{"id":"2435115d-5a71-4a85-aa79-7d4a5a82b40a","arxiv_id":"2502.04050","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"PartEdit trains part-specific text tokens to produce spatial masks in a frozen diffusion model, enabling localized text-based part edits.","lead":"PartEdit learns special text tokens that make a frozen diffusion model localize object parts such as heads, hoods, and seats, then uses those localizations to blend in text-described edits. The authors introduce a small part-editing benchmark and report that their method beats existing text-based editors in user studies 66 to 90 percent of the time.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Token localization on under-trained parts (car-hood, chair-seat) is unvalidated; the 'all metrics' claim may rest on averaging over parts that localize well.","rationale":"The paper's internal pipeline is coherent: a separate localization path keeps the part token's context fixed to [part, BG] padding, so the mask is not corrupted by edit-word context; Eq. 3-6 form a consistent method. The critical weakness is empirical rather than logical: the localization quality of the trained tokens on the evaluation benchmark is never quantified, and it is the sole input to the blending mask. The reader's weakest_assumption identifies exactly this transfer and generalization risk, and I concur. My review adds a concrete, checkable specificity: several part tokens are trained on only 10 custom images, and no per-part mIoU is reported for SDXL on the benchmark. Because Table 1 averages across parts, a single poorly-localizing part (e.g., car-hood or chair-seat) could inflate failure while the average still looks competitive. This does not invalidate the method; it means the central claim's evidence is conditional on localization quality that the paper does not demonstrate. Therefore the reader's CONDITIONAL verdict should stand unchanged.","tokens_in":102,"tokens_out":5480,"duration_ms":61636,"concrete_test":"Compute the mIoU between each trained part token's aggregated attention mask (as used in Eq. 4-5) and the manual ground-truth masks on all 60 PartEdit-Synth images, per part and per inference timestep. Then recompute Table 1's averages after excluding any part with mIoU below a chosen threshold (e.g., 50). If one or more parts fail localization, the headline 'all metrics' claim is an artifact of averaging across parts.","verdict_should_be":"UNCHANGED","load_bearing_attack":"PartEdit's central claim depends on the optimized part tokens producing reliable cross-attention localization on unseen images and across the benchmark's prompt distribution. This is load-bearing because every result in Table 1 is an average over 7 part types, several of which (car-hood, chair-seat) are trained from only 10 custom-annotated images (Appendix I), and the paper reports no per-part or absolute mask-quality metric on the evaluation set. Appendix J reports mIoU only for quadruped head under SD2.1 (71.7), not for the SDXL tokens or the other 6 parts. If localization fails for a part (e.g., car-hood with varied geometry and viewpoint), the feature blending in Eq. 6 edits the wrong region, yet the failure can be masked by the average over 60 images. The claim 'outperforms existing editing methods on all metrics' is therefore not robustly established without per-part localization validation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"PartEdit proposes a text-based image editing method for object parts using a frozen pre-trained diffusion model (SDXL). The key idea is to optimize a small number of textual part tokens (e.g., , ) by supervising their cross-attention maps with binary cross-entropy against part segmentation masks (Eq. 3). At inference, the optimized tokens are used in an edit prompt, their cross-attention maps are aggregated across layers (Eq. 4), thresholded adaptively with an Otsu-based scheme (Eq. 5), and used to blend features between the source and edit diffusion trajectories (Eq. 6). The paper introduces a new part-editing benchmark (PartEdit-Synth and PartEdit-Real), compares against several text-based and mask-based editing methods, reports quantitative metrics and user studies, and releases code and data. The central claims are that learned part tokens provide reliable localization, that the feature-blending pipeline performs seamless localized edits, and that PartEdit outperforms existing methods on all metrics.","tokens_in":21728,"tokens_out":4153,"duration_ms":48256,"significance":"If the claims hold, PartEdit is a practically useful contribution: it enables fine-grained part edits from text alone, without user-supplied masks, while keeping the base diffusion model frozen. The method is clearly specified, the equations for token optimization, mask aggregation, adaptive thresholding, and feature blending are easy to follow, and the ablations (timestep selection, layer selection, thresholding, padding, number of training images) are informative. The released code, benchmark, and evaluation protocol are valuable assets for follow-up work. The main risk is that the entire editing pipeline depends on the cross-attention localization maps transferring reliably to unseen images and prompts, and that risk is not directly quantified for most of the evaluated parts.","major_comments":[{"comment":"The central claim that PartEdit outperforms existing methods depends on the learned tokens producing reliable localization masks on unseen images, because the feature blending of Eq. (6) edits whatever region the tokens localize. However, the paper does not report per-part localization accuracy for the tokens used in the main experiments. Table 1 reports averages over 60 synthetic and 13 real images across seven part types, and Appendix J reports mIoU only for the quadruped-head token and only under SD2.1, not for SDXL. This is a load-bearing gap: parts such as car-hood and chair-seat are trained from only 10 custom-annotated images, and if their localization maps are poor, the edit will be applied to the wrong region while the averaged metrics can still look strong. Please add per-part localization metrics (e.g., mIoU against the annotated masks) for all seven tokens on the evaluation sets, report the per-part breakdown of the edit metrics in Table 1, and discuss which parts drive the aggregate gains.","section":"Section 4.3, Appendices I and J"},{"comment":"The paper repeatedly states that PartEdit 'outperforms existing editing methods on all metrics,' but Table 1 reports single point estimates without confidence intervals, standard deviations, or significance tests. This matters especially for the real-image benchmark, which contains only 13 images, and for the user studies: a 360-response 2AFC study can give a useful aggregate, but reporting a bare percentage such as 66% or 66.92% without an interval makes it impossible to tell whether the preference over that baseline is distinguishable from chance. Please report variability (per-image or bootstrap CIs) and, where feasible, significance tests for the quantitative metrics and user-study preferences, or temper the 'all metrics' claim accordingly.","section":"Table 1 and Section 4.3"},{"comment":"The editing pipeline assumes that the attention mask M_t from Eq. (4) is informative at every denoising step used for blending. The paper shows qualitative attention visualizations and ablations of the thresholding scheme, but it does not validate mask quality as a function of timestep or across different prompts and objects for the actual SDXL tokens. Since Eq. (6) is applied for all timesteps in [1, t_e] and t_e=50 in the default synthetic setting, an inaccurate mask at an early or late timestep can directly corrupt the output. Please include a quantitative analysis of mask quality over the blending horizon, or at least per-part failure cases, so that the reader can assess how often localization transfers to the benchmark distribution.","section":"Section 3.3, Eq. (6)"}],"minor_comments":[{"comment":"The notation says that for timestep t the mask M_t aggregates attention maps from the previous step t-1, but the indexing is not fully specified for t=1; please clarify the boundary case and align the notation with the inference loop described in the text.","section":"Eq. (4)"},{"comment":"There is a typo in the second paragraph: 'LAIOB-5B' should be 'LAION-5B'.","section":"Introduction"},{"comment":"The caption says 'feature bending' but the method performs feature blending; please correct the caption to match the terminology used in Eq. (6).","section":"Figure 3 caption"},{"comment":"The user-study instructions are reported in a quotation that appears to have an extra unmatched quotation mark; please check the formatting of the quoted instruction text.","section":"Appendix H"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a graphics/vision venue and the method is plausible, but the evaluation currently does not directly validate the load-bearing localization assumption for most part tokens. The missing per-part metrics and statistical support are fixable in revision, so I recommend major revision rather than rejection. I would also encourage the editor to ask the authors to make explicit how the 10 custom-annotated images for car-hood, chair-seat, and torso were selected, since those tokens are the most likely to generalize poorly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the core idea is clean and the paper does the work. They learn part tokens with BCE loss against part masks on a handful of images, then use those tokens to produce cross-attention masks at inference, and blend features between source and edit trajectories with an adaptive non-binary threshold. That combination is new relative to SLiMe and Open-Vocabulary Attention Maps, which stop at segmentation, and relative to P2P/iP2P, which do not localize parts. The equations are coherent, the ablations cover the main knobs (timesteps, layers, threshold, token padding), and the appendix includes a 5-fold cross-validation for quadruped-head mIoU. No circularity either: tokens are trained against masks and then evaluated on held-out images; that is a legitimate transfer test. The benchmark is small but it is a real attempt at a reusable evaluation protocol. If the code and data actually ship on the project page, that is a reproducible contribution.\n\nThe soft spots are about claims, not mechanism. Table 1 has no error bars or significance tests; the user studies report only the 'ours preferred' percentage, without raw counts or confidence intervals. 'Outperforms all methods on all metrics' is too strong for a benchmark of 60 synthetic plus 13 real images, especially with hyperparameters tuned on the same benchmark. The stress-test concern about under-validated localization is fair and lands: the load-bearing assumption is that a token trained on 10-20 images, some custom-annotated (car-hood, chair-seat), gives reliable cross-attention maps on unseen images. The paper never reports per-part localization quality on the evaluation set; Appendix J gives mIoU only for quadruped head under SD2.1, not SDXL and not the other six parts. Since every edit depends on these masks, the averaged wins in Table 1 could be carried by parts that localize well while hoods and seats do worse. That does not sink the method, but it means the central quantitative claim is not yet established. Minor: 'first' is asserted without a tight prior-art search, and token optimization for part localization already exists. Code is promised without a commit hash, so I cannot verify what ran.\n\nWho is this for: anyone building text-guided editing or using cross-attention supervision for fine-grained control. The limitations section is honest about style transfer and unrealistic edits. It deserves a serious referee; I would send it out and ask for per-part localization numbers, error bars on Table 1, and a check that the benchmark prompt distribution does not overlap the token training distribution.","headline":"PartEdit is a genuinely useful combination of token optimization and mask-based blending for text-only part editing; the method holds together, but the 'all metrics' claim outruns the evidence because localization is only validated for one part.","tokens_in":22299,"tokens_out":4455,"would_cite":true,"duration_ms":43490,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"PartEdit is a text-based method for editing individual object parts with pre-trained diffusion models: it learns part tokens whose cross-attention maps localize the part, then blends source and edited features through those maps so only…","keywords":["fine-grained image editing","object parts","diffusion models","cross-attention","token optimization","feature blending","text-based editing","part editing benchmark"],"falsifier":"Train a part token with the paper's 10–20-image protocol, then run edits on a held-out set of object categories, poses, and scenes that were absent from training. Compare the aggregated inference-time masks against manual part annotations: if mean IoU falls toward chance for small or articulated parts, or if background PSNR and SSIM on the edited images drop below existing mask-free baselines, the localization-transfer assumption fails.","tokens_in":21298,"feed_emoji":"🎨","tokens_out":6478,"duration_ms":62913,"temperature":0.7,"pith_summary":"This paper claims that text-only editing of individual object parts—heads, torsos, hoods, seats—is achievable with a pre-trained diffusion model, without user-supplied masks and without retraining the generation model. The method trains a special part token so that its cross-attention maps reproduce the part's segmentation mask during denoising; those maps then serve as soft blending masks that confine the edit to the part. The authors construct a part-editing benchmark and report that PartEdit beats previous text-editing methods on all measured metrics and is chosen by users 66–90 percent of the time in paired comparisons.","feed_headline":"PartEdit edits object parts with text alone","feed_subtitle":"Learned part tokens keep edits like 'spiderman torso' inside the torso, no user masks needed.","key_machinery":"The load-bearing mechanism is the optimized part token and its cross-attention maps. A trainable embedding, initialized randomly, is supervised with a binary cross-entropy loss against ground-truth part masks at chosen timesteps and UNet layers; the trained token then produces localization maps at every inference step. Those maps are aggregated across layers, min-max normalized, adaptively thresholded so that background is zeroed and a soft transition band remains, and used to blend source-path and edit-path features at each layer and timestep. This single mechanism both localizes the edit and preserves the unedited image.","core_discovery":"Pre-trained diffusion models do not reliably know where object parts are: the cross-attention map for \"head\" can fire on arms, and \"hood\" can fail to appear at all. PartEdit counters this by expanding the model's part vocabulary with learned tokens while keeping the model frozen. Each part token is optimized with a binary cross-entropy loss against part masks from existing part datasets or a small user-annotated set, over intermediate timesteps and selected UNet decoder layers. At inference the token's aggregated cross-attention maps become a per-step non-binary mask; an adaptive thresholding step plus per-layer, per-timestep feature blending between the source and edited trajectories applies the new attribute exactly where the part lies. The result is that edits such as \"spiderman torso\" or \"destroyed hood\" change only the requested region and blend into the original object.","pith_inferences":["Beyond the paper: if the localization transfer holds, part tokens effectively convert segmentation annotations into reusable text-level controls, suggesting that cross-attention can serve as a general-purpose bottleneck for part grounding in editing pipelines.","Beyond the paper: the authors' observation that transformer-based diffusion models already localize parts more accurately suggests PartEdit's token-learning mechanism could yield even better results on such backbones.","Beyond the paper: the benchmark's separate foreground and background metrics give future editors a concrete evaluation recipe: verify the edit landed where requested and that everything else is unchanged.","Beyond the paper: the reported failure to mix a new style into a part points to self-attention as the next target, hinting that style and content disentanglement within the edited region is the main open problem."],"forward_implications":["Any object part for which roughly 10–20 annotated images exist can be edited by text alone; masks are needed only during token training, not at edit time.","Because the diffusion model stays frozen, the user keeps the full generative knowledge of the base model and can apply arbitrary new attributes or identities to the trained part without retraining.","Edits stay localized: unedited regions are preserved better than with prompt-to-prompt and instruction-based editing methods on the new PartEdit benchmark.","Multiple part tokens can be combined at inference to edit several regions in one image, by jointly normalizing their attention maps.","The approach works for real photographs when paired with an inversion method and an automatic captioner, not only for generated images."],"supporting_citations":[{"why":"Supplies the token-optimization recipe of optimizing a prompt embedding while the model stays frozen, which the part-token training builds on.","marker":"[Zhou et al. 2022]"},{"why":"SDXL is the frozen pre-trained diffusion backbone whose cross-attention maps are supervised and later used for localization.","marker":"[Podell et al. 2024]"},{"why":"Prompt-to-Prompt is the main synthetic-editing baseline and the origin of the cross-attention manipulation approach extended here.","marker":"[Hertz et al. 2022]"},{"why":"PASCAL-Part supplies part segmentation masks used to train the part tokens.","marker":"[Donadello and Serafini 2016]"},{"why":"PartImageNet supplies part masks such as quadruped-head annotations for token optimization and cross-validation.","marker":"[He et al. 2022]"},{"why":"LEDITS++ provides the inversion used for real-image edits and is also the strongest real-image editing baseline.","marker":"[Brack et al. 2024]"},{"why":"BLIP-2 generates source captions for real images, enabling text-based part edits on real photographs.","marker":"[Li et al. 2023]"},{"why":"Provides the masked CLIP similarity metric used to evaluate edited and unedited regions.","marker":"[Sun et al. 2024]"},{"why":"The adaptive thresholding scheme in the blending mask builds on this histogram-based threshold selection.","marker":"[Otsu et al. 1975]"}],"fun_headline_variants":["PartEdit learns part tokens for fine-grained image edits","PartEdit: text-only part editing without user masks","Diffusion models finally understand object parts via PartEdit","PartEdit: learned tokens localize edits to exact object parts","PartEdit expands diffusion's part vocabulary for precise edits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline depends on a token trained on only 10–20 images producing cross-attention maps that correctly find the requested part on unseen images, prompts, objects, and backgrounds; if that localization transfer fails, the feature blending edits the wrong region.","fun_headline_variants_meta":{"raw":{"variants":["PartEdit learns part tokens for fine-grained image edits","PartEdit: text-only part editing without user masks","Diffusion models finally understand object parts via PartEdit","PartEdit: learned tokens localize edits to exact object parts","PartEdit expands diffusion's part vocabulary for precise edits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000172,"raw_usage":{"total_tokens":1248,"prompt_tokens":894,"completion_tokens":354,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":276}},"tokens_in":510,"tokens_out":354,"duration_ms":3838,"temperature":1.0,"reasoning_tokens":276,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T23:42:36.444114+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a part token with the paper's 10–20-image protocol, then run edits on a held-out set of object categories, poses, and scenes that were absent from training. Compare the aggregated inference-time masks against manual part annotations: if mean IoU falls toward chance for small or articulated parts, or if background PSNR and SSIM on the edited images drop below existing mask-free baselines, the localization-transfer assumption fails.","supporting_citations":[],"review_version":1}