{"id":"c3b9e436-f948-44aa-abb2-1db9e3f301c4","arxiv_id":"2505.05101","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"MDE-Edit adds two mask-guided attention losses to diffusion inference, improving object placement and color localization in multi-object image edits.","lead":"This paper proposes a training-free method for editing several objects in one photo at once, using two attention-based losses to keep each change inside its own region. It reports better text alignment and background preservation than four prior methods on multi-object benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LoMOE, the method behind LoMOE-bench, is absent from Table 1, so the 'outperforms SOTA' claim is untested against its most relevant baseline; a direct comparison would settle it.","rationale":"The reader's verdict is CONDITIONAL and already lists the missing LoMOE comparison among its concerns, but the reader's stated weakest assumption is the reliability of cross-attention maps as segmentation for OAL. That mechanism-level concern is real and untested, yet the paper's ablations do show that OAL and CCL improve results relative to their absence; the mechanism may be imperfect but is not directly disproven. My stress-test focuses on the strongest form of the paper's central claim, which is empirical superiority over SOTA. Omitting LoMOE from a benchmark named after that method is a direct threat: it is the one baseline most likely to be competitive on localized multi-object editing, and its absence means the headline claim is currently unsubstantiated on the benchmark designed for the task. This omission is fixable by running the missing comparison, so it does not warrant rejection, but it is precisely the kind of evidence gap that keeps the verdict at CONDITIONAL. The proposed concrete test is a head-to-head LoMOE comparison with proper statistics; if it passes, the central claim gains strong support, and if it fails, the claim must be revised.","tokens_in":12511,"tokens_out":10034,"duration_ms":108165,"concrete_test":"Run the official LoMOE implementation on the exact LoMOE-bench subset used for Table 1, with the same source/target prompts, SAM-generated edit masks, and metric scripts (CLIP Score, BG-LPIPS, BG-SSIM). Recompute all rows of Table 1 on this subset and compare. If LoMOE attains BG-LPIPS <= 0.086, BG-SSIM >= 0.936, or CLIP Score >= 0.290 on the overlap scenarios, the claim that MDE-Edit outperforms state-of-the-art methods is not supported. Also run each method and metric over at least three random seeds and report mean +/- std to determine whether the observed differences are statistically meaningful.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that MDE-Edit 'outperforms state-of-the-art methods in editing accuracy and visual quality' (Abstract, Section 5.2), supported by Table 1. Yet Section 5.1 says evaluation uses both OIR-bench and LoMOE-bench, while the comparison set in Section 5.2 is NTI+P2P, CDS, GnR, and OIR. LoMOE (Chakrabarty et al., ACM MM 2024) is never included, even though it is the method that introduced LoMOE-bench and is specifically designed for localized multi-object editing—the exact task MDE-Edit claims to advance. Thus the benchmark most targeted at the paper's contribution is evaluated without its originating method, leaving the strongest relevant baseline untested. If LoMOE achieves better CLIP Score, BG-LPIPS, or BG-SSIM on this benchmark, the headline conclusion fails. Reported numbers also lack error bars or significance tests, so even the apparent margins (e.g., BG-LPIPS 0.111 vs 0.106 for OIR) may not be robust. This is a load-bearing evidence gap: it concerns the central empirical claim, not a peripheral detail.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MDE-Edit, a training-free inference-stage optimization method for multi-object image editing with Stable Diffusion. The method uses two losses: Object Alignment Loss (OAL), which aligns averaged multi-layer cross-attention maps with SAM segmentation masks to enforce spatial accuracy, and Color Consistency Loss (CCL), which boosts edit-token attention within masks while suppressing leakage to other regions. The latent is updated by gradient descent inside the editing masks during the initial denoising steps. Experiments on OIR-bench and LoMOE-bench compare against NTI+P2P, CDS, GnR, and OIR, reporting higher CLIP Score, lower BG-LPIPS, and higher BG-SSIM; ablations show that both losses contribute to the final performance.","tokens_in":12812,"tokens_out":2733,"duration_ms":26697,"significance":"If the empirical claims hold, MDE-Edit is a plausible and lightweight alternative for localized multi-object editing, decoupling structural localization from appearance control in a training-free scheme. The dual-loss design is well motivated, and the ablations show consistent trends in the intended direction. The paper addresses a real limitation of existing mask-based and attention-based methods. However, the evidence for the central 'outperforms SOTA' claim is currently incomplete: the most relevant baseline (LoMOE) is missing from the comparison on its own benchmark, and the reported numbers are presented without error bars or significance testing. These issues are essential to verify before the performance claim can be accepted.","major_comments":[{"comment":"The paper evaluates on LoMOE-bench (Section 5.1) but omits LoMOE itself (Chakrabarty et al., ACM MM 2024) from the comparison set in Table 1. LoMOE is the method that introduced this benchmark and is specifically designed for localized multi-object editing, the exact task MDE-Edit claims to advance. Omitting it leaves the most relevant baseline untested; the claim of outperforming state-of-the-art methods is therefore not substantiated on this benchmark. Include LoMOE in Table 1 or explicitly justify its exclusion.","section":"Section 5.2, Table 1"},{"comment":"No error bars, confidence intervals, or significance tests are reported for any metric. Some reported margins are small (e.g., non-overlap BG-LPIPS 0.106 for MDE-Edit versus 0.111 for OIR), so it is unclear whether the improvements are statistically meaningful. The text states that the quantitative evaluation 'conclusively demonstrates' superiority, which is not supported without repeated runs and appropriate statistical testing.","section":"Section 5.1, Table 1"},{"comment":"OAL depends on the assumption that averaged cross-attention maps (averaged over layers and heads) align with SAM masks at the same spatial resolution. The paper does not specify how masks are resized to the cross-attention resolution, nor does it analyze how reliably this implicit segmentation localizes small or heavily overlapping objects. Since a wrong target in Eq. (5) would propagate errors into the latent update, this assumption is load-bearing. Please provide implementation details for mask preprocessing and an analysis of localization accuracy, at least for the overlapping-object cases shown in Figure 5.","section":"Section 4.2.1, Eq. (4)-(5)"}],"minor_comments":[{"comment":"The notation in Eq. (8) is ambiguous: z*_t appears on both sides of the equation. Clarify that the gradient update is computed on the editing-branch latent and then merged with the reconstruction-branch latent outside the mask.","section":"Section 4.3, Eq. (8)"},{"comment":"The text says 'Our code can be accessed in the supplementary material,' but no code repository, appendix, or supplementary document is included in the manuscript. Please add a concrete link or describe how the code can be obtained.","section":"Section 5.1"},{"comment":"LoMOE [3] is cited in the references and used as a benchmark, but it is not discussed in the related-work section. A brief discussion of its relationship to MDE-Edit would help position the contribution.","section":"Section 2, Related Work"},{"comment":"The source and target prompt text in the qualitative comparison figures is very small and difficult to read. Increasing the font size or placing the prompts in a separate readable table would improve clarity.","section":"Figures 4 and 5"},{"comment":"The qualitative descriptions of baseline failures (e.g., 'OIR often results in blurred details', 'CDS tends to affect neighboring regions') are not tied to specific examples in the figures. Please refer to specific rows or regions in Figures 4 and 5 to make these claims verifiable.","section":"Section 5.2, Paragraph 2"}],"recommendation":"major_revision","confidential_remarks":"The missing LoMOE baseline and the lack of significance testing are the main barriers to accepting the performance claim. If the authors can add a direct comparison against LoMOE and report variance over multiple runs (or at least provide evidence of statistical robustness), the paper would be substantially stronger. The code availability statement should also be made concrete."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: MDE-Edit is a credible, training-free multi-object editing method. The new bit is the pairing of an object-alignment loss (OAL) that fits averaged multi-layer cross-attention maps to SAM masks with a color-consistency loss (CCL) that amplifies edit-token attention inside masks and suppresses leakage. That combination is not in OIR, CDS, GnR, or NTI+P2P, and the ablations in Table 2 show each loss moves the metrics in the advertised direction. The qualitative figures are consistent with the story. So the core mechanism holds up on its own terms.\n\nThe soft spots are real and mostly evidentiary. The stress-test concern is correct: Section 5.1 says evaluation uses both OIR-bench and LoMOE-bench, but Table 1 compares only against NTI+P2P, CDS, GnR, and OIR. LoMOE is the method that introduced LoMOE-bench and is purpose-built for localized multi-object editing, so leaving it out leaves the strongest baseline for this exact task untested. The 'outperforms state-of-the-art' claim cannot stand without that comparison.\n\nAlso missing: error bars or significance tests. The gaps are sometimes small (BG-LPIPS 0.111 vs 0.106 in non-overlap), and without variance we don't know if OIR is actually behind. The paper says code is in the supplementary material, but the submission has no supplementary material, so that claim is unfulfilled. A few implementation details are under-specified: how SAM masks are resized to cross-attention resolution, and the exact optimization schedule (iterations per timestep, step size, number of timesteps optimized beyond the stated 20). These are addressable in revision.\n\nOn the circularity question: OAL and CCL directly push the attention maps toward the targets the method claims to control, so the ablations partly measure fitting to the objective. That said, the reported metrics (CLIP Score, BG-LPIPS, BG-SSIM) are external to the losses, so the circularity is not fatal; it just means the ablation numbers need to be read with that in mind.\n\nWho this is for: people working on attention-based editing and inference-stage control of diffusion models. They will find the loss formulation and the ablations useful. I would not yet cite the SOTA claim, but I would engage with the method.\n\nRecommendation: send it to peer review, but with a specific request to add the LoMOE baseline on LoMOE-bench, provide error bars, and release the code. If those are supplied and the comparison holds, this is a solid incremental contribution.","headline":"Plausible dual-loss editing method with internally consistent ablations, but the missing LoMOE baseline and lack of error bars undermine the SOTA claim until fixed.","tokens_in":13319,"tokens_out":2964,"would_cite":false,"duration_ms":29970,"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":"MDE-Edit claims that a training-free dual-loss optimization of the diffusion latent—aligning cross-attention to masks and amplifying attribute attention inside them—enables localized multi-object edits that improve both semantic alignment…","keywords":["multi-object image editing","diffusion models","cross-attention control","text-guided editing","masked latent optimization","object alignment loss","color consistency loss","training-free editing"],"falsifier":"Measure the overlap (IoU) between the averaged attention maps and the SAM masks on a set of small, overlapping objects; if optimizing OAL fails to improve editing quality precisely in the low-overlap cases, the localization claim is falsified.","tokens_in":12281,"feed_emoji":"🎨","tokens_out":5440,"duration_ms":49857,"temperature":0.7,"pith_summary":"This paper claims that a training-free, inference-stage optimization can make text-guided diffusion models edit several objects at once without the usual failures of attention misalignment and color bleeding. The proposed MDE-Edit adds two losses to the noise-latent denoising loop: Object Alignment Loss (OAL), which pulls layer-averaged cross-attention maps toward per-object segmentation masks, and Color Consistency Loss (CCL), which forces the attention of a new attribute token (for example \"red\") to concentrate inside the target mask. On non-overlapping and overlapping multi-object benchmarks, the paper reports higher CLIP Score and better background preservation than NTI+P2P, CDS, GnR, and OIR, with the largest margins on overlapping scenes. If the claim holds, precise multi-object editing needs no retraining or model weight changes, only gradient updates to the latent during sampling.","feed_headline":"MDE-Edit beats prior methods on multi-object edit benchmarks","feed_subtitle":"Training-free dual losses align attention to masks and bind colors tightly, lifting CLIP score and background fidelity.","key_machinery":"The load-bearing mechanism is the cross-attention map itself, extracted from the U-Net of Stable Diffusion and averaged over layers and heads. OAL treats that averaged map as a soft segmentation and matches it to a hard SAM mask with a dual BCE loss; CCL uses the same maps to compute a ratio between the edit token's attention and the total attention of shared tokens inside the mask. The combined gradient updates the latent noise only inside edit masks (Eq. 8), while attention injection from the reconstruction branch (Eq. 3) keeps shared-token structure intact.","core_discovery":"The central discovery is that the two failure modes of multi-object editing—wrong location and attribute leakage—can be decoupled and fixed separately at inference time. OAL treats the averaged cross-attention map for each edited token as an implicit segmentation and aligns it to the SAM mask with two binary cross-entropy terms, one on the raw map and one on its scale-normalized version. CCL maximizes, inside the mask, the attention of the new attribute token relative to the sum of all shared-token attentions, which suppresses color leaking into adjacent objects. The latent is then updated by gradient descent on the combined loss, and Eq. (8) re-blends the edited latent with the reconstruction latent outside the masks. In Table 1 the method reaches CLIP Score 0.282 (non-overlap) and 0.290 (overlap) versus best baselines 0.278 and 0.267, while improving BG-SSIM to 0.925/0.936 and BG-LPIPS to 0.106/0.086.","pith_inferences":["As an extension beyond the paper, OAL could be fed attention-derived masks instead of SAM masks, which would remove the external segmentation dependency; the paper does not test this.","As an extension beyond the paper, CCL's ratio formulation suggests a natural test on attribute swaps other than color (texture, pose, material), where attention leakage may behave differently.","As an extension beyond the paper, the method's dependence on mask quality means accuracy on small or heavily occluded objects is an open question; the paper does not quantify attention-mask overlap."],"forward_implications":["Because OAL and CCL act only on intermediate attention maps and the latent, the same recipe can be attached to any Stable-Diffusion-based editor that already provides an inversion trajectory.","The mask-blended gradient update in Eq. (8) keeps unedited regions untouched by construction, which explains the reported BG-LPIPS and BG-SSIM gains.","The dual-loss separation means object replacement and attribute recoloring can be scheduled with different loss weights, giving users independent control over geometry and appearance.","If the Table 1 numbers transfer to other backbones, the approach offers a generic inference-time plug-in for multi-object editing without retraining."],"supporting_citations":[{"why":"Provides the latent diffusion backbone and the U-Net whose cross-attention maps OAL and CCL consume.","marker":"[25]"},{"why":"Supplies the null-text inversion that maps the input image to the noise latent edited by MDE-Edit.","marker":"[18]"},{"why":"Generates the per-object segmentation masks used as OAL targets and as the blend mask in Eq. (8).","marker":"[12]"},{"why":"Defines the multi-object editing task and one baseline; its OIR-bench is used for evaluation.","marker":"[34]"},{"why":"Provides the NTI+P2P baseline and the cross-attention injection idea that Eq. (3) extends.","marker":"[9]"},{"why":"Provides the CDS baseline that must be beaten in the multi-object comparison.","marker":"[19]"},{"why":"Provides the GnR baseline that must be beaten in the multi-object comparison.","marker":"[30]"},{"why":"Defines CLIP Score, the semantic alignment metric reported in Table 1.","marker":"[10]"},{"why":"Defines LPIPS, used as the background perceptual fidelity metric BG-LPIPS.","marker":"[35]"},{"why":"Defines SSIM, used as the background structural fidelity metric BG-SSIM.","marker":"[33]"}],"fun_headline_variants":["Training-free dual-loss editing beats prior multi-object methods","Dual-loss inference optimization stops color bleeding in multi-object edits","MDE-Edit: align SAM masks and bind colors without training","Zero-training editing: dual losses untangle multi-object attention","MDE-Edit sets new benchmarks on multi-object editing accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that what the diffusion model pays attention to, averaged across its layers and heads, reliably marks where each object is in the image, even for small or overlapping objects.","fun_headline_variants_meta":{"raw":{"variants":["Training-free dual-loss editing beats prior multi-object methods","Dual-loss inference optimization stops color bleeding in multi-object edits","MDE-Edit: align SAM masks and bind colors without training","Zero-training editing: dual losses untangle multi-object attention","MDE-Edit sets new benchmarks on multi-object editing accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000629,"raw_usage":{"total_tokens":2941,"prompt_tokens":1013,"completion_tokens":1928,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":629,"completion_tokens_details":{"reasoning_tokens":1844}},"tokens_in":629,"tokens_out":1928,"duration_ms":15352,"temperature":1.0,"reasoning_tokens":1844,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:12:32.039063+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the overlap (IoU) between the averaged attention maps and the SAM masks on a set of small, overlapping objects; if optimizing OAL fails to improve editing quality precisely in the low-overlap cases, the localization claim is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the CDS baseline that must be beaten in the multi-object comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the GnR baseline that must be beaten in the multi-object comparison."}],"review_version":1}