{"id":"bafe0fa0-977e-448e-ab61-b143d3b59016","arxiv_id":"2411.19652","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Setting cross-attention maps to a fixed uniform distribution over text tokens reduces reconstruction error in DDIM inversion and improves text-guided editing when combined with adaptive mask blending.","lead":"This paper replaces the cross-attention maps inside a diffusion model's U-Net with uniform maps, one fixed value per text token, to make image reconstruction from DDIM inversion more stable. It also adds a per-timestep mask that blends a neutral reconstruction with the edited image, improving text-guided editing.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Paper asserts uniform attention maps 'maintain compatibility with the pretraining distribution' (Sec. 3.3.1) with no evidence; if false, the method's success may not generalize to other models/resolutions/schedules, undermining the general claim.","rationale":"The reader identified the same weakest assumption: the pretrained U-Net's behavior under replaced cross-attention maps is asserted but not verified. I agree this is the most load-bearing concern because it underwrites the method's core claim and its stated advantage over zero cross-attention. The paper's empirical results are internally consistent and the method is simple and plausible, but the absence of any distributional evidence leaves a real gap: if uniform attention is not actually compatible with the pretraining distribution, the reconstruction gains could be an artifact of the specific model, resolution, or sampling schedule. The proposed test (varying model family and step count, plus direct feature-distribution comparison) would settle whether the concern lands. The reader's CONDITIONAL verdict is appropriate; my stress-test does not change it, hence UNCHANGED. I do not see a more fundamental objection: the correlation analysis in Fig. 3 is suggestive rather than causal, and the editing hyperparameters are tuned, but these are secondary to the reconstruction claim. The uniform-attention trick is well-defined, has a parameter-free form (Eq. 7), and the reported numbers are substantial; the remaining risk is generalization, which the concrete test directly addresses.","tokens_in":15633,"tokens_out":5312,"duration_ms":44722,"concrete_test":"Run the Tab. 1 reconstruction protocol on PIE with SD 2.1 and 50-step DDIM (plus SDXL if feasible), comparing Uniform Attention Maps against Null Prompt and Zero Cross-Attention; if the improvement over baselines is not preserved, the method lacks generality. Additionally, for a fixed set of inputs, compute per-layer mean/variance and cosine distances of x^(l) under original, uniform, and zero cross-attention; if uniform features are not closer to original features than zero-attention features, the Sec. 3.3.1 pretraining-compatibility argument is empirically false.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism is Eq. 7, S_uniform = (1/N)·1_{M×N}, and the paper's justification for preferring it over zero cross-attention is the claim (Sec. 3.3.1) that uniform maps 'maintain compatibility with the pretraining distribution' of latent features x^(l). This claim is unsupported: the pretrained U-Net was trained with strongly non-uniform, query-dependent cross-attention maps, yet the paper provides no layer-wise statistics or causal test comparing the feature distributions under original, uniform, and zero attention. The success in Tabs. 1–5 is demonstrated only on SD 1.4 with 20-step DDIM (reconstruction) and SD 2.1 with DPM (composition), so the observed gains could be specific to these settings. If uniform attention is, in fact, as out-of-distribution as zero attention for other models or schedules, the method would silently fail outside the reported configuration, and the paper's general claim—'significantly enhancing image reconstruction fidelity'—would not hold. This is load-bearing because the method's theoretical motivation and its apparent robustness across prompts rest entirely on this compatibility assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Uniform Attention Maps (UAM) for diffusion-model inversion and editing: in Eq. (7), the cross-attention score map S_t^{(l)} is replaced by the uniform matrix S_uniform = (1/N) 1_{M^{(l)} x N}, and this replacement is used during DDIM inversion and reconstruction with value tokens taken from a chosen prompt. The paper also introduces an adaptive mask-guided editing algorithm (Sec. 3.3.2) that blends an auxiliary null-prompt/uniform-attention branch with the target branch using a timestep-dependent mask. Experiments on PIE, TF-ICON, and CelebA-HQ report improved reconstruction fidelity (Tabs. 1-2), composition (Tab. 5), and editing metrics (Tabs. 3-4) relative to several baselines, with ablations for the mask threshold lambda and mask timestep T_mask (Tab. 6). The authors attribute the gains to cross-attention misalignment between inversion and reconstruction (Sec. 3.2, Fig. 3) and assert that uniform maps 'maintain compatibility with the pretraining distribution' (Sec. 3.3.1).","tokens_in":15854,"tokens_out":7004,"duration_ms":69840,"significance":"If the empirical results hold, the method is a simple drop-in replacement that improves reconstruction fidelity across two Stable Diffusion versions and two solvers, and it modestly improves editing when combined with MasaCtrl or Prompt-to-Prompt. The paper provides code, and the reconstruction gains in Tab. 1 are large and consistent, which strengthens the plausibility of the central claim. However, the causal mechanism and the compatibility assertion are not yet supported, and some editing improvements are small and unreplicated. The contribution is therefore conditional on additional validation of the mechanism, generalization, and statistical significance; it is not yet ready for acceptance.","major_comments":[{"comment":"The causal claim that cross-attention discrepancies cause reconstruction errors is supported only by a correlation scatter plot between MSE(A_inv - A_rec) and MSE(z_0,inv - z_0,rec). This correlation does not establish that the cross-attention term is the driver, as both quantities could reflect a common confound such as general trajectory mismatch. To make the attribution load-bearing, the authors should supply an intervention: for example, hold the attention maps fixed between inversion and reconstruction while varying the value tokens, or run inversion with original attention maps and reconstruction with uniform maps (and vice versa), and show that reconstruction error tracks the A-term mismatch. As written, Eq. (7) is motivated by a claim that the experiments do not isolate.","section":"Sec. 3.2, Fig. 3"},{"comment":"The assertion that uniform attention maps 'maintain compatibility with the pretraining distribution' is unsupported. The paper provides no layer-wise statistics of the intermediate features x^{(l)} under original, uniform, and zero attention, and no evaluation on other model families, resolutions, or step counts beyond SD1.4 with 20-step DDIM and SD2.1 with DPM. The observed superiority of uniform over zero attention (Tab. 1) is consistent with the claim, but it does not demonstrate compatibility; many other fixed attention patterns could also outperform zero. Please either add direct distributional or causal evidence, or soften the claim to an empirical observation.","section":"Sec. 3.3.1"},{"comment":"The editing hyperparameters lambda (quantile) and T_mask are selected by ablations on the PIE benchmark (Tab. 6) and then used to report the headline editing results on the same PIE split (Tabs. 3-4). This is a form of test-set selection; as a result, the reported editing gains may not generalize to new images. The authors should either use a held-out validation set, report the sensitivity of Tables 3-4 over a range of lambda and T_mask, or justify that the chosen values are not tuned to the test set.","section":"Sec. 4.6, Tabs. 3-4, 6"},{"comment":"The editing improvements for the strongest baseline (DI) are small: for example, Structure Distance 24.70 to 24.60, PSNR 22.64 to 22.68, and SSIM 81.33 to 81.52 in Tab. 3. No standard deviations, confidence intervals, or paired significance tests are reported for any table, so it is unclear whether these differences are statistically reliable. In addition, Tab. 4 shows CLIP(Edited) decreasing from 22.44 (DDIM) to 22.23 (DDIM+Ours), which contradicts the text's statement that CLIP scores in edited regions show 'notable gains'. Please report variance across images or runs and correct the overclaim.","section":"Tabs. 3-4"}],"minor_comments":[{"comment":"The axis label 'MSE(A_inv - A_rec) / MSE(z0,inv - z0,rec) x 1e-3' is ambiguous; please spell out the quantities and units used on both axes.","section":"Fig. 3"},{"comment":"The phrase 'using the noise prediction from the previous timestep' should clarify the order; DDIM inversion updates z_t using the noise prediction at t-1, which is an approximation rather than an assumption of equality between adjacent timesteps.","section":"Sec. 3.2, Eq. (3)"},{"comment":"The symbol C is introduced for uniform cross-attention maps but the main text uses S_uniform; please unify the notation.","section":"Algorithm 1"},{"comment":"References [31] and [34] contain stray page numbers '8, 2' that appear to be artifacts; please correct the bibliography entries.","section":"References"},{"comment":"There are several typos, including 'the clean predicted imageˆz0,t' in Sec. 3.2, 'T arget' in Algorithm 1, and 'bench mark' in Sec. 4.1; a careful proofread is needed.","section":"Various"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's empirical core is simple and the reconstruction gains are substantial, so I do not recommend rejection. However, the reported mechanism is weakly supported, and the editing claims are partly based on small, unreplicated differences and test-set-selected hyperparameters. I would urge the editor to request the additional experiments outlined in the major comments. The paper may be acceptable after this revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the uniform-attention swap is a genuine, training-free improvement for DDIM inversion reconstruction, and the adaptive mask is a reasonable add-on. The analysis section oversells a correlation as causation, and the editing numbers rest on hyperparameters chosen on the same benchmark. Still, the central result is simple enough and consistently benchmarked that it deserves serious refereeing.\n\nWhat's new: replacing the U-Net's cross-attention maps with a fixed 1/N matrix during inversion (Eq. 7) is not in the cited literature, and the per-timestep adaptive mask blending (auxiliary branch + dilation + threshold) is also new. Reconstruction gains are consistent across PIE and CelebA-HQ, and the composition/editing tables show improvements over strong baselines like TF-ICON, P2P, and MasaCtrl. The ablation on lambda and Tmask (Tab. 6) is honest about sensitivity, and the code link is a plus.\n\nSoft spots: Fig. 3 is a scatter correlation between attention-term MSE and reconstruction MSE. The paper concludes cross-attention is 'responsible' for errors, but no intervention is done on other components, so the causal claim is not supported. The 'compatibility with pretraining distribution' argument in Sec. 3.3.1 is asserted, not measured; the stress-test note is right that there are no layer-wise statistics. What saves the empirical argument is the direct comparison with zero cross-attention, which is worse — that is evidence that uniform is not just removing guidance. But generalization across SD versions, solvers, and resolutions is shown only for a narrow slice (SD1.4/DDIM 20-step, SD2.1/DPM). The editing gains are real but modest, and the hyperparameters (lambda quantile, Tmask) are chosen via ablation on the PIE benchmark, the same set used for headline numbers, which is mild test-set leakage.\n\nBottom line: this is a practical, working idea with a weak supporting narrative, not a broken one. The reconstruction contribution is self-contained and reproducible. The editing part needs a cleaner evaluation protocol. A good referee should push for layer-wise diagnostics, a second model/scheduler, and a separation between validation and test. I'd take it for review.\n\nWho it's for: people doing tuning-free editing or DDIM inversion; they'll pick up the trick quickly.","headline":"The uniform-attention swap is a real, training-free improvement for DDIM inversion reconstruction, but the paper's causal analysis oversells a correlation and the editing results rest on benchmark-tuned hyperparameters.","tokens_in":16397,"tokens_out":2214,"would_cite":false,"duration_ms":20045,"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":"Replacing cross-attention with a fixed uniform matrix over text tokens improves DDIM reconstruction fidelity, and an adaptive mask-guided blend carries the gain into editing.","keywords":["diffusion models","DDIM inversion","image reconstruction","image editing","cross-attention","uniform attention maps","tuning-free editing","mask-guided editing"],"falsifier":"A direct disconfirmation would be to run the same DDIM inversion–reconstruction protocol on a different backbone, such as a transformer-based diffusion model or a higher-resolution latent U-Net, with uniform cross-attention maps: if structure distance does not drop relative to the standard attention baseline, the pretraining-compatibility claim fails. A sharper test of the proposed mechanism measures the actual adjacent-timestep mismatch $\\|\\epsilon_\\theta(z_t, t, c) - \\epsilon_\\theta(z_{t-1}, t-1, c)\\|$ under standard versus uniform attention, since the paper's causal story predicts this mismatch shrinks with uniform maps and a measurement showing it unchanged or larger would separate the mechanism from the metric gains.","tokens_in":15436,"feed_emoji":"🖼️","tokens_out":12310,"duration_ms":83795,"temperature":0.7,"pith_summary":"Text-guided editing of real photos with diffusion models first inverts the image into a noise trajectory and then denoises it under the new prompt; in practice the trajectory does not return to the original image, and this paper argues that the U-Net's cross-attention maps are the main cause. It documents a positive correlation between inversion–reconstruction mismatch in the cross-attention term and error in the predicted clean image, then proposes a blunt fix: replace every cross-attention score map with the constant matrix $S = \\frac{1}{N}\\mathbf{1}_{M\\times N}$, so the attention update collapses to the average of the text-token value vectors. On the PIE benchmark with 20 DDIM steps this lowers structure distance from $15.31\\times10^{-3}$ to $4.76\\times10^{-3}$ under the null prompt, and reconstruction becomes nearly independent of whether the conditioning text is the source prompt or the empty string. For editing, the paper adds an auxiliary branch and an adaptive mask that blends the faithful uniform-attention reconstruction into the target-prompt branch, improving structure distance, background preservation, and CLIP alignment across several baselines. If the claims hold, a training-free one-line change to attention could remove the reconstruction drift that currently forces per-image optimization in tuning-free editing.","feed_headline":"Uniform attention maps cut DDIM reconstruction error threefold","feed_subtitle":"Swap cross-attention for the average token value and the inverted image returns faithfully, with no retraining.","key_machinery":"The load-bearing object is the uniform cross-attention map of Eq. (7), $S^{(l)}_{\\mathrm{uniform}} = \\frac{1}{N}\\mathbf{1}_{M^{(l)}\\times N}$, a constant matrix of ones scaled by the number of conditioning tokens that replaces the learned softmax score map in every U-Net cross-attention layer. Under this substitution the attention term becomes $A^{(l)}_t = S^{(l)}_{\\mathrm{uniform}} V^{(l)}$, meaning each pixel receives the arithmetic mean of the prompt-derived value vectors, independent of the query and of the timestep. This removes the prompt-dependent attention update from both the inversion pass and the reconstruction pass, so the adjacent-timestep noise-prediction approximation on which DDIM inversion relies is no longer perturbed by cross-attention drift. The companion mechanism is the adaptive mask, $M = \\mathrm{dilate}(|\\hat{z}^{\\mathrm{tgt}}_{0,t} - \\hat{z}^{\\mathrm{src}}_{0,t}| \\le \\lambda)$, which decides per timestep where the faithful uniform-attention prediction replaces the target-branch prediction.","core_discovery":"The central claim is that non-uniform cross-attention, not the ODE discretization itself, is what breaks DDIM inversion–reconstruction in text-conditioned diffusion models, and that substituting a uniform average over conditioning tokens restores fidelity. With the standard update $\\tilde{x}^{(l)}_t = x^{(l)}_t + A^{(l)}_t$ where $A^{(l)}_t = \\mathrm{softmax}(Q^{(l)}_t (K^{(l)})^{\\top}/\\sqrt{d})\\, V^{(l)}$, the paper replaces the softmax score map with $S^{(l)}_{\\mathrm{uniform}} = \\frac{1}{N}\\mathbf{1}_{M^{(l)}\\times N}$, so every visual token receives the same mean of the $N$ value vectors. The authors argue this respects the pretraining distribution better than zeroing the attention term entirely: Zero Cross-Attention Maps remove all semantic guidance and degrade fine details, while the uniform average keeps the expected attention update but removes prompt-dependent variance. For editing, they propose adaptive mask-guided editing with three parallel branches: an auxiliary branch (null prompt plus uniform attention) that reconstructs faithfully, a source branch, and a target branch; at each timestep a mask $M = \\mathrm{dilate}(|\\hat{z}^{\\mathrm{tgt}}_{0,t} - \\hat{z}^{\\mathrm{src}}_{0,t}| \\le \\lambda)$ blends the auxiliary prediction into the target branch. The paper reports the best reconstruction numbers in its tables (structure distance $4.76\\times10^{-3}$ versus $15.31\\times10^{-3}$ for the null-prompt baseline) and consistent gains when wrapped around Prompt-to-Prompt, MasaCtrl, and TF-ICON editing pipelines.","pith_inferences":["Editorial inference: if the mechanism is correct, the uniform-map fix should transfer to other inversion schemes whose forward–reverse mismatch is driven by cross-attention drift, such as EDICT-style coupled inversions; a cheap test is to swap their attention maps for uniform ones and check whether reconstruction error falls the same way.","Editorial inference: the Fig. 3 evidence is correlational, so the causal claim that cross-attention misalignment drives reconstruction error can be settled by a layer-wise ablation that applies uniform maps in only a subset of U-Net layers and locates where the error reduction concentrates.","Editorial inference: because uniform attention averages the value tokens, the conditioning signal becomes effectively constant, which predicts that scrambled or meaningless prompts should reconstruct as faithfully as the true source prompt; this is directly testable and would delimit how much of reconstruction fidelity actually depends on text conditioning."],"forward_implications":["DDIM inversion–reconstruction on the PIE benchmark improves without training or per-image optimization, cutting structure distance from $15.31\\times10^{-3}$ (null prompt) to $4.76\\times10^{-3}$ at 20 steps.","Reconstruction becomes nearly prompt-independent: uniform attention with source-prompt values and with null-prompt values give almost identical fidelity (structure distance $4.67$ and $4.76$), so pipelines no longer need prompt tuning for reconstruction quality.","Editing methods that rely on attention control (Prompt-to-Prompt, MasaCtrl) gain structure-distance, background-preservation, and CLIP-score improvements when the adaptive mask blend is added.","The TF-ICON image composition pipeline improves under the same substitution, with lower LPIPS on foreground and background and a higher CLIP text score.","Uniform attention maps outperform both the source/null prompt baselines and Zero Cross-Attention Maps, supporting the paper's claim that keeping the value-token mean preserves pretraining compatibility that wholesale zeroing destroys."],"supporting_citations":[{"why":"Supplies the DDIM inversion procedure whose reconstruction drift the paper analyzes and repairs.","marker":"[29]"},{"why":"The per-image optimization baseline (Null-text Inversion) whose cost motivates the tuning-free uniform-attention fix.","marker":"[25]"},{"why":"Provides the PIE benchmark, the editing setup, and the evaluation metrics used for all main experiments.","marker":"[14]"},{"why":"Supplies the TF-ICON composition benchmark and the 'exceptional prompt' stabilizing idea the paper contrasts with constant attention.","marker":"[21]"},{"why":"Defines the latent diffusion U-Net with cross-attention (Stable Diffusion) that the method modifies.","marker":"[27]"},{"why":"The Prompt-to-Prompt cross-attention control method that the adaptive mask blend is evaluated on.","marker":"[10]"},{"why":"The MasaCtrl mutual self-attention control baseline that the mask blend is wrapped around.","marker":"[3]"},{"why":"Establishes the adjacent-timestep noise-prediction approximation in DDIM inversion that the analysis builds on.","marker":"[33]"}],"fun_headline_variants":["Uniform attention maps fix DDIM inversion-reconstruction drift","Swap cross-attention for uniform maps to boost image fidelity","DDIM inversion errors blamed on cross-attention, solved by uniform maps","No retraining: uniform attention maps boost diffusion fidelity","Uniform attention maps improve both reconstruction and editing in diffusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the pretrained U-Net continues to denoise accurately when its cross-attention score maps are replaced by a fixed uniform matrix; the paper asserts 'compatibility with the pretraining distribution' on the strength of its benchmark results but supplies no layer-wise statistics or causal test, so the method could silently fail on other model families, resolutions, or sampling schedules.","fun_headline_variants_meta":{"raw":{"variants":["Uniform attention maps fix DDIM inversion-reconstruction drift","Swap cross-attention for uniform maps to boost image fidelity","DDIM inversion errors blamed on cross-attention, solved by uniform maps","No retraining: uniform attention maps boost diffusion fidelity","Uniform attention maps improve both reconstruction and editing in diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000324,"raw_usage":{"total_tokens":1901,"prompt_tokens":1111,"completion_tokens":790,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":727,"completion_tokens_details":{"reasoning_tokens":709}},"tokens_in":727,"tokens_out":790,"duration_ms":7192,"temperature":1.0,"reasoning_tokens":709,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:59:07.835232+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct disconfirmation would be to run the same DDIM inversion–reconstruction protocol on a different backbone, such as a transformer-based diffusion model or a higher-resolution latent U-Net, with uniform cross-attention maps: if structure distance does not drop relative to the standard attention baseline, the pretraining-compatibility claim fails. A sharper test of the proposed mechanism measures the actual adjacent-timestep mismatch $\\|\\epsilon_\\theta(z_t, t, c) - \\epsilon_\\theta(z_{t-1}, t-1, c)\\|$ under standard versus uniform attention, since the paper's causal story predicts this mismatch shrinks with uniform maps and a measurement showing it unchanged or larger would separate the mechanism from the metric gains.","supporting_citations":[{"cited_title":"Denoising diffusion implicit models","cited_arxiv_id":null,"evidence_quote":"Supplies the DDIM inversion procedure whose reconstruction drift the paper analyzes and repairs."},{"cited_title":"Null-text inversion for editing real images using guided diffusion models","cited_arxiv_id":null,"evidence_quote":"The per-image optimization baseline (Null-text Inversion) whose cost motivates the tuning-free uniform-attention fix."},{"cited_title":"Pnp inversion: Boosting diffusion-based editing with 3 lines of code","cited_arxiv_id":null,"evidence_quote":"Provides the PIE benchmark, the editing setup, and the evaluation metrics used for all main experiments."},{"cited_title":"TF-ICON: diffusion-based training-free cross-domain image composi- tion","cited_arxiv_id":null,"evidence_quote":"Supplies the TF-ICON composition benchmark and the 'exceptional prompt' stabilizing idea the paper contrasts with constant attention."},{"cited_title":"High-resolution image synthesis with latent diffusion models","cited_arxiv_id":null,"evidence_quote":"Defines the latent diffusion U-Net with cross-attention (Stable Diffusion) that the method modifies."},{"cited_title":"Prompt-to-prompt image editing with cross-attention control","cited_arxiv_id":null,"evidence_quote":"The Prompt-to-Prompt cross-attention control method that the adaptive mask blend is evaluated on."},{"cited_title":"Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and edit- ing","cited_arxiv_id":null,"evidence_quote":"The MasaCtrl mutual self-attention control baseline that the mask blend is wrapped around."},{"cited_title":"EDICT: exact diffusion inversion via coupled transformations","cited_arxiv_id":null,"evidence_quote":"Establishes the adjacent-timestep noise-prediction approximation in DDIM inversion that the analysis builds on."}],"review_version":1}