{"id":"4909921c-6969-4760-bf0c-3ddcd9748bdd","arxiv_id":"2501.04631","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A feed-forward diffusion model generates fully disentangled clothed avatars by representing body, hair, and clothing in separate layers of a Gaussian-based UV feature plane.","lead":"LayerAvatar generates a clothed 3D avatar from noise in seconds with its body, hair, clothes, and shoes kept as separate layers that can be edited and transferred. It combines a layered UV map of 3D Gaussians with a single-stage diffusion model trained on multi-view images.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No held-out split is reported for the THuman2.0 FID evaluation, so the headline generation-quality claim may reflect memorization rather than generalization.","rationale":"The reader's stated weakest assumption is the dependence on Sapiens segmentation and SMPL-X fitting for per-component ground truth. That is a real concern, and the paper itself acknowledges it in Sec. 8.1(1). However, the method already includes robustness mechanisms (Huber losses, multi-view consistency, filtering of noisy samples), and the qualitative decomposition and component-transfer results provide some evidence that the disentanglement pipeline works despite label noise. The training/evaluation overlap issue is more load-bearing because it threatens the only quantitative support for the 'high-quality' part of the central claim. The paper describes training on THuman2.0 and then evaluating on THuman2.0 without ever stating a train/test split. For a generative model, FID is a distributional metric; if the reference distribution is the training distribution, an overfit model can score artificially well. This is not an accusation of dishonesty; it is a missing control that is standard for generative-model evaluation. The concern is addressable, so it does not warrant rejection, but it does mean the current evidence is insufficient to accept the headline claim at face value. The layered representation and single-stage training scheme are otherwise coherent, and the ablations in Tab. 2 support the value of the layered design, so the appropriate verdict remains CONDITIONAL pending the requested evaluation protocol.","tokens_in":17722,"tokens_out":7328,"duration_ms":78472,"concrete_test":"Request or reconstruct the exact scan IDs used for THuman2.0 training and for FID evaluation, or rerun the evaluation with an explicit held-out split: hold out a fixed subset of scans before fitting UV feature planes and training the UNet, generate a large batch of avatars, render them, and compute FID against only the held-out scans. Also compute the recall and nearest-neighbor distance from generated renders to the training renders as a memorization check. If the FID degrades substantially from 12.50 or the ranking against E3Gen changes under this controlled protocol, the high-quality generation claim is not currently supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of generating 'high-quality ... in seconds' rests on the quantitative comparison in Tab. 1a, where LayerAvatar reports FID 12.50 on THuman2.0. The paper never states that the scans used to compute this FID are disjoint from the scans used to fit the layered UV feature planes and train the diffusion model. In Sec. 4 (Dataset), 500 THuman2.0 scans are sampled and rendered from 54 views, and this same dataset is described as part of the composite training set. Section 4.1 then evaluates 'on THuman2.0 dataset' with FID. If the FID reference images are drawn from the training scans, a generative model that memorizes training examples can achieve an artificially low FID. The comparison is further complicated by the fact that baseline FID values are adopted from other papers (footnotes in Tab. 1), which may use different reference sets and rendering protocols. This is load-bearing because it directly affects the 'high-quality' half of the abstract's claim: even with perfect semantic masks and perfect component disentanglement, the headline result is not established unless the FID is computed against a held-out reference set. The issue is distinct from the Sapiens-segmentation concern identified by the reader; it is an evaluation-protocol gap that can be closed without changing the method.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LayerAvatar, a feed-forward diffusion-based method for generating clothed avatars with disentangled body, hair, shoes, top, and bottom components. The key representation is a layered UV feature plane storing Gaussian attributes for three component groups (body; hair+shoes; top+bottom), with semantic labels used to separate components at render time. The method trains a single-stage diffusion model on multi-view images with reconstruction, segmentation, and occlusion constraints, and claims high-quality generation in seconds, support for animation and novel-view synthesis, and component transfer. Experiments compare holistic generation quality on THuman2.0, layer-wise quality on Tightcap, and component quality against optimization-based methods, with ablations studying the layered representation and single-stage training.","tokens_in":17965,"tokens_out":4156,"duration_ms":39524,"significance":"If the quantitative claims are established, this is a meaningful step: a feed-forward generator that outputs disentangled, animatable avatars with component transfer is practically useful, and the layered UV feature plane is a well-motivated representation that combines the editability of layers with the quality of Gaussian splatting. The paper also provides ablations that support the layered representation over single-layer alternatives, and it ships a reproducible-looking pipeline with explicit loss terms and hyperparameters. However, the current evaluation protocol leaves the headline claims of 'high-quality' and 'superior performance' insufficiently supported, so the significance depends on closing the measurement gaps below.","major_comments":[{"comment":"The FID evaluation on THuman2.0 does not specify any held-out split. Section 4 describes sampling 500 THuman2.0 scans, rendering 54 views each, and using them for training, and then describes a composite training set of 1954 selected scans from THuman2.0, THuman2.1, and CustomHuman. Section 4.1 then evaluates 'on THuman2.0 dataset' and reports FID 12.50, but it is never stated that the scans used to compute the FID reference set are disjoint from the scans used to fit the layered UV feature planes and train the diffusion model. If the reference images are drawn from the training scans, the FID can be artificially low due to memorization. This directly affects the abstract's claim of 'high-quality' generation. Please state the exact split and, if necessary, recompute the FID on a held-out subset of THuman2.0 subjects.","section":"Sec. 4 (Dataset) and Sec. 4.1 (Evaluation of Generation Quality)"},{"comment":"The baseline FID values for EVA3D, StructLDM, and E3Gen are adopted from their respective papers (footnotes *, †, ⋆), which may use different reference image sets, rendering protocols, and evaluation subsets. A lower FID under a different protocol is not evidence of superiority. The claim in Sec. 4.1 that 'Our method outperforms all baselines on THuman2.0' is therefore not supported by a controlled comparison. Please re-run the baselines under the same evaluation protocol used for LayerAvatar, or at least clearly restrict the claim to the adopted numbers and discuss protocol differences.","section":"Table 1a and footnotes"},{"comment":"The L-PSNR value for the proposed method is reported as '>40' rather than as an exact number, with the explanation that the two masked layers are nearly identical. This is not a precise metric value and it makes the disentanglement comparison in Table 1b difficult to interpret; the text states the method 'surpasses other methods' in L-PSNR, but a threshold without a distribution cannot support a quantitative superiority claim. Please report the actual mean L-PSNR (and ideally its standard deviation or per-sample distribution) for the Tightcap layer-wise evaluation.","section":"Table 1b and Sec. 4.1"},{"comment":"The component supervision for THuman2.0, THuman2.1, and CustomHuman derives entirely from Sapiens-predicted semantic segmentation, as stated in Sec. 3.4 ('The ground truth of silhouette masks is estimated based on the semantic segmentation results predicted by Sapiens'). The authors acknowledge this in Sec. 8.1(1). Since the same kind of predicted masks are also used to verify disentanglement in the THuman2.0-based results, the paper would be strengthened by a quantitative sensitivity analysis of the segmentation noise, beyond the qualitative Figure J. For example, report FID or component-mask IoU when training with clean masks (e.g., on Tightcap) versus predicted masks, or evaluate the final generation quality on a subset with manually verified masks.","section":"Sec. 3.4 and Sec. 7.3"}],"minor_comments":[{"comment":"The sentence 'The elimination of identity shifting demonstrates that our method achieves full disentanglement' overstates the implication: absence of identity shifting is one proxy for disentanglement, not a proof of full disentanglement. Consider softening the wording.","section":"Sec. 4.1"},{"comment":"No error bars, confidence intervals, or number of random seeds are reported for any FID, KID, L-PSNR, or user-study result. Given the variability of generative model metrics, at least a note on single-seed reporting or a variance estimate would improve reproducibility.","section":"Table 1"},{"comment":"The label 'Clothed AvatarHuman BodyExteriorComponentsSegmentation maskConstraint Loss' in Figure 2 appears to be a formatting artifact; please fix the figure layout so that each label is legible and attached to the correct part.","section":"Sec. 7.2 and Figure 2"},{"comment":"There is a typo in '3G Gaussians' which should read '3D Gaussians'.","section":"Sec. 8.1(2)"},{"comment":"The main text says the UV feature plane is split 'channel-wise', but the supplementary says the three layers are concatenated width-wise into a tensor of size 12 × 128 × 384. Please clarify the exact layout (channel dimension and width dimension) in the main text for consistency.","section":"Sec. 3.2 and Sec. 7.1"}],"recommendation":"major_revision","confidential_remarks":"The core method is plausible and the representation is interesting, but the quantitative evaluation needs substantial strengthening before the paper can be accepted. The held-out split issue is the most serious: the paper must state explicitly whether the THuman2.0 FID is computed on scans that were used in training, and if so, redo the evaluation on a disjoint set. The adopted-baseline FID numbers and the truncated L-PSNR also need to be addressed. The novelty claim of being 'the first feed-forward diffusion-based method' for component-disentangled avatars would benefit from a more careful positioning against HumanLiff and other layer-wise generators, though this is not a blocking issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nLayerAvatar is a genuinely new piece of systems work: it takes E3Gen's single UV feature plane for Gaussian avatars and splits it into semantic layers (body, hair+shoes, top+bottom), then trains a single-stage diffusion model over that layered latent. That combination is, as far as I can tell, not in the prior work they cite, and the ablations against single-layer variants support the design choice. The disentanglement results in the figures are visibly better than the baselines, and the component transfer application is a real payoff. If the representation and training objective are taken at face value, this is a solid contribution to feed-forward avatar generation.\n\nThe soft spots are all in the evaluation protocol, not in the method's internal logic. The big one: the headline FID of 12.50 on THuman2.0 is computed without any stated split between training scans and the reference set. The training set is built from THuman2.0/2.1/CustomHuman; Section 4 describes sampling 500 THuman2.0 scans, and Section 4.1 just says \"on THuman2.0.\" If the FID reference images come from the same scans used to fit the UV planes, a model that memorizes training views can get an artificially low FID. The stress-test catches this correctly. It is load-bearing for the \"high-quality\" half of the abstract, and it is fixable by reporting a held-out split, or by retraining on a separate subset. Related minor issues: baseline FIDs are adopted from other papers with potentially different rendering protocols; L-PSNR is truncated to \">40\" without giving the actual number; and the dependence on Sapiens segmentation masks is acknowledged in Sec. 8.1 but not quantified (no error analysis on mask noise, even though they show one robustness figure). None of these are fatal, and the central method does not collapse.\n\nWho this is for: anyone working on 3D human generation, editable avatars, or compositional generative models. It deserves a serious referee pass; the evaluation issues are exactly the kind that peer review should catch and require the authors to fix. I would not desk-reject this. It should go to review, and I'd expect a conditional accept after the authors supply a proper test split and baseline protocol description.\n\nRecommendation: send to peer review, with an emphasis on evaluation protocol.","headline":"Layered UV feature plane is a real step forward for feed-forward avatar generation, but the headline FID is unproven until the authors report a held-out training/evaluation split.","tokens_in":18499,"tokens_out":1709,"would_cite":true,"duration_ms":15818,"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":"This paper introduces LayerAvatar, a feed-forward diffusion method that generates clothed 3D avatars whose body, hair, shoes, and top/bottom clothing are stored in separate layers of a Gaussian-based UV feature plane, so components can be…","keywords":["LayerAvatar","clothed avatar generation","component disentanglement","layered UV feature plane","3D Gaussian splatting","diffusion model","component transfer","SMPL-X"],"falsifier":"Take a group of subjects in tight, skin-colored, or patterned clothing, render their multi-view images, run the same segmentation and SMPL-X fitting, and compare the isolated body layer against the true body surface obtained from an under-clothing scan or body-fitting method. If the body layer systematically adopts clothing color or geometry wherever segmentation confuses fabric with skin, the claimed disentanglement fails. A cheaper quantitative version is to measure L-PSNR between body and clothing layers on such subjects; if it falls toward the single-layer baseline value, the layered representation only separates components when the supervision is clean.","tokens_in":17477,"feed_emoji":"👤","tokens_out":9638,"duration_ms":79985,"temperature":0.7,"pith_summary":"LayerAvatar is a method for generating clothed 3D avatars whose parts are separated at creation time: the body, hair, shoes, and top and bottom clothing are stored in different layers of a shared UV feature plane built around 3D Gaussians. The paper positions it as the first feed-forward diffusion method for this task, aiming to show that the layered representation lets a single model produce a fully disentangled, animatable avatar in seconds rather than the hours that per-component optimization methods need. The methods it argues against either treat the avatar as one undivided object, which blocks editing, or optimize each component separately, which is slow. A sympathetic reader would care because, if the claim holds, swapping clothes, hair, or shoes between avatars, and re-posing them, become direct operations on the representation instead of separate reconstruction problems.","feed_headline":"LayerAvatar splits body, hair, and clothes into editable avatar layers","feed_subtitle":"New generation method stores each part in its own layer, so avatars are made in seconds and pieces swap directly.","key_machinery":"The layered UV feature plane carries the argument. Instead of scattering Gaussians in an unstructured 3D field, the method initializes Gaussian primitives on per-component SMPL-X templates and writes their attributes (position offset, opacity, rotation delta, scale delta, color) as local features in a shared 2D UV plane divided into three layers with semantic labels. Two lightweight shared MLP decoders, one for geometry and one for texture, convert the plane into attribute maps, and each Gaussian samples its own attributes by bilinear interpolation. Because overlapping components occupy different layers, they get independent feature spaces, which is what lets the diffusion model separate body from clothing; because the templates carry SMPL-X skinning weights, the same representation deforms into novel poses and body shapes by linear blend skinning.","core_discovery":"The central claim is that component disentanglement can be baked into the generative representation rather than imposed afterward. LayerAvatar represents a clothed avatar as a set of Gaussian primitives $G_{\\text{avatar}} = \\{G_{\\text{body}}, G_{\\text{top}}, G_{\\text{bottom}}, G_{\\text{hair}}, G_{\\text{shoes}}\\}$, each component anchored to an SMPL-X-based template and mapped into a three-layer UV feature plane with semantic labels: the innermost layer holds the body, the second holds hair and shoes, and the third holds top and bottom clothing. A single-stage diffusion model is trained on these layered planes, simultaneously fitting the planes from multi-view images and learning the denoising prior, with per-component rendering losses, a semantic segmentation loss, and two occlusion-specific constraints (an inner-body mask term and a skin-color prior from the hands) to keep the heavily occluded body layer plausible. The paper reports state-of-the-art FID on THuman2.0, layer-wise L-PSNR above 40 on Tightcap, decomposition without identity shifting, and successful component transfer across subjects.","pith_inferences":["A testable extension is to add more layers for accessories such as glasses or bags; the architecture treats layers as independent UV planes, so capacity and disentanglement should scale with layer count rather than requiring a redesign.","Because generation happens in a latent UV space, a conditioning mechanism (text prompt, pose, or identity image) could be attached to the denoiser to make the same representation support controlled generation; the paper does not explore this.","The fidelity of the body layer will probably track the quality of the segmentation supervision; comparing results on datasets with true component meshes, such as the Tightcap split used here, against the Sapiens-mask supervision on THuman2.0 would isolate this dependence.","The paper leaves implicit that a collision-avoidance post-process on the body and clothing Gaussians could turn the generated layers into physically plausible assets, which matters for animation and simulation downstream."],"forward_implications":["A single avatar is generated in about two seconds, because the diffusion model produces the whole layered representation in one pass and no per-subject optimization is needed.","Components transfer directly: the paper demonstrates upper clothes, pants, hair, and shoes being moved between avatars of different body shapes while keeping detail.","Generated avatars support animation: because each layer is attached to SMPL-X templates, novel gestures, facial expressions, and body shapes are obtained by warping and linear blend skinning.","Decomposition is stable: rendering any single layer does not cause identity shifting, which the paper uses as evidence that the layers truly separate the avatar rather than sharing entangled features.","The representation also generalizes to multi-layer outfits and dress/skirt types, and to single-image reconstruction of unseen subjects, suggesting the same pipeline will scale to larger, more diverse training sets."],"supporting_citations":[{"why":"Supplies the primitive representation whose attributes are stored on UV planes and rendered efficiently by splatting.","marker":"3D Gaussians [30]"},{"why":"Predicts the semantic segmentation maps that are merged into five component labels and used as ground truth for per-component masks and the segmentation loss.","marker":"Sapiens [31]"},{"why":"Provides the body model whose templates anchor all component Gaussians and whose blend shapes and skinning weights drive pose and expression deformation.","marker":"SMPL-X [52]"},{"why":"Contributes the single-stage training scheme in which UV feature plane fitting and diffusion training happen jointly, with the diffusion loss acting as a prior for occluded regions.","marker":"[6]"},{"why":"Establishes the idea of storing Gaussian attributes in a UV feature plane for feed-forward avatar generation, which LayerAvatar extends to a layered, component-disentangled version.","marker":"E3Gen [87]"},{"why":"Is the closest layer-wise generation baseline; its preprocessed Tightcap split is used to compare layer-wise quality and disentanglement.","marker":"HumanLiff [24]"},{"why":"Is the main dataset for holistic generation evaluation and part of the training data.","marker":"THuman2.0 [81]"},{"why":"Provides separate meshes for cloth and shoes, enabling direct rendering of per-component multi-view images and silhouettes for layer-wise evaluation.","marker":"Tightcap [9]"}],"fun_headline_variants":["LayerAvatar bakes part disentanglement into generation","Generate avatars with separate editable layers via LayerAvatar","LayerAvatar splits body, hair, and clothes into layers","Disentangled avatar generation: LayerAvatar stacks components","Instant avatar generation with editable parts: LayerAvatar"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise, which the paper itself flags in Section 8.1(1), is that the semantic segmentation maps used as ground truth correctly separate body, hair, top, bottom, and shoes, and that the fitted SMPL-X body template lies inside the clothing; if either is wrong, the layers receive incorrect supervision and the disentanglement collapses.","fun_headline_variants_meta":{"raw":{"variants":["LayerAvatar bakes part disentanglement into generation","Generate avatars with separate editable layers via LayerAvatar","LayerAvatar splits body, hair, and clothes into layers","Disentangled avatar generation: LayerAvatar stacks components","Instant avatar generation with editable parts: LayerAvatar"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000232,"raw_usage":{"total_tokens":1495,"prompt_tokens":957,"completion_tokens":538,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":463}},"tokens_in":573,"tokens_out":538,"duration_ms":5685,"temperature":1.0,"reasoning_tokens":463,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:28:32.627169+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a group of subjects in tight, skin-colored, or patterned clothing, render their multi-view images, run the same segmentation and SMPL-X fitting, and compare the isolated body layer against the true body surface obtained from an under-clothing scan or body-fitting method. If the body layer systematically adopts clothing color or geometry wherever segmentation confuses fabric with skin, the claimed disentanglement fails. A cheaper quantitative version is to measure L-PSNR between body and clothing layers on such subjects; if it falls toward the single-layer baseline value, the layered representation only separates components when the supervision is clean.","supporting_citations":[],"review_version":1}