{"id":"38abee2b-78b9-47f8-8be3-498f4ce9cf14","arxiv_id":"2608.09322","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Token-wise surprisal selects editable regions, source key/value memory and hard projection preserve the rest, yielding record background-preservation scores on PIE-Bench without masks or fine-tuning.","lead":"ATDEdit is an inference-time method for editing images with diffusion transformers. It finds which image tokens should change, updates only those, and locks the rest to the source image, reporting better background preservation on PIE-Bench than published baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (7) constructs the target latent with an unscaled displacement, inconsistent with the rectified-flow interpolation in Eq. (6); the steering update is therefore evaluated off-manifold and lacks a correctness guarantee.","rationale":"The reader's weakest assumption is that token-wise conditional surprisal reliably identifies editable regions. I disagree that this is the single most load-bearing concern. Even if the surprisal perfectly identified the editable set, the steering update computed from Eq. (7) has no correctness guarantee because the target latent z_tar is not the rectified-flow interpolation for the current editable state x_edit. Under the paper's own definition in Eq. (6), the correct same-noise target latent is z_src + (1-t)(x_edit - x_src), not z_src + (x_edit - x_src). At intermediate noise levels (e.g., t = 0.5), the error is half the edit displacement, which is substantial. The paper carefully scopes other formal claims in Section 3.7, but does not address this construction, and the reference to FlowEdit does not resolve the inconsistency. Because the same target velocity is used to steer the editable-token updates, the Table 1 results could depend on out-of-distribution model behavior rather than the described mechanism. A concrete re-derivation or a re-run with the corrected interpolation would settle whether this concern is load-bearing. My read keeps the overall verdict CONDITIONAL, since the empirical numbers might still be reproducible despite the flawed construction, but the condition should include a check of this identity.","tokens_in":13189,"tokens_out":21664,"duration_ms":219221,"concrete_test":"Replace Eq. (7) with the corrected same-noise target latent z_tar_t = z_src_t + (1-t)(x_edit_t - x_src), keeping all other components fixed, and rerun the PIE-Bench evaluation. If the preservation metrics (PSNR, LPIPS) or CLIP-Edited change by more than a small margin (e.g., PSNR > 0.5 dB, LPIPS > 0.01, CLIP > 0.2), the reported results are artifacts of the off-manifold evaluation. Additionally, for a perfectly trained linear flow model, derive analytically whether the corrected rule yields the intended transport from x_src toward the target, and show the uncorrected rule does not.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.3 defines the source latent via the rectified-flow interpolation (Eq. (6)): z_src_t = (1-t)x_src + t epsilon. Equation (7) then sets the coupled target latent to z_tar_t = z_src_t + (x_edit_t - x_src). For the target state x_edit_t, the same-noise rectified-flow interpolation would be (1-t)x_edit_t + t epsilon = z_src_t + (1-t)(x_edit_t - x_src). The paper's construction omits the (1-t) factor, so z_tar_t lies off the interpolation manifold by t(x_edit_t - x_src). This off-manifold latent is fed into the target velocity v_theta in Eq. (13), producing the steering estimate that drives the update in Eq. (14). The paper supplies no theorem or empirical check that the model remains well-calibrated at these points; the 'Common-noise scope' in Section 3.7 only bounds variance, not bias. Since the central empirical claim (Table 1) depends on this update, the method's stated mechanism is not established. This is an internal inconsistency, not a matter of consensus.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes ATDEdit, a training-free, inference-time method for text-guided editing with diffusion transformers. The method computes a token-level \"conditional surprisal\" from same-state source/target velocity predictions to identify editable tokens, applies target-conditioned updates only to those tokens, substitutes source key/value memory at keep-token positions, and hard-resets keep-token latent rows to their source values. The authors derive scoped theoretical statements: common-noise coupling reduces estimator variance under nonnegative covariance, hard projection yields exact zero drift on selected keep-token rows in exact arithmetic, and top-quantile selection maximizes captured score mass under a fixed budget. On PIE-Bench, they report the strongest background-preservation numbers in their Table 1 comparison, including 27.44 dB PSNR and 0.055 LPIPS, while maintaining competitive semantic alignment.","tokens_in":13423,"tokens_out":10254,"duration_ms":103862,"significance":"If the reported results are robust, the contribution is practically useful: a mask-free, fine-tuning-free editing procedure that works with a modern rectified-flow DiT and improves background preservation. The manuscript is unusually careful about the scope of its formal claims: it repeatedly states that hard projection is a token-state guarantee, not pixel-level invariance; that common-noise variance reduction requires nonnegative covariance; and that top-quantile selection is only a fixed-budget score-mass statement. These explicit scope limitations are a strength. The main significance is nonetheless conditional on the empirical evaluation being reproducible and on the off-manifold target latents not degrading model calibration.","major_comments":[{"comment":"The coupled target latent is defined as z^{tar}_t = z^{src}_t + (x^{edit}_t - x^{src}_t). For the rectified-flow interpolation in Eq. (6), the same-noise target latent corresponding to the current editable state x^{edit}_t would be (1-t)x^{edit}_t + tε = z^{src}_t + (1-t)(x^{edit}_t - x^{src}_t). Eq. (7) therefore evaluates v_θ at points that are not the standard interpolation for x^{edit}_t; the deviation is t(x^{edit}_t - x^{src}_t). The velocity difference in Eq. (13), which drives the update in Eq. (14), is thus obtained from inputs where the model may be miscalibrated. The common-noise variance bound in Section 3.7 concerns variance, not bias, and does not cover this. Since the central empirical claim depends on this update, the paper should either provide evidence that the model remains calibrated at these inputs (e.g., compare against the consistent interpolation from Eq. (6), or measure clean-estimate error on typical edit states during sampling) or explicitly state the construction as an unvalidated heuristic and temper the corresponding claims.","section":"§3.3, Eq. (7)"},{"comment":"The method's free parameters—the editable-fraction schedule ρ(t), EMA coefficient β, detection window [t_lo, t_hi], mask-update interval κ, Monte Carlo draws n_avg, and dilation rounds r—are not reported anywhere in the manuscript. Figure 5 studies only [t_lo, t_hi], on a single \"Chicken → Duck\" example, and the caption explicitly states that the plotted values are not dataset-level PIE-Bench averages. Without the exact hyperparameter values used for Table 1, the results cannot be reproduced, and the reported preservation gains could depend on choices tuned on a single example. Please report the full hyperparameter configuration and provide a dataset-level sensitivity study, or at least a small validation split, for the most influential parameters.","section":"§4.1 and Table 1"},{"comment":"The core mechanism is the online editable-token mask derived from Token-wise Conditional Surprisal, but the manuscript never reports any direct measure of mask quality. PIE-Bench provides ground-truth edited-region masks; Section 4.1 says these are used only for evaluation metrics. Reporting, for example, the IoU between the final keep/edit mask and the ground-truth edited region, or a localization precision/recall curve across time, would show whether the surprisal ranking is actually finding the editable region rather than indirectly inferring this from PSNR/LPIPS. The current evidence leaves open the possibility that the preservation gains come from freezing a large fraction of tokens rather than from correct localization. Section 5 correctly admits the method's reliance on DiT calibration; a direct mask-quality metric would address this admitted limitation.","section":"§3.4 and Table 1"},{"comment":"The abstract and conclusion claim the \"strongest reported preservation metrics\" without qualification, but Table 1 mixes backbones: DiffEdit and InfEdit run on SD1.4, KVEdit and Stable Flow on FLUX, and only FlowEdit, FlowAlign, and ATDEdit on SD3. The authors acknowledge this in the body (\"we treat FlowEdit and FlowAlign as the primary same-backbone SD3 comparisons\"), but the headline claim is still stated without that caveat. Please qualify the abstract and conclusion, e.g., \"strongest among the compared methods on this benchmark\" or \"strongest among the same-backbone SD3 comparisons,\" and ideally add a same-backbone re-run of at least one representative baseline. The table also reports point estimates without error bars; reporting variance over Monte Carlo seeds or small hyperparameter perturbations would strengthen the comparison.","section":"Abstract and Table 1"}],"minor_comments":[{"comment":"The text references \"Eq. (15)\" as the Euclidean projection several paragraphs before Eq. (15) is introduced in Section 3.6; please use a forward-reference or restate the projection formula locally.","section":"§3.2"},{"comment":"Please state explicitly whether the source memory (K^src, V^src) is computed at z^src_t or at the target latent z^tar_t; this affects the interpretation of Eq. (12).","section":"§3.5"},{"comment":"The axes are labeled \"Tlo\" and \"Thi\" while the text uses t_lo and t_hi; please unify the notation.","section":"Figure 5"},{"comment":"The score divides by d (the feature dimension) but not by the number of Monte Carlo draws; please clarify whether the 1/d factor is an intentional feature-channel normalization.","section":"Eq. (9)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is honest and unusually well scoped, and the empirical comparison is fairly extensive. My recommendation hinges on whether the authors can (i) report exact hyperparameters and dataset-level sensitivity, and (ii) address the off-manifold target-latent concern with either a correction or a direct empirical validation. If these are resolved, the paper would be a solid fit for the venue. One additional consideration: the abstract's \"strongest reported preservation metrics\" should be qualified to avoid cross-backbone overclaiming."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"ATDEdit is a genuinely new combination of known pieces: online token-wise conditional surprisal for locating edits, source key/value memory substitution, and hard projection of keep-token rows. The writing is clear and the formal claims are honestly scoped. The projection statement is exactly the Euclidean projection onto the keep-set constraint, and the top-quantile claim is just fixed-budget score maximization; the paper says both explicitly and does not oversell them. That is a real strength.\n\nThe empirical story is plausible but not fully secured. Table 1 mixes backbones; the same-backbone comparison against FlowEdit and FlowAlign is the only fair one, and ATDEdit wins preservation metrics there. But there are no error bars, no code or exact hyperparameters, and the ablation table is point estimates with a note that differences should be read descriptively. For a claimed SOTA result, that is thin. Figure 5 is a one-sample sensitivity sweep, so the window parameters are not really validated.\n\nThe stress-test on Eq. (7) is correct. The paper couples the target latent as z_tar_t = z_src_t + (x_edit_t - x_src), but same-noise rectified-flow interpolation would give z_src_t + (1-t)(x_edit_t - x_src). As written, z_tar_t is off the interpolation manifold by t(x_edit_t - x_src), and the target velocity is evaluated at those off-manifold points. The paper gives no argument or empirical check that the SD3 velocity field stays calibrated there. This does not invalidate the variance-reduction claim, since the same noise realization is still shared, but it does mean the steering update has no correctness guarantee and the method's stated mechanism is not fully established. This is a real internal inconsistency, not a matter of taste.\n\nThe self-citations to the authors' earlier editing methods are not a problem by themselves; what is more noticeable is that those methods are cited but not compared, which is easy for a reader to flag. The conclusion also honestly lists limitations: computational overhead, reliance on DiT calibration, and poor fit for global edits.\n\nWho is this for? Anyone working on training-free DiT image editing. It deserves peer review: the core idea is coherent, the write-up is honest, and the off-manifold coupling is either fixable or at least addressable with an experiment showing it does not matter. I would not desk-reject it. With code, error bars, and a response on the Eq. (7) issue, it could be a solid contribution.","headline":"ATDEdit is a coherent and honestly scoped test-time editing recipe, but the target-latent coupling in Eq. (7) is off-manifold as written and the headline numbers need error bars and code.","tokens_in":13998,"tokens_out":2650,"would_cite":true,"duration_ms":26949,"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":"ATDEdit makes diffusion editing token-asynchronous, using a surprisal score to update only tokens that should change and reporting 27.44 dB PSNR on unedited regions.","keywords":["text-guided image editing","diffusion models","diffusion transformers","asynchronous token decoding","inference-time editing","background preservation","token-wise conditional surprisal","hard projection"],"falsifier":"Run ATDEdit on a global edit (e.g., 'turn day to night' where nearly every token should change) and compare CLIP-Edited against a synchronous SD3 baseline under the same step budget; if the local-editing framework freezes tokens that should change and CLIP-Edited drops substantially below the synchronous baseline, that outcome would contradict the claim that surprisal reliably identifies the editable set.","tokens_in":12952,"feed_emoji":"🎨","tokens_out":7763,"duration_ms":71227,"temperature":0.7,"pith_summary":"The paper proposes ATDEdit, an inference-time framework for text-guided image editing with Diffusion Transformers that treats each sampler step as a parallel but globally coupled update of a token matrix. It argues that switching the text condition for every token at once causes global drift, and that editing should be asynchronous: each token row follows its own condition-switching and update rule. The method discovers editable tokens online through Token-wise Conditional Surprisal, a same-state velocity-difference score, then constrains the remaining keep tokens by supplying source key/value memory and hard-projecting their latent rows back to the source. On PIE-Bench, ATDEdit reports the strongest background-preservation numbers in its comparison (27.44 dB PSNR, 0.055 LPIPS), without external masks, user masks, or fine-tuning; if correct, this gives a training-free way to keep local edits local in modern DiT-based editors.","feed_headline":"Token surprisal keeps diffusion edits local: 27.44 dB PSNR","feed_subtitle":"Edits only high-surprisal image tokens and freezes the rest, beating prior editors on background fidelity with no masks or fine-tuning.","key_machinery":"The load-bearing object is Token-wise Conditional Surprisal, a token-level score $s_i(t)$ that averages, over $n_{\\mathrm{avg}}$ common-noise draws, the $\\ell^1$ norm of the difference between predicted clean latents under target and source conditions at the same noisy latent; under rectified-flow parameterization this equals the magnitude of the same-state velocity difference. The score is smoothed by EMA, top-quantile selected in a mid-noise window, and optionally dilated to form the editable mask $E(t)$. The complementary machinery is the keep-token constraint: a partial attention memory boundary that substitutes source key/value rows at keep positions (leaving queries and attention weights target-conditioned) and a Euclidean projection that resets keep-token latent rows to their source values. Together they implement the asynchronous update, and the paper proves the projection is the exact Euclidean projection onto the affine keep-token constraint set.","core_discovery":"ATDEdit reformulates diffusion editing as asynchronous token decoding: instead of a synchronous target-conditioned update applied to all tokens, the framework estimates an editable token set $E(t)$ at each timestep using Token-wise Conditional Surprisal, defined as the Monte Carlo average of the $\\ell^1$ magnitude of the same-state $\\hat{x}_0$ prediction difference when only the condition switches from $c_{\\mathrm{src}}$ to $c_{\\mathrm{tar}}$. Selected editable tokens receive target-conditioned corrections under a common-noise parallelogram coupling, while keep tokens are constrained by an attention memory boundary that replaces their keys/values with source-branch values and by a hard projection that exactly resets their latent rows to the source state. The paper is explicit about the scope of its formal results: the projection gives a token-state zero-drift guarantee in exact arithmetic, common noise reduces estimator variance under nonnegative covariance, and top-quantile mask selection optimizes only captured surprisal mass under a fixed cardinality budget. The empirical claim is that on PIE-Bench, this yields the strongest reported preservation metrics in the comparison — PSNR 27.44 dB, MSE 2.70 $\\times 10^{-3}$, SSIM 0.9303, LPIPS 0.055 — while CLIP-Edited remains competitive with the same-backbone SD3 baselines.","pith_inferences":["A testable extension is to let the user steer the editable fraction $\\rho(t)$ per region or per semantic class, which would make the surprisal-based mask a controllable editing dial rather than a fully automatic score.","The same asynchronous-token formulation could be carried to video editing, where the token grid is replaced by a spatiotemporal token set; the variance-reduction and zero-drift arguments would apply equally, but the mask discovery would need to handle motion.","Because the paper's formal guarantees are scoped to latent rows, a natural diagnostic is to compare pixel-level fidelity against the zero-drift guarantee: a large mismatch would localize how much of the background-preservation gain comes from the projection versus the KV memory boundary.","The robustness of the method hinges on whether surprisal ranking stays accurate under distribution shifts; stress-testing with prompts that are semantically similar to the source (e.g., near-synonym swaps) would show whether the editability signal degrades before the pixel metrics do."],"forward_implications":["Background preservation on PIE-Bench improves across Structure Distance, PSNR, MSE, SSIM, and LPIPS without any ground-truth or user-provided mask at inference time.","The online, time-varying mask lets the protected region adapt as denoising evolves, addressing a limitation of static mask-based editors like DiffEdit.","Because ATDEdit changes only the inference dynamics and not the backbone, it can be applied to any DiT-based rectified-flow sampler without fine-tuning.","The hard projection provides an exact latent-row zero-drift guarantee for selected keep tokens, making the preservation mechanism directly controllable in token space even though pixel-level invariance is explicitly not claimed.","The common-noise parallelogram coupling is a reusable variance-reduction device for any estimator built from paired source/target DiT evaluations."],"supporting_citations":[{"why":"Supplies the parallelogram coupling and the FlowEdit baseline that ATDEdit extends and compares against.","marker":"[14]"},{"why":"Provides the PIE-Bench dataset and evaluation protocol, the basis for the paper's empirical claim.","marker":"[10]"},{"why":"The SD3 backbone on which ATDEdit is implemented and against which same-backbone comparisons are made.","marker":"[7]"},{"why":"Key/value memory reuse idea that the attention memory boundary is inspired by and compared with.","marker":"[40]"},{"why":"DiffEdit is the mask-based editing baseline that motivates the online mask-discovery contribution.","marker":"[5]"},{"why":"MasaCtrl attention control provides a baseline that the proposed memory boundary is contrasted with.","marker":"[3]"},{"why":"DiT architecture provides the token-indexed representation that makes asynchronous token decoding possible.","marker":"[24]"},{"why":"Common random numbers result underlying the variance-reduction inequality for the common-noise coupling.","marker":"[38]"}],"fun_headline_variants":["Async token decoding edits only high-surprisal tokens","Surprisal-guided token editing: 27.44 dB PSNR on PIE-Bench","No masks, no fine-tuning: token surprisal keeps edits local","ATDEdit: async token decoding for local edits","Preserving background while editing: token surprisal selection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that token-wise conditional surprisal, computed from same-state source/target velocity differences, reliably identifies the tokens that should change under the new prompt; if the DiT velocity field is miscalibrated or the surprisal ranking is wrong, the hard projection will freeze semantically editable regions or allow background drift, and the reported PSNR and LPIPS gains will not transfer to new tasks.","fun_headline_variants_meta":{"raw":{"variants":["Async token decoding edits only high-surprisal tokens","Surprisal-guided token editing: 27.44 dB PSNR on PIE-Bench","No masks, no fine-tuning: token surprisal keeps edits local","ATDEdit: async token decoding for local edits","Preserving background while editing: token surprisal selection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001683,"raw_usage":{"total_tokens":6729,"prompt_tokens":1062,"completion_tokens":5667,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":678,"completion_tokens_details":{"reasoning_tokens":5588}},"tokens_in":678,"tokens_out":5667,"duration_ms":36972,"temperature":1.0,"reasoning_tokens":5588,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:18:24.082482+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ATDEdit on a global edit (e.g., 'turn day to night' where nearly every token should change) and compare CLIP-Edited against a synchronous SD3 baseline under the same step budget; if the local-editing framework freezes tokens that should change and CLIP-Edited drops substantially below the synchronous baseline, that outcome would contradict the claim that surprisal reliably identifies the editable set.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Common random numbers result underlying the variance-reduction inequality for the common-noise coupling."}],"review_version":1}