{"id":"40432719-cf61-4766-93fe-7e92eb547bbb","arxiv_id":"2608.04622","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"DAC-Pose combines an LLM-based semantic reasoning agent and a discrepancy-aware visual encoding agent to generate humans in new poses, but the reported gains are tiny and the DAVE agent likely sees the target image at inference.","lead":"A new method for pose-guided human image editing uses two AI agents, one that writes text descriptions of body parts hidden in the source photo and one that encodes how the pose changed, to guide a diffusion model. The authors report small improvements on two standard benchmarks, but the method may use the target image at test time, which would invalidate the comparison.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The DAVE agent's use of the target image in Eq. (14)–(15) is either test-time leakage or a major unspecified gap, so the reported SOTA numbers do not validate the proposed paradigm.","rationale":"The reader's weakest_assumption and my load-bearing concern coincide: Eq. (14) makes the DAVE agent depend on the target image It, whose availability at test time is unaddressed. This is not a stylistic or novelty dispute; it is a direct threat to the validity of every quantitative result in Table 1 and the DAVE ablation in Table 2. If the ground truth leaks at inference, the reported gains over IMAGPose and PCDMs are not evidence for the agent-driven paradigm. The manuscript itself contains no limitation statement or training-only qualifier around Eqs. (14)–(17), and Section 4.2 presents the numbers as standard test-set evaluations. I checked whether the concern could be defused by the surrounding text: the DAVE description repeatedly says 'training,' but the method section never separates inference behavior; Section 3.3.1 says 'during training' for the registration, yet Eq. (14) is written as part of the method without a test-time substitute. Given the paper's decisive reliance on these experiments, the honest verdict is REJECT with high correctness risk, matching the reader's assessment. The test I propose directly settles the question because it compares the claimed numbers against a version of the pipeline that provably does not touch It at inference. I also considered whether the missing error bars and GPT-5.5 dependence alone would justify REJECT, and they reinforce the reproducibility problem; however, the target-image issue is the single load-bearing flaw: it affects the central claim's soundness, not merely its precision. My recommendation is UNCHANGED because the reader already reached REJECT and identified the same weakest assumption; I find no additional concern that would move the verdict in the other direction.","tokens_in":13475,"tokens_out":1601,"duration_ms":15828,"concrete_test":"Run the released code (or a minimal reimplementation) on a DeepFashion test split with the ground-truth target image hidden: replace b in Eq. (14) with the VAE encoding of a self-generated first-pass output or with a pose-derived condition, then recompute Table 1 and Table 2. If SSIM/LPIPS/FID change materially from the reported values, Eq. (14) leaks; if they do not, the paper must still specify the inference-time construction of b.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's DAVE agent computes the registered discrepancy encoding c from the source embedding a and the target-region embedding b = E_vae(It) in Eqs. (14)–(15). At inference time, It is precisely the unknown image the model must produce, so 'target-region embedding' derived from It is not available without access to the ground truth. The paper never states that Eq. (14) is used only during training, nor does it describe any alternative target embedding (e.g., from the target pose or from an intermediate generation) at test time. Since Section 4.2 reports DAC-Pose's final test-set numbers with no caveat, the central claim of Table 1 is unsupported: either the evaluation leaks the ground truth target into the discrepancy-aware condition, or a critical component of the test-time pipeline is left unspecified. Every downstream conclusion—including the ablation showing that removing DAVE costs 0.0266 SSIM—depends on this ambiguity. If Eq. (14) is training-only, the paper must also explain how DAVE's benefits persist at inference; as written, the method as described cannot be applied to a held-out target image.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes DAC-Pose, a dual-agent framework for pose-guided human image generation. The Prior Semantic Reasoning (PSR) agent uses an MLLM to generate textual semantic priors for unseen or view-shifted regions, while the Discrepancy-Aware Visual Encoding (DAVE) agent computes a discrepancy encoding between source and target embeddings and fuses it into a denoising diffusion U-Net. The authors report state-of-the-art SSIM, LPIPS, and FID on DeepFashion and Market-1501, along with ablations showing that both agents contribute. I read the paper with particular attention to the test-time protocol for the DAVE agent, since the discrepancy encoding appears to be computed from the target image.","tokens_in":13697,"tokens_out":4950,"duration_ms":54101,"significance":"The agent-based framing, especially the combination of explicit spatial constraints and textual semantic priors, is timely and could be useful for extreme pose transfer. The paper also promises code release, which is a positive feature for reproducibility. However, the central quantitative claim is not supported as written because the DAVE agent's discrepancy encoding is computed from the target image, and the reported gains over the strongest baseline are extremely small and lack any uncertainty quantification. If the target-image dependence were resolved and the evaluation repeated with proper statistical testing, the framework could still be of interest, but the current evidence does not establish state-of-the-art performance.","major_comments":[{"comment":"The target image It enters the DAVE pipeline as b = E_vae(It) and is used to form c = CA(a, b, b) and then g = CA(c, d, d). At inference time, It is precisely the unknown image that the model is supposed to produce, yet the paper does not state that Eq. (14) is training-only, nor does it specify how b or c is obtained at test time (for example, from the target pose, from an intermediate generation, or from a separate estimator). The Section 4.2 results and Table 2 ablations are reported without any such caveat, so the reported metrics are consistent with the model being conditioned on the ground-truth target. This is a load-bearing gap: either the evaluation leaks information from the ground truth, or a critical part of the test-time architecture is unspecified. The authors must clarify the exact training and inference protocols and, if necessary, re-evaluate the method without target-image information.","section":"Section 3.3.1, Eqs. (14)-(16)"},{"comment":"The reported advantage over the strongest prior work is numerically very small: on DeepFashion, SSIM is 0.7572 versus 0.7561, LPIPS is 0.1274 versus 0.1284, and FID is 5.8547 versus 5.8738. No error bars, confidence intervals, or repeated-run statistics are provided, and no significance test is performed. Because diffusion sampling is stochastic and the pipeline includes LLM-based description generation, these differences are within plausible run-to-run variation. The headline claim of consistent superiority therefore needs a statistical foundation before it can be accepted.","section":"Section 4.2, Table 1"},{"comment":"The target pose map is defined as Pt = Epose(It), using the target image itself. If the target pose is part of the task input, the authors should say so explicitly and distinguish it from any later use of It in DAVE. If the target pose is derived from the ground-truth target image during evaluation, that is standard for this benchmark but should be stated clearly. The current wording blurs the boundary between the provided pose condition and the unknown target appearance, compounding the concern raised in the first major comment.","section":"Section 3.2.1, Eq. (1)"},{"comment":"Several components are left unspecified. The viewpoint keypoint weights omega_m are described only as satisfying a sum-to-one constraint; the quantization function Q_view in Eq. (6) is not defined; and the reliability analysis function F_rel in Eq. (12) has no description at all. These are not cosmetic details: they determine the spatial semantic constraint C_spa that drives the PSR agent. Without specifying them, the method cannot be reproduced from the manuscript, and the paper's reproducibility claims are weakened.","section":"Section 3.2.1, Eq. (4), and Section 3.2.2, Eq. (12)"}],"minor_comments":[{"comment":"The term 'DA VE' is typeset with an unusual space in many places; please use 'DAVE' consistently.","section":"Throughout"},{"comment":"The sentence beginning 'Despite their notable success under mild pose variations...' appears twice in near-identical form at the end of the subsection; please remove the duplicate.","section":"Section 2.1"},{"comment":"Figure 2 shows a 'Target Image' input and a 'Source Reconstruction Loss' that are not explained in the main text; these should be explicitly reconciled with the training and inference descriptions.","section":"Figure 2"},{"comment":"The paper names GPT-5.5 as the multimodal foundation model but gives no version, decoding parameters, or prompting scheme; please provide these details or a reference to a reproducible configuration.","section":"Section 4.1"}],"recommendation":"reject","confidential_remarks":"For the editor: the target-image leakage in the DAVE module is a direct correctness problem, not a disagreement with community consensus. I do not view this as a borderline case. The paper could warrant a fresh look if the authors demonstrate a valid test-time protocol and rerun the comparisons with error bars, but as it stands the central claim is unsupported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. It is a serious attempt to combine an LLM-based semantic reasoning agent with a discrepancy-aware visual encoder for pose-guided human generation, and it is clearly written, with extensive comparisons and ablations. The problem framing—cross-view pose inversion versus out-of-view body completion—is useful, and the qualitative results are plausible. Credit is also due for benchmarking against a wide range of baselines and for including an ablation that separates the PSR and DAVE components.\n\nNow the soft spot, and it is load-bearing. The DAVE agent's discrepancy encoding is computed from the target image It via Eq. (14)–(15). At inference, It is exactly the image the model is supposed to produce. The paper says the registration is performed \"during training,\" but then feeds the resulting c into the denoising fusion in Eq. (16), which is used in generation. No alternative target embedding is described for test time. So either the reported test numbers leak the ground truth into the condition, or a critical part of the pipeline is left unspecified. The ablation showing that removing DAVE costs 0.0266 SSIM depends entirely on this ambiguity. This is not a minor glitch; it undermines the central claim.\n\nBeyond that, the metric gaps over IMAGPose are tiny—0.0011 SSIM, 0.001 LPIPS, 0.019 FID—and reported without error bars or significance tests. The dependence on a proprietary GPT-5.5 API without versioning or prompt details further hurts reproducibility. The reliability analysis Frel is also undefined. These are secondary but real issues.\n\nI agree with the reader's take. The paper deserves a serious referee because the DAVE ambiguity might be fixable (e.g., restricting the discrepancy encoding to training only, or using a principled test-time proxy), and the topic is timely. But as written, the central result is unsupported, and the paper should not be accepted in its current form.","headline":"The DAVE agent's use of the target image at inference is either test-time leakage or an unspecified gap, and the thin metric gains do not compensate.","tokens_in":14242,"tokens_out":1805,"would_cite":false,"duration_ms":21791,"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":"DAC-Pose claims that a collaborative dual-agent system—one agent reasoning in language about unseen body regions, the other encoding pose-induced spatial discrepancies—improves pose-guided human image generation and beats all compared…","keywords":["pose-guided person image generation","AI agents","semantic reasoning","discrepancy-aware visual encoding","latent diffusion","cross-view pose inversion","out-of-view body completion","DeepFashion"],"falsifier":"Run the released DAC-Pose inference code on a hold-out DeepFashion split and check whether the VAE encoder ever receives the ground-truth target image $I_t$: if it does, the reported gains may measure target leakage rather than agent reasoning. A cleaner test replaces the target embedding in Eq. (14) with a prediction estimated from the source image and target pose; if SSIM and FID fall to roughly the \"without DAVE\" ablation levels, the discrepancy encoding is not performing the claimed work.","tokens_in":13237,"feed_emoji":"🧍","tokens_out":5541,"duration_ms":61644,"temperature":0.7,"pith_summary":"The paper claims that pose-guided human image generation can be improved by replacing a single passive pipeline with a collaborative pair of agents: one that reasons in language about body regions invisible in the source view, and one that explicitly measures the spatial mismatch between source and target poses. It says this dual-agent design lets a diffusion model synthesize clothing patterns, identity features, and newly exposed body parts under drastic viewpoint changes instead of hallucinating them. On DeepFashion and Market-1501, the reported numbers show DAC-Pose ahead of all fourteen comparison methods on SSIM, LPIPS, and FID. The reason this matters is that virtual try-on, digital avatars, and film production depend on preserving appearance when the pose changes radically.","feed_headline":"Two-agent framework tops 14 methods on pose-guided human generation","feed_subtitle":"Semantic reasoning plus mismatch-aware visual encoding improves all three metrics on two standard benchmarks.","key_machinery":"The central machinery is the collaborative dual-agent loop. The PSR agent converts pose pairs into spatial semantic constraints (a yaw-shift score derived from keypoint distances around a nose-neck reference axis, plus an LLM-detected extended body region) and then into three MLLM-generated text descriptions fused by reliability analysis into a conditioning embedding. The DAVE agent computes a discrepancy encoding $c = \\text{CA}(a,b,b)$ from source and target VAE embeddings and fuses it as $g = \\text{CA}(c,d,d)$ to steer the denoising UNet. The paper's argument is that semantic deduction supplies what the source view does not show, while discrepancy encoding supplies where the two views differ, and that both are needed.","core_discovery":"DAC-Pose reformulates single-view human generation as a dual-agent system. The Prior Semantic Reasoning agent uses DWPose skeletons to estimate yaw shift and detect expanded body parts, then has a multimodal LLM produce three text descriptions (source appearance, extended region, perspective difference) that are merged by reliability analysis into a semantic embedding. The Discrepancy-Aware Visual Encoding agent encodes source and target images with a frozen VAE and computes a discrepancy encoding via cross-attention, then fuses it with the reference-branch feature to condition the denoising UNet. With these two conditions, the paper reports the best SSIM, LPIPS, and FID on DeepFashion and Market-1501, and ablations show removing either agent degrades the numbers, with DAVE removal causing the larger drop.","pith_inferences":["If the inference-time DAVE gap is closed by predicting the target embedding from the source image and target pose, the reported margin over IMAGPose is likely to shrink; the paper's current formulation leaves this unresolved.","The PSR agent's quality depends on the multimodal LLM's descriptions, so a test that feeds oracle or ground-truth captions could separate the benefit of textual priors from the benefit of the fusion machinery.","The same dual-agent pattern could transfer to video human animation or virtual try-on, where large viewpoint changes are the dominant failure mode, but this is an extension the paper does not demonstrate."],"forward_implications":["DAC-Pose reports the best SSIM, LPIPS, and FID on both DeepFashion and Market-1501, so an agent-based conditioning design can compete with and slightly surpass strong diffusion baselines.","Removing either agent lowers measured performance, which implies the semantic priors and discrepancy constraints are complementary rather than redundant.","Because DAVE removal hurts more than PSR removal, spatial discrepancy encoding carries a larger share of the visual-fidelity improvement under the paper's settings.","The framework generalizes across resolutions and environments, since gains appear on 256×176 DeepFashion and 128×64 Market-1501."],"supporting_citations":[{"why":"Supplies the Stable Diffusion latent diffusion backbone that DAC-Pose builds its denoising process on.","marker":"[9]"},{"why":"PCDMs is the strongest pure-vision diffusion baseline that DAC-Pose compares against and clearly exceeds on FID.","marker":"[4]"},{"why":"IMAGPose is the closest quantitative competitor whose SSIM, LPIPS, and FID DAC-Pose edges out.","marker":"[12]"},{"why":"CFLD is a coarse-to-fine latent diffusion baseline representing the texture-fidelity approach DAC-Pose extends.","marker":"[5]"},{"why":"DWPose extracts the source and target skeletons that feed the PSR agent's spatial geometry analysis.","marker":"[32]"},{"why":"Animate Anyone provides the lightweight Pose Guider used as the PoseNet conditioning module.","marker":"[28]"},{"why":"PIDM is a denoising-diffusion baseline whose identity and texture handling DAC-Pose compares against.","marker":"[13]"},{"why":"PoCoLD is a pose-constrained latent diffusion baseline representing the mapping-relationship reasoning paradigm.","marker":"[11]"}],"fun_headline_variants":["DAC-Pose: dual agents collaborate for better human generation","Two-agent AI framework improves pose-guided human generation","Semantic and visual agents beat prior art in pose transfer","DAC-Pose outperforms 14 methods on two benchmarks","Agent-driven approach excels at drastic viewpoint changes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The DAVE agent's discrepancy encoding is computed from the target image $I_t$, but at inference the target image is the unknown output, and the paper does not say how the target embedding is obtained at test time.","fun_headline_variants_meta":{"raw":{"variants":["DAC-Pose: dual agents collaborate for better human generation","Two-agent AI framework improves pose-guided human generation","Semantic and visual agents beat prior art in pose transfer","DAC-Pose outperforms 14 methods on two benchmarks","Agent-driven approach excels at drastic viewpoint changes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000695,"raw_usage":{"total_tokens":3139,"prompt_tokens":936,"completion_tokens":2203,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":2125}},"tokens_in":552,"tokens_out":2203,"duration_ms":18640,"temperature":1.0,"reasoning_tokens":2125,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:26:33.221863+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released DAC-Pose inference code on a hold-out DeepFashion split and check whether the VAE encoder ever receives the ground-truth target image $I_t$: if it does, the reported gains may measure target leakage rather than agent reasoning. A cleaner test replaces the target embedding in Eq. (14) with a prediction estimated from the source image and target pose; if SSIM and FID fall to roughly the \"without DAVE\" ablation levels, the discrepancy encoding is not performing the claimed work.","supporting_citations":[{"cited_title":"Rombach, A","cited_arxiv_id":null,"evidence_quote":"Supplies the Stable Diffusion latent diffusion backbone that DAC-Pose builds its denoising process on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"PCDMs is the strongest pure-vision diffusion baseline that DAC-Pose compares against and clearly exceeds on FID."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"IMAGPose is the closest quantitative competitor whose SSIM, LPIPS, and FID DAC-Pose edges out."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CFLD is a coarse-to-fine latent diffusion baseline representing the texture-fidelity approach DAC-Pose extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DWPose extracts the source and target skeletons that feed the PSR agent's spatial geometry analysis."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Animate Anyone provides the lightweight Pose Guider used as the PoseNet conditioning module."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"PIDM is a denoising-diffusion baseline whose identity and texture handling DAC-Pose compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"PoCoLD is a pose-constrained latent diffusion baseline representing the mapping-relationship reasoning paradigm."}],"review_version":1}