{"id":"e838ff63-3bcb-4675-b013-d480d073a148","arxiv_id":"2505.04320","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"The authors propose a FLUX-based multi-turn editing pipeline combining dual-objective LQR guidance and adaptive attention masking that maintains image identity over successive text edits, with the best FID score in fourth-round tests.","lead":"This paper presents a method for editing images in multiple rounds, keeping the original picture recognizable while applying a series of text-based changes such as adding a hat, changing a dog to a cat, and then adding a scarf. Because reliable iterative editing would make interactive photo retouching and product design tools practical, the method is worth attention, though its tests cover only four editing rounds.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Attention-mask localization is not validated: if the map-sum threshold in Eqs. 11-15 fails to track the edited region, the claimed editability/consistency balance, on which the central claim rests, is unsupported.","rationale":"The paper's central claim is that combining flow matching inversion, dual-objective LQR, and adaptive attention masking yields multi-turn consistency and editability. The weakest link is the masking component: it is the only mechanism responsible for localizing edits, and its selection criterion is justified by two visual examples and a prior observation rather than by quantitative localization accuracy. This is load-bearing because if the mask does not track the requested region, the method has no mechanism to prevent the LQR from over-constraining the edit or to prevent unwanted structural change; the reported FID and CLIP gains would not transfer reliably to new images or edit types. The concern does not, by itself, falsify the central claim: Table 2's ablation shows that medium activation outperforms high activation and no attention on FID, which is weak but real evidence that activation level matters. However, that ablation does not test localization, and because all hyperparameters are hand-picked without a held-out set and no error bars or significance tests are reported, the quantitative advantage could be configuration-specific. I also noted an internal inconsistency between Table 1 and Table 3: Table 1 lists FireFlow at FID 7.970 for the fourth turn, while Table 3 lists FireFlow at FID 12.375 for the same round, the value Table 1 assigns to FireFlow-v. That inconsistency is concerning for reporting quality, but it does not change the FID ordering, so I do not treat it as the primary issue. A mask-overlap test would settle whether the localization premise is sound. Given the reader already returned CONDITIONAL and my analysis reinforces that condition without overturning the empirical support, the appropriate verdict remains conditional; no adjustment is needed.","tokens_in":17299,"tokens_out":6320,"duration_ms":63510,"concrete_test":"Use the PIE-Bench edit-region masks (which the benchmark provides) to compute, for each of the 10 editing types and all four editing turns, the IoU between the mask M_k generated by Eq. (14) under the paper's settings (i=10..14, hfactor=2.0, rfactor=0.8, and the reported or reconstructed tau) and the annotated edited region at the sampling timesteps where the mask is applied. Report mean IoU per editing type, plus the fraction of edits whose thresholded attention-average overlaps the annotated region by at least 0.3. As a second check, rerun Table 2 with the mask replaced by the ground-truth PIE-Bench mask: if the paper's 'Ours' configuration is substantially worse than the ground-truth-mask variant on FID, CLIP-T, or CLIP-I, then the hand-selected attention maps are not reliably localizing edits.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.2 builds the entire localization mechanism on the premise that the sum of FLUX double-block self-attention maps, s_{k,l}, is a reliable spatial indicator of the edit region. After rescaling (Eq. 11), sorting by activation sum (Eq. 12), averaging ranks 10-14 (Eq. 13), thresholding (Eq. 14), and applying the resulting mask to the next step's attention (Eq. 15), the method assumes that the masked region is exactly where the edit should happen and everywhere else should be preserved. The only evidence offered is two visual examples in Fig. 5 and an external observation [47]. No IoU or region-overlap measurement against PIE-Bench's provided edit-region masks is reported, despite the benchmark containing such masks; no per-edit-type breakdown is given; and the indices i=10, j=14, hfactor, rfactor, and tau are hand-set without a held-out validation set. The ten PIE-Bench editing types include global changes such as background and style edits, for which a localized medium-low activation mask is conceptually mismatched. If the mask does not track the target region, the mechanism either suppresses the requested edit (the LQR already pulls sampling toward the original image via Eq. 9 with lambda=0.7) or lets the edit bleed into non-target structure, breaking the very trade-off that the central claim depends on. Supplement C.2 concedes that first-round edits are suboptimal, which is consistent with this fragility. The core assertion of 'significantly improves edit success rates and visual fidelity' therefore rests on an unvalidated localization assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a multi-turn image editing framework built on the FLUX.1-dev rectified flow model. The two stated technical contributions are (i) a dual-objective Linear Quadratic Regulator (LQR) guidance that conditions sampling on both the original image and the previous turn's result, with the goal of preventing error accumulation across editing rounds, and (ii) an adaptive attention highlighting mechanism that selects 'medium-low' activated self-attention maps from the FLUX double blocks to build a spatial mask that localizes edits to salient regions while preserving the rest of the image. The method is evaluated on a GPT-4-Turbo-extended version of PIE-Bench, with quantitative metrics (FID, CLIP-T, CLIP-I, CLIP-Edit, structure distance) and qualitative examples. The central claim is that the framework 'significantly improves edit success rates and visual fidelity compared to existing methods' for iterative, multi-turn editing.","tokens_in":17552,"tokens_out":10783,"duration_ms":100998,"significance":"If the claims are substantiated, the paper would provide a useful tool for interactive, iterative image editing, an area that is underexplored relative to single-turn editing. The dual-objective LQR is a natural and technically sound extension of the single-objective LQR in RF-Inversion, and the idea of exploiting layer-wise activation levels in DiT self-attention to localize edits is interesting and potentially generalizable. The paper also ships a multi-turn benchmark extension that could support future work. However, the significance is currently limited by the partial support for the headline claims: the best-FID result at the fourth turn is offset by non-best CLIP-T and CLIP-I scores, the attention-mask mechanism is not quantitatively validated against ground-truth edit regions, and a number of technical details in the equations need correction. The framework is defensible, but the current evidence does not yet establish a consistent multi-turn advantage.","major_comments":[{"comment":"The mask construction as written is mathematically inconsistent. Eq. (12) sorts the scalar activation values a_{k,l} = sum(s_{k,l}), so the sorted quantities a'_{k,l} are scalars, not 2D attention maps. Eq. (13) then averages these scalars, producing a scalar \\bar{v}_{i:j}; thresholding a scalar in Eq. (14) cannot yield a spatial mask M_k as used in Eq. (15). To implement the method described in the text, Eq. (13) must average the 2D attention maps s'_{k,l} associated with the selected ranks, not their scalar activation sums. This is a load-bearing error: without a spatial mask, the adaptive attention guidance cannot localize edits, and the claimed editability/consistency trade-off is undefined.","section":"Sec. 5.2, Eqs. (12)-(15)"},{"comment":"The central novelty, the adaptive attention mask, is never validated against ground truth. PIE-Bench provides edit-region masks, and the supplement even uses them to compute CLIP-Edit (Sec. D.1), yet the paper reports no IoU or region-overlap between M_k and these masks, and no per-edit-type breakdown. The ten PIE-Bench categories include global edits such as background and style changes, for which a localized 'medium-low' activation mask is conceptually mismatched. The block range i=10, j=14, hfactor, rfactor, and tau are all hand-set without sensitivity analysis or a held-out validation set. The evidence for the mask's localization quality is limited to two visual examples in Fig. 5. Without quantitative localization validation, the claimed balance between editability and multi-turn coherence is unsupported.","section":"Sec. 5.2 and Sec. 6.3"},{"comment":"The claim that the framework 'significantly improves edit success rates and visual fidelity compared to existing methods' is overstated relative to the reported numbers. At the fourth editing turn, the proposed method achieves the best FID (5.553) but not the best CLIP-T (FireFlow-v: 28.281 vs 26.831) or the best CLIP-I (RF-Solver: 0.906 vs 0.894). Supplement Table 3 shows that in rounds 1-3 the method has higher FID than RF-Inv. (e.g., round 1: 2.554 vs 1.854) and lower CLIP-T than FireFlow, and Supplement C.2 concedes that first-round editing is suboptimal. All results are point estimates without error bars or significance tests, and Table 1 contains two rows labeled 'Ours' (15 and 8 steps) without explaining which configuration corresponds to the method described in Sec. 5. A claim of 'significantly improves' requires either stronger quantitative support or a more carefully qualified presentation.","section":"Abstract, Sec. 6.3, Table 1, and Supplement Table 3"},{"comment":"The guidance terms in Eqs. (8) and (9) appear to have a sign error. In Eq. (5) and in the original RF-Inversion formulation, the LQR guidance is \\eta(v_t(X_t|X1) - v_t(X_t)), with a minus sign between the conditional and unconditional velocity fields. In Eqs. (8) and (9), the corresponding terms are written with a plus sign: v_t(X_t|X0) + v_t(X_t) and v_t(X_t|X_dual) + v_t(X_t). If taken literally, this would push the trajectory away from the conditioning image rather than toward it, which contradicts the stated goal of anchoring sampling to the original image and the previous-turn result. Please clarify whether this is a typesetting error and correct the equations; the ablation and Fig. 3 depend on this guidance mechanism.","section":"Sec. 5.1, Eqs. (8)-(9)"}],"minor_comments":[{"comment":"The abstract contains a grammatical error: 'a adaptive attention highlighting method' should read 'an adaptive attention highlighting method'.","section":"Abstract"},{"comment":"The notation v_t(X_t|x') for the conditional velocity field is not explicitly defined. State whether v_t(X_t|X_dual) is the analytic field (X_dual - X_t)/(1-t) from Proposition 1 or a network-conditioned variant; the sign analysis in the paper depends on this definition.","section":"Sec. 5.1, Eqs. (6)-(9)"},{"comment":"The index k is overloaded: it denotes the denoising step in Eq. (10)-(15) but also refers to the editing turn in Eqs. (8)-(9) (e.g., 'the k-th editing step'). Please use distinct symbols to avoid confusion.","section":"Sec. 5.2, Eq. (10)"},{"comment":"The symbol \\bar{v}_{i:j} is easily confused with the velocity field v; rename the averaged attention map, for example to \\bar{A}_{i:j} or \\bar{s}_{i:j}, especially given that scalar activation sums are also denoted a'.","section":"Sec. 5.2, Eq. (13)"},{"comment":"The ablation terminology is confusing: 'low' activation is defined as blocks 12-17, 'medium' as blocks 6-10, and 'high' as the top 5, but Eq. (12) sorts blocks in ascending order of activation, so blocks 12-17 are actually high-activation blocks in that ordering. Clarify the relationship between the ablation definitions and the main 'medium-low' selection (i=10, j=14).","section":"Sec. 6.4"},{"comment":"The captions of several qualitative figures say 'Quantitative Results' (e.g., Fig. 12, 'Quantitative Results on Natural Animals'); these should say 'Qualitative Results'.","section":"Fig. 12 and Supplement Figs. 11, 13, 14"},{"comment":"The FID metric is cited to reference [51], which is the LPIPS paper (Zhang et al., 'The unreasonable effectiveness of deep features as a perceptual metric'). FID should be cited to Heusel et al., 2017; please correct the reference.","section":"References [51]"},{"comment":"The paper acknowledges that evaluation is limited to four editing turns because of instruction-generation errors. The abstract and introduction claim 'multi-turn' and 'iterative refinement' more generally; it would be helpful to state this four-turn limit explicitly in the main text so readers calibrate the scope of the claims.","section":"Supplement Sec. C.1"},{"comment":"The two rows labeled 'Ours' (one with 15 steps and one with 8 steps) are not explained in the table or in the surrounding text. Please label them clearly (e.g., 'Ours (15 steps)' and 'Ours (8 steps)') and specify which configuration is reported in the ablation study.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal, but the novelty relative to RF-Inversion is partly incremental: the dual-objective LQR is essentially a weighted-average re-derivation of the single-objective case, and the main new idea is the attention-mask mechanism. The latter needs careful validation, and the equations currently contain a scalar/2D-map inconsistency that must be fixed. The lack of code or dataset release is a further concern for reproducibility. I would condition acceptance on a thorough revision that addresses the mask construction, adds quantitative mask localization experiments, includes error bars or significance tests, and softens the claims to match the evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real engineering contribution to multi-turn editing, but the paper oversells itself and the attention-mask story needs work before I'd trust it.\n\nWhat's actually new: the multi-turn setting itself is underexplored, and the paper makes a sensible case that single-step inversion/editing methods accumulate drift. Combining a dual-objective LQR with a second-order flow solver is a reasonable extension of RF-Inversion and FireFlow, and the ablation in Table 2 does show the dual objective helps FID. The adaptive attention scheme — selecting medium-low activated FLUX blocks to build a mask — is an interesting idea, and a few qualitative examples look good. The LLM-extended PIE-Bench is a practical contribution, though the generation is only as good as GPT-4 Turbo's instructions.\n\nThe soft spots are where the stress-test note lands. The mask localization is the load-bearing part of the editability/consistency balance, and there is no quantitative evidence that blocks 10–14 activate where the edit should happen. PIE-Bench ships edit-region masks; an IoU report per editing type is the obvious missing experiment. Without it, the hand-set indices and threshold parameters feel fragile, especially for global edits where a medium-low local mask is conceptually mismatched. The supplement admitting first-round edits are suboptimal does not help.\n\nThe empirical claim 'significantly improves' is also overstated. At turn four you lead only on FID; both FireFlow-v (CLIP-T) and RF-Solver (CLIP-I) beat you on the other metrics. With no error bars or significance tests, 'balanced' is a fair summary; 'significantly improves' is not. The four-turn cap is acknowledged but still limits the central claim about multi-turn consistency.\n\nThe LQR proposition is correct but minor — it's a weighted-average reduction of a standard LQR — so the real novelty has to come from the attention mechanism, which is exactly the least-validated part.\n\nBottom line: this deserves peer review, not desk rejection. The problem is relevant, the framework is coherent, and the ablations are directionally right. But the authors need to validate the mask against ground-truth edit regions, report variance, extend the evaluation beyond four turns, and soften the abstract. I'd want to see code and data released before citing their numbers.","headline":"A workable multi-turn editing recipe with a promising dual-objective LQR, but the attention-mask localization is under-validated and the headline claims overstate the numbers.","tokens_in":18196,"tokens_out":2492,"would_cite":false,"duration_ms":23989,"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":"A control loop with two anchors keeps an edited image stable across many edit rounds while keeping edits localized.","keywords":["multi-turn image editing","flow matching inversion","dual-objective LQR","adaptive attention guidance","rectified flow","FLUX DiT","PIE-Bench","image editing consistency"],"falsifier":"Run the method on the full PIE-Bench multi-turn dataset and compare each generated editing mask against the ground-truth edited-region masks; if the average intersection-over-union of the mask with the true edit region is near zero, or if the FID at turn four does not beat the baselines when more than one seed is used, the attention-localization premise is not supported. Also rerun with the mask modules removed after using the attention mask and vice versa.","tokens_in":16986,"feed_emoji":"🖌️","tokens_out":3334,"duration_ms":28327,"temperature":0.7,"pith_summary":"The paper argues that single-step image editing methods fail when users want to refine an image over multiple rounds: each round accumulates errors, drifts from the original, and eventually produces artifacts. The authors propose a multi-turn editing framework that anchors every edit to both the original image and the previous edit result using a dual-objective LQR controller in the flow-matching latent space, and then uses a mask built from mid-level attention maps to keep each edit focused on the intended object. They report that at the fourth editing turn their method achieves the best FID (5.553) among all compared methods while keeping CLIP-I and CLIP-T scores balanced, meaning the image stays both faithful to the source and aligned with the prompt. A sympathetic reader would care because interactive tools for design, retouching, and content creation need this kind of iterative, stable editing.","feed_headline":"Two anchors keep iterative image edits stable for many rounds","feed_subtitle":"A dual-objective LQR plus adaptive attention masks lets users refine images turn after turn without drift or artifacts.","key_machinery":"The central object is the dual-objective LQR update for sampling: at edit turn $k$, the guidance target is $X_{\\text{dual}} = X_{0,0} + \\lambda(X_{k-1,0} - X_{0,0})$, where $X_{0,0}$ is the original image and $X_{k-1,0}$ is the previous edit. This target is plugged into the LQR-derived conditional velocity $v_t(x_t \\mid X_{\\text{dual}}) = (X_{\\text{dual}} - x_t)/(1-t)$, which steers the flow back toward the original while allowing the previous edit to influence the change. The second component is the adaptive attention mask $M_k$, computed by sorting the 19 double-block self-attention maps by their activation sums, averaging maps 10 through 14 (the medium-low activation ones), and thresholding the average to produce a binary mask. This mask is applied element-wise to the attention matrix at the next step, localizing the edit. Together they balance the tension between preservation and editability that limits single-anchor LQR methods.","core_discovery":"The central claim is that multi-turn image editing can be made stable by combining three components: second-order flow-matching inversion for accurate image-to-noise mapping, a dual-objective linear quadratic regulator (LQR) that guides sampling toward a weighted blend of the original image and the previous edit result, and an adaptive attention mask that reweights only the image regions the edit should affect. The dual-objective LQR is derived by showing that the multi-target optimal control problem reduces to a single-target LQR whose effective target is the weighted average of the individual targets. The attention mask is formed by taking the sum of selected medium-low activation self-attention maps from the FLUX double blocks, averaging them, and thresholding, then element-wise multiplying this mask into the next attention computation. The paper reports that this combination prevents error accumulation across edit turns, preserves the original image's structure and distribution, and still performs the requested edits, with the best FID at turn four.","pith_inferences":["The dual-anchor idea is likely transferable to video editing: treating the first frame as the fixed anchor and the previous edited frame as the drifting target could reduce temporal drift across many frames.","The hand-picked layer range (10 to 14) for the attention mask could instead be learned from ground-truth edit masks, which would likely make the method work across more diverse edit types and prompts.","Because the LQR guidance alone can over-regularize, the authors' attention mask is what controls editability; a user-defined or prompt-derived mask might allow even more precise control in mixed edits such as changing one region's color while keeping another unchanged.","A fair test of the mask quality would be to measure the intersection over union between the generated mask and the human-annotated edit region in PIE-Bench; the authors only show visual examples."],"forward_implications":["If the method works as claimed, interactive image editors can support multi-turn refinement without users having to restart from scratch after each failed edit.","The dual-objective LQR formulation gives a principled way to anchor generation to both a fixed reference and a drifting target, which should reduce error accumulation over long sequences.","The finding that medium-low activation self-attention maps are the ones that localize fine details suggests that layer-wise attention analysis can guide where edits are applied in rectified flow transformers.","The new multi-turn extension of PIE-Bench provides a benchmark for future work to measure drift, editability, and quality over successive editing rounds.","The method needs only 15 steps for inversion and sampling, comparable to FireFlow's 8 steps and far fewer than the 28-50 steps used by several baselines."],"supporting_citations":[{"why":"RF-Inversion supplies the single-objective LQR control baseline that the dual-objective LQR extends.","marker":"[40]"},{"why":"FireFlow supplies the second-order ODE solver and the acceleration scheme used implement the flow matching inversion and sampling.","marker":"[11]"},{"why":"RF-Solver is the second-order solver baseline whose accumulated error the method aims to reduce.","marker":"[44]"},{"why":"Xu et al. provides the observation that FLUX's lower-left self-attention quadrant encodes text-to-image spatial influence, which motivates the mask selection.","marker":"[47]"},{"why":"PIE-Bench and Direct Inversion provide the single-turn editing benchmark and inversion strategy used to build the multi-turn evaluation dataset.","marker":"[22]"},{"why":"Rectified flow provides the ODE formulation that underlies the inversion and sampling process.","marker":"[30]"}],"fun_headline_variants":["Adaptive attention keeps iterative image edits on track","Stable multi-turn editing via flow matching and LQR","Multi-turn image editing without drift or artifacts","Consistent multi-turn image editing with no drift"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The assumption that the sum of a FLUX double-block self-attention map reliably indicates where the edit should happen, so that averaging the medium-low activated maps and thresholding gives a correct editing mask.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive attention keeps iterative image edits on track","Stable multi-turn editing via flow matching and LQR","Multi-turn image editing without drift or artifacts","Consistent multi-turn image editing with no drift"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000475,"raw_usage":{"total_tokens":2326,"prompt_tokens":884,"completion_tokens":1442,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":1382}},"tokens_in":500,"tokens_out":1442,"duration_ms":11039,"temperature":1.0,"reasoning_tokens":1382,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:31:34.042387+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the method on the full PIE-Bench multi-turn dataset and compare each generated editing mask against the ground-truth edited-region masks; if the average intersection-over-union of the mask with the true edit region is near zero, or if the FID at turn four does not beat the baselines when more than one seed is used, the attention-localization premise is not supported. Also rerun with the mask modules removed after using the attention mask and vice versa.","supporting_citations":[{"cited_title":"Semantic Im- age Inversion and Editing using Rectified Stochastic Differ- ential Equations, 2024","cited_arxiv_id":null,"evidence_quote":"RF-Inversion supplies the single-objective LQR control baseline that the dual-objective LQR extends."},{"cited_title":"FireFlow: Fast Inversion of Rectified Flow for Image Semantic Editing, 2024","cited_arxiv_id":null,"evidence_quote":"FireFlow supplies the second-order ODE solver and the acceleration scheme used implement the flow matching inversion and sampling."},{"cited_title":"Taming Rectified Flow for Inversion and Editing, 2024","cited_arxiv_id":null,"evidence_quote":"RF-Solver is the second-order solver baseline whose accumulated error the method aims to reduce."},{"cited_title":"Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow, 2022","cited_arxiv_id":null,"evidence_quote":"Rectified flow provides the ODE formulation that underlies the inversion and sampling process."}],"review_version":1}