{"id":"fc1cfdf5-3713-4600-b753-0217b6750830","arxiv_id":"2412.21079","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Edicho makes edits consistent across in-the-wild image pairs by injecting explicit pixel correspondences into the attention and classifier-free guidance steps of a pretrained diffusion model, with no training.","lead":"Edicho is a training-free method that edits two or more real-world photos consistently, using a pixel correspondence map to share features between images while a diffusion model redraws them. It works as a plug-in for existing editing tools such as ControlNet and BrushNet, which matters for product photos, marketing images, and 3D reconstruction from edited views.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The EC metric is partially self-fulfilling: Edicho's hard copying of source features and latents mechanically raises CLIP similarity, so the quantitative outperformance in Table 1 is not independent evidence of better consistency.","rationale":"The paper makes a useful and plausible engineering contribution: explicit correspondence is a sensible prior for multi-image editing, and the ablations in Fig. 6 and Fig. S1 support the two proposed components. I do not see an internal inconsistency in the main equations, and the DIFT-accuracy concern raised by the reader is legitimate but is explicitly acknowledged as a limitation by the authors. The more load-bearing issue is that the primary quantitative evidence for the central claim, the EC metric, is not an independent measure of consistency. Because Edicho's attention and CFG modifications directly copy source features and latents, high EC is a designed consequence of the algorithm. This does not disprove the method, but it means Table 1 cannot support the headline comparison against implicit-correspondence baselines. The random-correspondence ablation would settle whether the metric is sensitive to the actual correspondence quality, and adding per-pair error bars would address the missing statistical grounding. User studies are present but underreported; without details on image sets, question design, and significance, they cannot substitute for a controlled metric. The verdict should remain CONDITIONAL: the method is plausible and the mechanism is coherent, but the quantitative claim needs an independent consistency metric before the comparative superiority is accepted.","tokens_in":13299,"tokens_out":8307,"duration_ms":90880,"concrete_test":"Run Edicho with the correspondence C_{i,j} replaced by a random permutation of the same correspondence (or by an identity, warp-free copy that sends every target query to the source query at the same grid position), keeping all other components and hyperparameters fixed. If the EC score on the test set stays near 0.9355 or 0.8931 while qualitative consistency degrades, the EC metric is insensitive to correspondence accuracy and the quantitative superiority claim is an artifact of feature copying. Independently, recompute EC with standard errors over individual image pairs and add a no-correspondence copy-source baseline to Table 1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim rests on Table 1, where editing consistency (EC) is defined as CLIP feature similarity between the edited images. Edicho's own mechanism makes high EC scores almost inevitable: Eq. (4) replaces the target's self-attention output with features attending only to source keys and values after warping queries, and Eq. (8) injects source latents into the target's unconditional CFG branch. Both operations force the target's internal representation, and therefore the decoded image, to move toward the source output in any feature space, including CLIP. A larger EC value can thus reflect stronger feature copying rather than more perceptually consistent edits. The implicit-correspondence baselines (MasaCtrl, Cross-Image-Attention) also borrow source keys and values, but Edicho adds query warping and latent injection, so the reported margin (0.9355 vs. 0.9258 for global editing) cannot be attributed to explicit correspondence without a control. The user study in App. E is more direct, but it is reported only as percentages without error bars or significance testing, and the exact protocol is underspecified. Since the claim of outperforming implicit-correspondence methods is a core part of the abstract and Section 4.2, this metric circularity is the most load-bearing weakness; DIFT misalignment is a real but secondary limitation that the authors already concede in the Conclusion.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents Edicho, a training-free framework for consistent image editing across multiple images. The method first computes explicit semantic correspondence between input images using a pre-trained extractor (DIFT). It then injects this correspondence into a pre-trained diffusion model in two ways: (i) Corr-Attention, which warps the target's query features to the source according to the correspondence and borrows source keys and values in self-attention (Eqs. 3–4), and (ii) Corr-CFG, which fuses source latents into the unconditional branch of classifier-free guidance (Eqs. 7–9) to enhance consistency. Edicho is designed to be plug-and-play with existing editing models such as ControlNet and BrushNet. The paper reports qualitative results, a user study, and quantitative TA/EC scores for local and global editing, plus applications to customization and 3D reconstruction.","tokens_in":13576,"tokens_out":6626,"duration_ms":55907,"significance":"Edicho addresses a relevant problem, and the core idea—using explicit correspondence to guide attention and CFG—is well motivated and distinct from prior implicit-correspondence approaches. The method is training-free, model-agnostic, and appears to work across diverse in-the-wild inputs. The qualitative results in Figs. 4–5 are often convincing, and the user study (App. E) provides some evidence of preference. However, the quantitative evaluation is currently too weak to establish the claimed improvements: the EC metric is partially circular with the editing mechanism, Table 1 lacks error bars and sample sizes, and the user study lacks statistical detail. For these reasons, the empirical support for the central claim requires strengthening, though the underlying approach is plausible and likely to be useful.","major_comments":[{"comment":"The editing consistency (EC) metric measures CLIP feature similarity between edited images. Edicho's design directly promotes such similarity by copying source features into the target: Eq. (4) makes the target's attention output at each selected step attend only to source keys and values, and Eq. (8) injects source latents into the target's unconditional branch. Both operations push the target's internal representation, and therefore the decoded image, toward the source in any feature space, including CLIP. Consequently, the reported EC margins (e.g., 0.9355 vs. 0.9258 for global editing) cannot be interpreted as independent evidence of better perceptual consistency. To support the claim of outperforming implicit-correspondence baselines, please add control experiments (e.g., ablating the query warping or latent injection and reporting EC), compute a consistency metric that is not trivially increased by feature copying, and report error bars and significance tests.","section":"§4.2, Table 1"},{"comment":"The quantitative evaluation is statistically underspecified. Table 1 reports single point estimates with no indication of the number of test image pairs, no variance, and no significance testing, yet the text in §4.2 calls the improvements 'significant.' The user study in App. E reports only overall vote percentages (over 60% for the proposed method) from 30 participants answering up to 20 questions, but does not state the number of underlying image sets, the randomization procedure, or the variability across participants. Please provide the full evaluation protocol, per-participant breakdown, and appropriate statistical tests so the central quantitative claims can be verified.","section":"§4.1–4.2, App. E"},{"comment":"The method's correctness rests on the accuracy of the precomputed correspondence C_ij from DIFT. If the correspondence is misaligned due to large pose changes, occlusions, or lighting differences, the warped queries in Eq. (3) and the latent injection in Eq. (8) transport mismatched content, which the authors partially concede in the Conclusion ('sometimes the generated textures would be inconsistent due to the correspondence misalignment'). The manuscript does not quantify the frequency or severity of such failures, nor does it provide a sensitivity analysis to correspondence quality. Please add a failure analysis or a sensitivity study (e.g., varying the correspondence extractor, or artificially corrupting C_ij) to delineate when the method is reliable.","section":"§3.3–3.4, Conclusion"}],"minor_comments":[{"comment":"The notation for the fusing function T is inconsistent: Eq. (7) calls T with three arguments (two noise predictions and C_ij), while Eq. (8) defines T with four arguments (z_i, z_j, C_ij, γ). Please align the notation and explicitly define the injection function Inj.","section":"Eqs. (7)–(8)"},{"comment":"The section title contains a typo: 'Classifer-free Guidance' should be 'Classifier-free Guidance'.","section":"Section 3.4 title"},{"comment":"The reference entry for Adobe Firefly contains a typo: 'Adobe reseachers' should be 'Adobe researchers'.","section":"Reference [45]"},{"comment":"The implicit correspondence is extracted at different layer/step pairs for different cases (e.g., (1,10), (2,15), (4,25)), so the qualitative comparison in Fig. 2 is not apples-to-apples. Please use a fixed set of layers/steps or report the best-performing configuration for the implicit method.","section":"Fig. 2 and App. D"},{"comment":"The phrase 'Where the Warpfunction' should be 'where the Warp function' for correct capitalization and spacing.","section":"Eq. (3)"},{"comment":"The hyperparameters λ=0.8 and γ=0.9 are set without any sensitivity analysis; a small ablation over these values would help readers understand the method's robustness.","section":"App. B"}],"recommendation":"major_revision","confidential_remarks":"The main obstacle to acceptance is the circularity of the EC metric, which makes the central quantitative claim difficult to evaluate. I would like the editor to ensure the authors treat this concern seriously and provide non-circular evidence. The paper is otherwise within scope and the idea is solid."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the core idea is genuine and clean: instead of relying on implicit attention-derived correspondences like MasaCtrl, StyleAligned, and Cross-Image-Attention, Edicho pre-computes explicit correspondence (via DIFT or similar) and injects it into both the self-attention (warping queries, borrowing source K/V, Eq. 3-4) and the unconditional CFG branch (latent injection with weighted fusion, Eq. 8). That is a real algorithmic contribution, and it is clearly explained. Second, the qualitative results are the paper's real evidence. The figures show consistent edits across pose, lighting, and background variation, and the ablations—warping queries vs. outputs, unconditional-only vs. both-branch CFG—are informative and support the design choices. The user study (over 60% preference, 500 votes) is decent supporting evidence, though the protocol is under-specified.\n\nThe soft spots are real but mostly fixable. The biggest one: the EC metric is partially self-fulfilling. Eq. (4) hard-copies source keys/values and Eq. (8) injects source latents into the unconditional branch, so any feature-based similarity measure—CLIP included—will mechanically rise. Table 1 reports single point estimates with no error bars, significance tests, or stated test-set size, so the margin over MasaCtrl (0.9355 vs. 0.9258) could partly reflect feature copying rather than perceptual consistency. The DIFT misalignment concern is secondary but honest; the authors themselves concede inconsistent textures under correspondence misalignment in the Conclusion. Also missing: code release (promised), hyperparameter sensitivity (lambda, gamma, step/layer ranges), and a fixed-seed comparison to isolate the method's effect.\n\nNone of this sinks the paper. The central mechanism is plausible, the method is training-free and model-agnostic, and the qualitative demonstrations are strong. The quantitative evaluation just needs to be upgraded to something that doesn't entangle the metric with the mechanism—e.g., human preference with confidence intervals, or a perceptual metric that doesn't rely on CLIP similarity of the output images themselves.\n\nWho should read this: anyone working on multi-image editing, appearance transfer, or correspondence-guided diffusion. It deserves a serious referee, but with a clear request for stronger evaluation. I'd cite it once the code is out and the evaluation is tightened.","headline":"Edicho's explicit-correspondence-guided attention and CFG is a clean, novel twist on zero-shot consistent editing; the mechanism is plausible and the qualitative work is strong, but the quantitative evidence is thinner than the claims.","tokens_in":14119,"tokens_out":622,"would_cite":true,"duration_ms":7820,"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":"Edicho claims that consistent multi-image editing in the wild is achievable training-free by injecting explicit, pre-computed correspondence into self-attention and classifier-free guidance.","keywords":["consistent image editing","training-free editing","diffusion models","explicit correspondence","self-attention manipulation","classifier-free guidance","zero-shot image editing","diffusion-based correspondence"],"falsifier":"Replace the correspondence map generated by the extractor with a random or corrupted map and run Edicho on the same inputs; if editing-consistency scores (e.g., the paper's EC metric) do not visibly degrade, then the claimed mechanism is not carrying the result, whereas a sharp drop would confirm that the method's success is tied to the accuracy of explicit correspondence.","tokens_in":1622,"feed_emoji":"🖼️","tokens_out":3009,"duration_ms":77947,"temperature":0.7,"pith_summary":"The paper sets out to show that consistent editing of multiple real-world photographs—so that the same prompt produces matching changes across images—can be achieved without any training or fine-tuning, by letting a pre-computed correspondence map between the images steer the denoising process of an off-the-shelf diffusion model. The proposed method, Edicho, alters two places in the diffusion pipeline: self-attention, where query features are warped and source keys and values are borrowed, and the unconditional branch of classifier-free guidance, where source latents are injected. The authors argue that this explicit, pre-estimated correspondence is fundamentally more reliable than the implicit correspondences inferred from attention weights on which prior training-free methods rely, and that the scheme plugs into existing editing backbones. If the claim holds, users can apply the same edit to many photos—product shots, event photos, character designs—and get uniform results despite differences in pose, lighting, and background.","feed_headline":"Zero-shot edits stay consistent across wild photos","feed_subtitle":"Edicho injects pre-computed correspondences into attention and CFG, so one prompt edits many photos uniformly.","key_machinery":"The mechanism is correspondence-warped self-attention combined with correspondence-guided classifier-free guidance. In the attention module, the query feature of the target image is warped into the source image's coordinate frame using the prestored correspondence map, and attention reads keys and values from the source, so coherent features are transferred across images. In the guidance module, the unconditional branch of the denoising step is modified by randomly injecting a portion of the source noisy latent into the target's unconditional latent, keeping the conditional branch untouched; this design choice is intended to preserve the generative prior. Both manipulations are applied only in a middle range of denoising steps and from a given attention layer onward, preserving the untouched generative prior.","core_discovery":"The central claim is that injecting explicit pre-estimated correspondence into the diffusion denoising process yields consistent edits across in-the-wild images, outperforming implicit-correspondence methods. Concretely, Edicho computes a correspondence map between source and target images, then in self-attention warps the target query by that map and queries the source keys and values, so that each edited location attends to its matching source location rather than to wherever attention happens to land. In classifier-free guidance, it modifies only the unconditional branch, fusing source latents into the target's unconditional noise estimate, which is reported to preserve the model's generative priors while improving consistency. When plugged into a Stable Diffusion backbone with existing editing backbones, this produces text-aligned edits that are consistent in texture, object counts, and style across image pairs and sets, as measured by CLIP-based scores and user studies.","pith_inferences":["If the mechanism is as robust as claimed, a natural extension is to apply it to video frames or multi-view images of the same scene, where correspondence is easier to compute and temporal consistency is a similar problem.","The method effectively decouples what to edit from where to edit consistently, suggesting that any edit operator—not just text-to-image backbones—could be made consistent by the same two injection points.","A possible failure mode not fully explored in the paper is occlusion: because the warp assumes one-to-one correspondence, images with occlusions may copy features from the occluder; a visibility-aware warp would be a testable improvement.","The fixed denoising-step and layer range is likely task-dependent; an adaptive schedule chosen based on where the correspondence extractor is most reliable could improve results further."],"forward_implications":["The method is model-agnostic: it can be layered onto any Stable-Diffusion-based editing backbone without retraining, so future editing models inherit the consistency mechanism.","Explicit correspondence prediction is more reliable than implicit attention-derived correspondence, so consistent editing should be built on precomputed matching rather than on attention similarity.","Consistent edits can serve as training data for concept customization and as inputs for 3D reconstruction, extending the usefulness of the editing outputs.","Because the method is zero-shot, it can be applied to arbitrary image sets without per-set optimization, making batch photo editing practical.","The method's consistency quality is directly tied to the accuracy of the correspondence extractor, so improving extractors should directly improve editing consistency."],"supporting_citations":[{"why":"Produces the explicit correspondence maps between image pairs that Edicho injects into attention and CFG.","marker":"[50]"},{"why":"Baseline method that infers correspondence implicitly from attention weights; Edicho's comparison point for the explicit-vs-implicit argument.","marker":"[1]"},{"why":"Baseline tuning-free mutual self-attention editing method that Edicho compares against in global editing.","marker":"[7]"},{"why":"Baseline shared-attention method for style-aligned generation that Edicho compares against.","marker":"[18]"},{"why":"Inspires the choice to fuse features only in the unconditional branch of classifier-free guidance.","marker":"[36]"},{"why":"Provides the inpainting reference network used for local editing tasks.","marker":"[25]"},{"why":"Provides the conditioning reference network used for global editing tasks.","marker":"[60]"},{"why":"Supplies the pretrained latent diffusion backbone whose attention and CFG modules are modified.","marker":"[47]"}],"fun_headline_variants":["Edicho: one prompt edits many wild photos consistently","Consistent in-the-wild edits via explicit correspondences","Training-free trick aligns diffusion edits across images","Attention warping keeps edits uniform across photos"],"cache_read_input_tokens":16256,"weakest_assumption_plain":"The method assumes the pre-computed correspondence map is accurate and remains aligned with the diffusion model's latent feature space at the layers and steps where editing is applied, so that warped queries and injected latents copy the right features.","fun_headline_variants_meta":{"raw":{"variants":["Edicho: one prompt edits many wild photos consistently","Consistent in-the-wild edits via explicit correspondences","Training-free trick aligns diffusion edits across images","Attention warping keeps edits uniform across photos"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000268,"raw_usage":{"total_tokens":1567,"prompt_tokens":842,"completion_tokens":725,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":458,"completion_tokens_details":{"reasoning_tokens":665}},"tokens_in":458,"tokens_out":725,"duration_ms":7239,"temperature":1.0,"reasoning_tokens":665,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:02:52.955934+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the correspondence map generated by the extractor with a random or corrupted map and run Edicho on the same inputs; if editing-consistency scores (e.g., the paper's EC metric) do not visibly degrade, then the claimed mechanism is not carrying the result, whereas a sharp drop would confirm that the method's success is tied to the accuracy of explicit correspondence.","supporting_citations":[{"cited_title":"Emergent correspondence from image diffusion","cited_arxiv_id":null,"evidence_quote":"Produces the explicit correspondence maps between image pairs that Edicho injects into attention and CFG."},{"cited_title":"Cross-image attention for zero-shot appearance transfer","cited_arxiv_id":null,"evidence_quote":"Baseline method that infers correspondence implicitly from attention weights; Edicho's comparison point for the explicit-vs-implicit argument."},{"cited_title":"MasaCtrl: Tuning-free mutual self-attention control for consistent image synthesis and editing","cited_arxiv_id":null,"evidence_quote":"Baseline tuning-free mutual self-attention editing method that Edicho compares against in global editing."},{"cited_title":"Style aligned image generation via shared atten- tion","cited_arxiv_id":null,"evidence_quote":"Baseline shared-attention method for style-aligned generation that Edicho compares against."},{"cited_title":"Null-text inversion for editing real images using guided diffusion models","cited_arxiv_id":null,"evidence_quote":"Inspires the choice to fuse features only in the unconditional branch of classifier-free guidance."},{"cited_title":"Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion","cited_arxiv_id":null,"evidence_quote":"Provides the inpainting reference network used for local editing tasks."},{"cited_title":"Adding conditional control to text-to-image diffusion models","cited_arxiv_id":null,"evidence_quote":"Provides the conditioning reference network used for global editing tasks."},{"cited_title":"High-resolution image synthesis with latent diffusion models","cited_arxiv_id":null,"evidence_quote":"Supplies the pretrained latent diffusion backbone whose attention and CFG modules are modified."}],"review_version":1}