reader's weakest_assumption focused on the data-construction proxies (HED sketches, neural-painting, random texture patches). That is a legitimate external-validity concern, but the more immediately decisive, internally verifiable problem is the duplicated baseline numbers between Tables 1 and 2. Exact equality of two separate baselines' FIDCLIP and LPIPS values to four decimals across different tasks cannot be a rounding artifact; it indicates the numbers were copied. This directly undermines the headline claim of surpassing previous state-of-the-art methods. The FIDCLIP direction ambiguity adds further uncertainty. Since the quantitative core of the paper is not verifiable from the submission and the duplication is load-bearing, the reader's REJECT verdict remains appropriate. I partially agree with the reader: the proxy-data issue is real but secondary; the table inconsistency is the decisive correctness risk.","tokens_in":13205,"tokens_out":2819,"duration_ms":27368,"concrete_test":"Request the per-sample evaluation logs for Tables 1 and 2, or re-run the evaluation protocol on released checkpoints if available. Verify whether the DressCode and ChatGarment rows reproduce independently; specifically, recompute at least one FIDCLIP and LPIPS value from actual DressCode/ChatGarment outputs. If the values still match the Table 1 baselines exactly, or if they cannot be reproduced at all, the quantitative SOTA comparison is invalid. Also clarify whether FIDCLIP is higher- or lower-is-better and recompute with the correct direction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central SOTA claim is supported by the quantitative comparisons in Tables 1 and 2. In Table 2, DressCode is assigned FIDCLIP=0.4524 and LPIPS=0.6803, while ChatGarment is assigned FIDCLIP=0.4648 and LPIPS=0.7001 — values identical to those reported for Tripo and Hunyuan3D in Table 1. Since these are different methods, evaluated on a different task (sewing-pattern draping vs. sketch+texture 3D generation), exact equality to four decimal places cannot be coincidental. This indicates the baseline entries were copied rather than independently measured. The summary in Sec. 4.3 that 'Fashion-3DLR consistently surpasses all baseline methods' therefore lacks valid quantitative support. In addition, the FIDCLIP metric is described as 'higher score indicates better alignment' while the tables use a downward arrow, so even the direction of the metric is ambiguous. The data-proxy assumption about HED sketches and random texture patches is a further generalization risk, but the immediate load-bearing defect is the duplicated quantitative evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Fashion-3DLR, a controllable 3D garment generation framework that takes paired design elements (e.g., sketch and texture/brush) as input, fuses them via a Garment Feature Fusion Diffusion Transformer (GFF-DiT), generates garment latents with a rectified flow transformer, and decodes them into 3D Gaussian Splatting (3DGS) and mesh representations. The framework also supports physics-based 3DGS cloth simulation and mesh-based virtual try-on. The central claims are that Fashion-3DLR surpasses prior state-of-the-art in generation quality and controllability, and that it is the first end-to-end pipeline to produce simulation-ready garments from paired design elements.","tokens_in":13389,"tokens_out":6594,"duration_ms":56902,"significance":"If the results were valid, this would be a useful contribution to intelligent fashion design, enabling designers to control 3D garment geometry and appearance from 2D sketch and texture inputs. The GFF-DiT fusion mechanism is a sensible extension of prior pairwise fusion work, and the integration with 3DGS physics simulation is relevant. However, the current evidence is not sufficient to support the claims. The duplicated baseline numbers in Tables 1 and 2 invalidate the quantitative SOTA comparison, and the proposed validation of fabric dynamics is self-referential. The paper's strengths are the conceptual pipeline and qualitative results, but these alone do not meet the standard for acceptance.","major_comments":[{"comment":"Table 2 reuses the baseline entries from Table 1 verbatim: DressCode (FIDCLIP=0.4524, LPIPS=0.6803) copies Tripo's row, and ChatGarment (0.4648, 0.7001) copies Hunyuan3D's row. These are different methods evaluated on a different task (sewing-pattern draping vs. sketch+texture 3D generation). Exact equality to four decimals cannot be coincidental. As a result, the conclusion in Sec. 4.3 that 'Fashion-3DLR consistently surpasses all baseline methods' lacks valid quantitative support.","section":"Sec. 4.3, Table 2"},{"comment":"The FIDCLIP metric is defined as 'a higher score indicates better alignment', yet Tables 1–3 label it with a downward arrow, indicating lower is better. This contradiction makes the reported improvements ambiguous; if the definition is correct, the table entries may actually indicate worse performance. Please clarify the metric direction and, if necessary, rerun the evaluation.","section":"Sec. 4.1, Metric definitions"},{"comment":"The controllability and diversity claims are based entirely on automatically derived proxies for design elements: HED edge maps for sketches, neural-painting stylization for brush, random 32×32 patches for textures, and GPT-4o captions. No evidence is given that these proxies match how designers specify elements. If the proxies differ from real inputs, the reported test-set metrics (computed on the same proxy-derived pairs) would overestimate real-world controllability. The paper should include at least a small user study or evaluation with genuine designer-provided inputs.","section":"Sec. 4.1, Data construction"},{"comment":"The fabric-specific parameters θ = {k_b, k_s, k_sh, ρ} are optimized to match a commercial cloth simulator's output (Eq. 7). This validates the 'fabric-specific dynamics' claim only against the simulator used to generate the reference, not against physical measurements. The claim of 'realistic, fabric-specific dynamics' (Sec. 1) is therefore not supported by the presented evidence; the authors should temper the claim or validate against real fabric video.","section":"Sec. 3.4, Physical parameter identification"},{"comment":"None of the quantitative tables include error bars, number of seeds, or significance tests, and the user study (Sec. 4.4) reports only aggregate percentages without statistical detail. The paper also states no code/data/checkpoints will be released. This lack of experimental rigor, especially together with the duplicated baseline numbers, makes it impossible to verify the reported results or assess the significance of the improvements.","section":"General experimental rigor"}],"minor_comments":[{"comment":"The operators ⊕ and ⊗ are reversed relative to standard convention: the text says '⊕ denotes multiplication and ⊗ indicates element-wise addition', which makes the modulation equation difficult to parse. Please fix.","section":"Eq. (1) and surrounding text"},{"comment":"Eq. (6) appears malformed: 'A_p(t)=F_p(t), A_p, F_p(t)⊤' is unclear and should be rewritten with proper notation for the deformation gradient and covariance update.","section":"Eq. (6)"},{"comment":"Table 1 cites [4] for Hunyuan3D-3.0, but reference [4] is the HunyuanImage 3.0 technical report (a 2D image model). Please verify the correct baseline and citation.","section":"Table 1 and References"},{"comment":"The claim of being 'the first end-to-end pipeline' is too strong given the lack of comparison to recent garment-specific 3D generation methods (e.g., GarmentDreamer, Garment3DGen) and the limited evaluation. Suggest softening this claim.","section":"Abstract and Sec. 1"},{"comment":"Placeholder conference information ('Woodstock, NY', 'June 03–05, 2018', 'Conference acronym XX') should be updated or removed. Also, minor typos: 'We introduce' should be lowercase in the Abstract; 'surpass' should be 'surpasses'.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The duplicated baseline entries in Tables 1 and 2 (identical to four decimal places for different methods and tasks) are a serious integrity concern and invalidate the central quantitative claim. Even if this is a copy-paste error, the current manuscript cannot be accepted without a complete re-evaluation of all baselines, a clear statement of metric directions, and a validation of the proxy-based inputs and fabric-dynamics modeling. The lack of code/data release further prevents verification. I recommend rejection, though the authors may resubmit after addressing these issues."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, the paper's actual idea—turn a sketch and a texture patch into a 3D garment, either as 3D Gaussians or a mesh, with non-watertight topology—is a real integration task I haven't seen in the cited prior work. The qualitative figures back that up: the generated garments look structurally plausible and preserve the texture. Second, the quantitative comparison is not trustworthy. Table 2 gives DressCode and ChatGarment FIDCLIP and LPIPS values identical to the Tripo and Hunyuan3D rows in Table 1. Those are different methods evaluated on a different task; exact equality to four decimals is not a coincidence. The summary line that Fashion-3DLR 'consistently surpasses all baseline methods' therefore has no valid numerical support. The FIDCLIP metric description says higher is better while the table arrow points down; that is a separate clarity problem.\n\nWhat the paper does well beyond the idea: the GFF-DiT bidirectional modulation is a sensible way to fuse heterogeneous conditions, and the authors are honest in the conclusion about the current limits (Gaussians can't interact with mesh bodies; mesh garments lack sewing patterns). The physics section is a reasonable adaptation of PhysGaussian, though the fabric parameters are fitted to a commercial simulator's output and no real-fabric validation is attempted. The data construction—HED edges as sketches, random 32x32 patches as textures, GPT-4o captions—is a real generalization risk for actual designer workflows, but it's not a fatal flaw on its own.\n\nThe load-bearing defect is the duplicated numbers. No code, data, checkpoints, or error bars are provided, so there is no independent way to verify even the corrected claims. This reads like a solid systems paper that was rushed into a submission format without doing the baselines properly. The ablated version of each module is traceable to prior work; nothing here suggests the authors are hiding a flaw in the method itself.\n\nWho is this for? People working on fashion AIGC and 3D asset generation from heterogeneous image conditions. The paper deserves a serious referee, but only after the baselines are actually run and the metric direction is fixed. I would not cite it as is, but I would keep an eye on the revision.\n\nRecommendation: send it to peer review, but flag the table duplication as a mandatory fix, not a stylistic suggestion.","headline":"The task formulation is genuinely new, but the SOTA claim collapses on the copied baseline numbers; the paper should be revised before it can be judged.","tokens_in":13962,"tokens_out":1919,"would_cite":false,"duration_ms":19124,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single pipeline turns a sketch and a fabric texture into a simulation-ready 3D garment, fusing both elements in a shared latent space.","keywords":["intelligent fashion design","3D generative model","image-to-image translation","fashion synthesis","3D garment generation","diffusion transformer","3D Gaussian splatting","cloth simulation"],"falsifier":"Take a set of hand-drawn fashion sketches and scanned fabric swatches from a public garment dataset, run Fashion-3DLR on those without any automatic-proxy preprocessing, and compare the output silhouettes and texture fidelity to outputs from the edge-map/random-patch pairs; if the alignment metrics degrade to the level of an un-conditioned baseline, the data-equivalence assumption that underwrites the controllability claim is false.","tokens_in":13045,"feed_emoji":"👗","tokens_out":12592,"duration_ms":105688,"temperature":0.7,"pith_summary":"This paper sets out to show that 3D garment generation can be controlled by the same two design elements a fashion designer actually uses—a structural sketch and an appearance reference such as a fabric texture or brush painting. It introduces Fashion-3DLR, an end-to-end model that bridges the semantic gap between these heterogeneous 2D inputs, fuses them into a shared latent space, and decodes the result into both 3D Gaussians and meshes that preserve the sketch's silhouette and the texture's material look. If the premise holds, designers could generate, refine, and physically simulate garment concepts without manual topology modeling or sewing-pattern reconstruction, and the same asset could serve virtual try-on. The paper reports that its outputs are non-watertight and wearable, support material-specific dynamics for cotton, silk, wool, and nylon, and outperform the comparison methods on the metrics they define.","feed_headline":"Turn a sketch and a texture into a simulation-ready 3D garment","feed_subtitle":"Designers can skip manual 3D modeling: the same pipeline drives material-aware simulation and virtual try-on.","key_machinery":"The load-bearing mechanism is the Garment Feature Fusion Diffusion Transformer (GFF-DiT), a diffusion transformer that fuses two heterogeneous design elements into one conditional latent via symmetric bidirectional modulation: sketch features are normalized and then rescaled by scale-and-shift factors predicted from texture features, and vice versa, so either modality can steer the fusion while both are preserved. The fused condition is injected as keys and values into a rectified flow transformer, where the noisy garment latent serves as the query; after denoising, convolutional up-sampling converts the latent into 3D Gaussian splats or a mesh. The physical-simulation extension rests on a p","core_discovery":"The central claim is that pairing a sketch and a texture (or oil-brush painting) is a sufficient condition set for generating a complete, wear-ready 3D garment in two exchangeable output formats. To achieve this, the authors propose a Garment Feature Fusion Diffusion Transformer (GFF-DiT) that bidirectionally modulates the two inputs—each modality is normalized and rescaled by affine factors predicted from the other—so that neither element dominates and the structural and appearance semantics are preserved together. The fused condition is then fed, through cross-attention, into a rectified flow transformer that denoises a noisy latent into a garment latent, which is upsampled into voxelized","pith_inferences":["A fair test of the controllability claim would be to run the trained model on hand-drawn designer sketches and real fabric photographs, since the paper's training and evaluation pairs are built from automatic proxies (edge maps, stylization filters, and random patches); if performance drops sharply on authentic inputs, the practical gains may be smaller than the reported metrics suggest.","The same bidirectional-fusion architecture is not inherently garment-specific; it could plausibly be retrained for other pairwise heterogeneous conditions, such as a blueprint plus a material palette for furniture, or character line-art plus a color palette, and would likely face the same semantic-gap problem.","Because the fusion is symmetric and operates before geometry generation, the framework offers a natural way to disentangle structure from appearance—fixing the texture and varying the sketch (or vice versa) explores shape or material variations independently, which the paper demonstrates qualitatively but does not formalize as an editing API.","The paper itself closes by naming two boundaries: the Gaussian garment does not yet interact with mesh-based human bodies, and the mesh garment does not encode sewing-pattern structure; those are the places where the demonstrated gains stop and where the garment would need to be re-integrated into a full production pipeline."],"forward_implications":["A sketch and a fabric texture (or brush painting) become sufficient control signals to generate a complete 3D garment that preserves the sketch's silhouette and the texture's appearance.","The generated 3D Gaussian garment can be simulated directly, without an intermediate mesh, and the motion parameters can be tuned per fabric (cotton, silk, wool, nylon) while keeping fine geometric details such as thin straps.","The same latent representation can be decoded into a mesh for virtual try-on, and the mesh inherits the input sketch's flat hems, sleeve lengths, and openings rather than collapsing into a watertight shell.","The reported ablation shows the fusion module is the source of the method's controllability: removing GFF-DiT degrades all evaluated quality and alignment metrics."],"fun_headline_variants":["Sketch plus texture equals a 3D garment that simulates","Two fashion elements, one controllable 3D garment","AI turns sketch and texture into simulation-ready 3D clothes","From 2D sketch to physical 3D garment in one flow","Garment AI: pair a sketch and texture, get a virtual try-on"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The entire controllability claim rests on the assumption that automatically generated stand-ins for design elements—edge maps for sketches, automated stylizations for brush strokes, and random cropped patches for fabric texture—faithfully capture how a human designer would draw, paint, and specify fabric.","fun_headline_variants_meta":{"raw":{"variants":["Sketch plus texture equals a 3D garment that simulates","Two fashion elements, one controllable 3D garment","AI turns sketch and texture into simulation-ready 3D clothes","From 2D sketch to physical 3D garment in one flow","Garment AI: pair a sketch and texture, get a virtual try-on"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000285,"raw_usage":{"total_tokens":1553,"prompt_tokens":818,"completion_tokens":735,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":645}},"tokens_in":562,"tokens_out":735,"duration_ms":7582,"temperature":1.0,"reasoning_tokens":645,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T03:20:39.186875+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of hand-drawn fashion sketches and scanned fabric swatches from a public garment dataset, run Fashion-3DLR on those without any automatic-proxy preprocessing, and compare the output silhouettes and texture fidelity to outputs from the edge-map/random-patch pairs; if the alignment metrics degrade to the level of an un-conditioned baseline, the data-equivalence assumption that underwrites the controllability claim is false.","supporting_citations":[],"review_version":1}