{"id":"81b6f4f5-445f-4e78-9db6-2748bf23cec8","arxiv_id":"2507.07591","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A multi-view diffusion hair-transfer system transfers a reference hairstyle onto a portrait and renders the edited person from many consistent viewpoints.","lead":"Stable-Hair v2 turns a photo of a person plus a reference hairstyle into a short multi-view video of that person wearing the reference cut, using a diffusion model trained with synthetic multi-view data. It targets digital humans and virtual avatars, where consistent hair from many angles is hard to generate.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table II contradicts the paper's own claim: on the multi-view benchmark, Stable-Hair v2 has worse FID, PSNR, and SSIM than HairFusion+SV3D, so the headline 'significantly outperforming existing methods' is not supported by the reported numbers.","rationale":"The reader's weakest-assumption focused on the synthetic-data circularity: the face-finetuned SV3D model is trained on synthetic GRAM-HD videos, and the same pipeline produces the 2,500-set multi-view benchmark, so the quantitative evaluation may not transfer to real multi-view captures. That is a valid external-validity concern. However, I find a more immediate and load-bearing problem: the reported Table II numbers contradict the paper's own prose on the multi-view task. The proposed method is worse than HairFusion+SV3D on FID, PSNR, and SSIM, and better by small margins on CLIP-I and IDS. Since the central contribution is multi-view hair transfer, the paper must either correct the table or revise the claim. This is an internal inconsistency, not a matter of consensus, so it cannot be waved off by appealing to the novelty of the setup. The secondary self-referential benchmark concern strengthens the case for caution but is not needed to establish that the strong claim is currently unsupported. I recommend keeping the reader's CONDITIONAL verdict: the method may still be a useful new-application contribution, but the paper needs corrected, independent evaluation before its central performance claims can be accepted.","tokens_in":17540,"tokens_out":3147,"duration_ms":35790,"concrete_test":"Recompute the multi-view quantitative results in Table II using the released code and the paper's protocol: regenerate or reuse the 2,500-set benchmark, compute FID under the standard lower-is-better convention, and report paired bootstrap 95% confidence intervals over the 50 test samples for CLIP-I, FID, PSNR, SSIM, and IDS. If the reproduced FID for Stable-Hair v2 remains ~32.170 while HairFusion+SV3D is ~31.203 and PSNR/SSIM remain lower, the claim of 'significantly outperforming' is false; if the numbers flip or the intervals overlap, the paper should correct Table II and explicitly soften the abstract claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative evidence for the paper's main claim is internally inconsistent. In Section IV.C.2, the text says: 'Our method not only addresses the limitations of existing approaches but also demonstrates superior performance in terms of FID and IDS.' But Table II reports for the multi-view task: CLIP-I: Ours 0.411 vs HairFusion+SV3D 0.406; FID: Ours 32.170 vs HairFusion+SV3D 31.203 and HairCLIPV2+SV3D 29.652 (FID is lower-is-better); PSNR: Ours 26.347 vs HairFusion+SV3D 27.846; SSIM: Ours 0.490 vs HairFusion+SV3D 0.564; IDS: Ours 0.683 vs HairFusion+SV3D 0.659. Thus, on three of the five headline metrics, including FID, the proposed method is worse than the stronger baseline, and on CLIP-I and IDS the advantage is small. The abstract's statement that the method 'significantly outperforming existing methods' is not supported by these numbers, even before considering the more general concern that the benchmark is generated by the same synthetic pipeline used for training. This is not a disagreement with an outside consensus; it is an internal contradiction between the table and the prose. If the FID entries are correct, the multi-view SOTA claim is false as stated; if they are mislabeled or computed with a different convention, the table needs correction and the comparison should be rerun.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Stable-Hair v2, a multi-view hair transfer framework built on a latent diffusion backbone. The main contributions are: a training-data generation pipeline that produces triplets of source, reference, and bald images together with multi-view pairs, using a diffusion-based bald converter, a data-augmented inpainting model, and a face-finetuned SV3D model; a pose-controllable latent IdentityNet for view control; a hair extractor with cross-attention; temporal attention layers for view consistency; and a three-stage training strategy. The authors claim this is the first work to apply multi-view diffusion models to hair transfer and that it significantly outperforms existing methods, based on single-view and multi-view quantitative comparisons, qualitative visualizations, user studies, and ablations.","tokens_in":17953,"tokens_out":3414,"duration_ms":37552,"significance":"The single-view results, the detailed data-generation pipeline, and the modular three-stage training strategy are useful contributions, and the authors make the code publicly available. If the multi-view claims held, this would be a meaningful step toward practical avatar and digital-human applications. However, the central multi-view superiority claim is directly contradicted by the paper's own Table II, and the quantitative evaluation is performed on a dataset generated by the same synthetic pipeline used for training, which substantially weakens the 'real-world' and state-of-the-art assertions as currently written.","major_comments":[{"comment":"The text states that 'Our method not only addresses the limitations of existing approaches but also demonstrates superior performance in terms of FID and IDS.' This is not supported by Table II: with FID (lower is better), Stable-Hair v2 scores 32.170, worse than HairFusion+SV3D (31.203) and HairCLIPV2+SV3D (29.652). The method is also worse on PSNR (26.347 vs. 27.846 for HairFusion+SV3D) and SSIM (0.490 vs. 0.564). Only CLIP-I (0.411 vs. 0.406) and IDS (0.683 vs. 0.659) show small advantages over the stronger baseline. The abstract's claim of 'significantly outperforming existing methods' is therefore internally inconsistent with the reported multi-view numbers. The authors must either correct the table or the prose, or rerun the multi-view evaluation; as presented, the multi-view SOTA claim is false on three of the five headline metrics.","section":"Section IV.C.2, Table II"},{"comment":"The quantitative multi-view evaluation uses 'a dataset generated by our training data pipeline' (2,500 multi-view sets). Because the proposed model is trained on the same synthetic SV3D/GRAM-HD distribution used to create this test set, while the comparison pipelines (HairCLIPV2/SV3D, HairFusion/SV3D) are applied post hoc to real images, the comparison is biased in favor of the proposed method. The paper does not report whether test identities were held out from the approximately 20,000 training sets, nor does it validate the synthetic multi-view data against real multi-view portrait captures. The authors should evaluate on real multi-view data or at minimum provide disjoint train/test splits and an analysis of the synthetic-to-real gap.","section":"Section IV.B and Section IV.C.2"},{"comment":"The entire training and evaluation pipeline depends on the face-finetuned SV3D model, which is fine-tuned on roughly 20,000 GRAM-HD-generated multi-view face videos and then used to produce 21 views for both source and bald images. The manuscript offers no validation that these generated multi-view pairs are view-aligned, identity-preserving, or geometrically accurate. Since the method is claimed to work in 'real-world' settings, the authors should provide evidence that the synthetic multi-view data is a faithful proxy for real multi-view captures, for example by comparing generated views with real multi-view portrait data or by reporting pose/identity accuracy metrics on held-out real images.","section":"Section III.B.3 and Section IV.A"}],"minor_comments":[{"comment":"The LDM loss in Eq. (1) is written as E(||epsilon - epsilon_theta(zt,t,c)||), missing the squared L2 norm and the subscript 2 used in the standard formulation; please align with the original LDM objective.","section":"Equation (1)"},{"comment":"The naming is inconsistent: the module is called 'Pose-Controllable Latent IdentityNet' in the text but 'Latent ControlNet' and 'pixel' vs. 'latent' in Fig. 9 and Table III. Please unify the terminology.","section":"Section III.C.1 and Fig. 9 / Table III"},{"comment":"The symbol epsilon is used both for the noise in Eq. (1) and for the 'noise augmentation of camera pose' in Eq. (2); this is confusing and should be clarified with distinct notation.","section":"Equation (2)"},{"comment":"The test set is described as '50 distinct samples' in Section IV.B but the quantitative experiments in Section IV.C.2 mention 2,500 sets of multi-view images; the relationship between these numbers should be clarified.","section":"Section IV.A vs. Section IV.B"},{"comment":"References [15] and [19] appear to be the same paper (Chung et al., AAAI 2025) with different titles and venues; please deduplicate and provide the correct citation.","section":"References [15] and [19]"}],"recommendation":"major_revision","confidential_remarks":"The contradiction in Table II is serious: the central claim of multi-view superiority is not supported by the reported numbers. If the authors cannot supply corrected results or a genuine real-data evaluation, the paper should not be accepted. The synthetic evaluation set also needs to be addressed transparently. The 'first work' claim should be checked carefully against recent multi-view hair/avatar literature before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is the first diffusion-based multi-view hair transfer framework, and that novelty is real. The pipeline—bald converter, data-augmented inpainting, face-finetuned SV3D for multi-view triplets, then stage-wise training of a pose-controllable latent IdentityNet, hair extractor, and temporal attention—is coherent and the ablations support the design choices. Single-view results in Table I are consistently strong, and the code is public. That part deserves credit.\n\nThe soft spot is load-bearing. In Section IV.C.2, the text says the method shows \"superior performance in terms of FID and IDS,\" but Table II reports FID 32.170 for Ours versus 29.652 for HairCLIPV2+SV3D and 31.203 for HairFusion+SV3D. Lower FID is better, so Ours is worst. PSNR and SSIM are also worse than HairFusion+SV3D. Only CLIP-I and IDS are best, and the margins are thin. The abstract's \"significantly outperforming existing methods\" is simply not supported by the numbers as printed. This is an internal contradiction, not a matter of interpretation.\n\nThe second issue is the evaluation set. Section IV.B states the quantitative multi-view experiments use \"a dataset generated by our training data pipeline.\" Training and evaluation share the same synthetic SV3D-generated distribution, and the fine-tuned SV3D itself was trained on synthetic GRAM-HD videos. There is no independent real multi-view capture anywhere in the loop. So the multi-view numbers can't establish real-world view consistency, which is the paper's central claim. That is a serious circularity, though not a fabrication—the method might well work in practice, but the evidence here doesn't prove it.\n\nAlso worth noting: the paper itself acknowledges suboptimal background reconstruction in multi-view, which explains the PSNR/SSIM shortfall, but then the overall superiority claim should be qualified accordingly.\n\nWho this is for: people working on portrait editing, avatar creation, and virtual try-on. The architecture will be useful even if the evaluation is reworked. It deserves a serious referee, but only with major revision. The quantitative comparison must be corrected, error bars added, and the multi-view evaluation rerun on real or at least independent data, ideally with direct baselines like HairNeRF and StyleGAN-Salon.","headline":"Genuinely new multi-view hair transfer framework, but the paper's own Table II contradicts its headline claim of superiority.","tokens_in":18425,"tokens_out":1744,"would_cite":false,"duration_ms":19340,"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":"Diffusion model transfers hairstyles across multiple views.","keywords":["multi-view hair transfer","diffusion models","avatar generation","hairstyle transfer","view consistency","pose conditioning","temporal attention","synthetic training data"],"falsifier":"Render the fine-tuned multi-view model's 21 outputs for real FFHQ images and compare them against ground-truth multi-view portrait captures of the same people, measuring head-pose error and identity similarity; if the synthetic pairs show systematic pose or identity drift, the training-data foundation of the method collapses. A simpler check is to run the full pipeline on images with extreme or unusual viewpoints and count how often the multi-view output loses identity or produces inconsistent head shapes.","tokens_in":17287,"feed_emoji":"💇","tokens_out":4084,"duration_ms":42986,"temperature":0.7,"pith_summary":"The paper tries to establish that multi-view hair transfer—placing a reference hairstyle onto a source face while preserving identity and background from every angle—can be done with a diffusion model trained on synthetic multi-view data. It argues this is the first such system and reports that it outperforms existing single-view hair transfer methods that are piped through a multi-view generator. The practical payoff, if true, is that digital humans and avatars could get consistent hair styling from one reference photo instead of per-view editing. The paper also presents a synthetic data pipeline that produces bald images, reference hairstyles, and view-aligned source–bald pairs for training and evaluation.","feed_headline":"Diffusion model transfers hairstyles across multiple views","feed_subtitle":"Synthetic multi-view training data keeps faces stable while hair moves from one portrait to another.","key_machinery":"The load-bearing machinery is the combination of a synthetic multi-view training pipeline and a staged training strategy. The face-finetuned multi-view diffusion model (SV3D fine-tuned on 3D GAN facial videos) supplies view-aligned source–bald pairs, and the polar–azimuth pose embedding inside a Latent IdentityNet supplies viewpoint control. A trainable Hair Extractor, whose self-attention features become keys and values for new cross-attention layers, supplies hairstyle content, and temporal self-attention blocks added to the U-Net supply inter-frame consistency. Training is staged so pose control is learned on bald images first, then hair transfer, then temporal smoothing.","core_discovery":"The paper's central claim is that a latent diffusion model can generate photorealistic, view-consistent hair transfer by learning in three stages: first a pose-controllable Latent IdentityNet learns to rotate bald heads using polar-azimuth pose embeddings; then a Hair Extractor injects reference hair features through cross-attention layers; and finally temporal attention layers smooth transitions between consecutive frames. Training data comes from a face-finetuned multi-view diffusion model that produces 21 views for each source and bald image, where the fine-tuning itself uses roughly 20,000 multi-view facial videos synthesized by a 3D-aware GAN. The paper reports quantitative gains over existing methods on single-view and multi-view hair transfer, and user-study wins on transfer accuracy, preservation, and naturalness.","pith_inferences":["If the synthetic-to-real gap is small, the same data pipeline could extend from front-facing views to full 360-degree views by adding more training viewpoints from the 3D GAN.","The staged training strategy may matter more than the specific hair domain; it could transfer to multi-view makeup, beard, or accessory editing where pose and attribute content compete.","The reported background-reconstruction weakness suggests a testable extension: explicitly conditioning on background features should reduce inter-frame flicker, measurable with frame-difference heatmaps.","One can probe whether the Hair Extractor truly learns hairstyle rather than identity by feeding reference images with identical hairstyles on very different faces and checking whether transfer remains invariant."],"forward_implications":["Hair styling for avatars and digital humans could be generated from a single reference photo with consistent results across viewpoints.","The staged recipe—pose control on bald heads first, then hair, then temporal smoothing—can be reused for other attribute-transfer tasks where pose and content compete.","The synthetic data pipeline removes the need for a large real multi-view hair dataset, making multi-view transfer practical where such data is unavailable.","End-to-end multi-view transfer improves consistency metrics over the baseline of single-view transfer followed by multi-view generation, establishing a benchmark for this task.","The temporal attention layer keeps head motion close to ground truth, a step toward video-level hair transfer."],"supporting_citations":[{"why":"Supplies the pretrained multi-view diffusion model that the paper face-finetunes and uses to generate multi-view training and evaluation image pairs.","marker":"[17]"},{"why":"Generates the roughly 20,000 multi-view facial videos used to fine-tune the multi-view diffusion model on human faces.","marker":"[64]"},{"why":"Provides the diffusion-based hair transfer approach that the bald converter follows and that Stable-Hair v2 extends to multiple views.","marker":"[14]"},{"why":"Provides the Non-Hair FFHQ dataset used to pre-train the diffusion-based bald converter.","marker":"[16]"},{"why":"Serves as the latent diffusion backbone for the multi-view hair transfer model.","marker":"[24]"},{"why":"Supplies the temporal attention module design that is inserted into the U-Net for inter-frame consistency.","marker":"[67]"},{"why":"Motivates the polar-azimuth pose conditioning used in the multi-view generation framework.","marker":"[57]"},{"why":"Used to extract 3DMM head-motion parameters for the pose-accuracy evaluation against ground truth.","marker":"[72]"}],"fun_headline_variants":["Stable-Hair v2: multi-view hair transfer diffusion","Diffusion model transfers hairstyles across viewpoints","Multi-view hair transfer with Stable-Hair v2","View-consistent hair swapping via diffusion model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole system rests on the assumption that synthetic multi-view face videos generated by a 3D GAN, and then used to fine-tune a multi-view diffusion model, produce pose-aligned and identity-preserving pairs accurate enough to train and evaluate a real-world hair transfer model, with no validation against real multi-view portrait captures.","fun_headline_variants_meta":{"raw":{"variants":["Stable-Hair v2: multi-view hair transfer diffusion","Diffusion model transfers hairstyles across viewpoints","Multi-view hair transfer with Stable-Hair v2","View-consistent hair swapping via diffusion model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000647,"raw_usage":{"total_tokens":2979,"prompt_tokens":957,"completion_tokens":2022,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":1961}},"tokens_in":573,"tokens_out":2022,"duration_ms":15365,"temperature":1.0,"reasoning_tokens":1961,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:36:41.413504+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render the fine-tuned multi-view model's 21 outputs for real FFHQ images and compare them against ground-truth multi-view portrait captures of the same people, measuring head-pose error and identity similarity; if the synthetic pairs show systematic pose or identity drift, the training-data foundation of the method collapses. A simpler check is to run the full pipeline on images with extreme or unusual viewpoints and count how often the multi-view output loses identity or produces inconsistent head shapes.","supporting_citations":[{"cited_title":"SV3D: Novel multi-view synthesis and 3D generation from a single image using latent video diffusion,","cited_arxiv_id":null,"evidence_quote":"Supplies the pretrained multi-view diffusion model that the paper face-finetunes and uses to generate multi-view training and evaluation image pairs."},{"cited_title":"Gram-hd: 3d-consistent image generation at high resolution with generative radiance manifolds,","cited_arxiv_id":null,"evidence_quote":"Generates the roughly 20,000 multi-view facial videos used to fine-tune the multi-view diffusion model on human faces."},{"cited_title":"Stable- hair: Real-world hair transfer via diffusion model,","cited_arxiv_id":null,"evidence_quote":"Provides the diffusion-based hair transfer approach that the bald converter follows and that Stable-Hair v2 extends to multiple views."},{"cited_title":"Hairmapper: Removing hair from portraits using gans,","cited_arxiv_id":null,"evidence_quote":"Provides the Non-Hair FFHQ dataset used to pre-train the diffusion-based bald converter."},{"cited_title":"Animatediff: Animate your personalized text-to- image diffusion models without specific tuning,","cited_arxiv_id":null,"evidence_quote":"Supplies the temporal attention module design that is inserted into the U-Net for inter-frame consistency."},{"cited_title":"Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set,","cited_arxiv_id":null,"evidence_quote":"Used to extract 3DMM head-motion parameters for the pose-accuracy evaluation against ground truth."}],"review_version":1}