{"id":"1fd384e5-dd8d-46d7-8140-2d968f89d2f4","arxiv_id":"2506.20155","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ReEdit transfers exemplar-based edits to new images by conditioning Stable Diffusion on a LLaVA-written caption plus a CLIP edit-direction vector, with no per-example optimization.","lead":"ReEdit is a new image editing method that copies an edit from a before-and-after image pair onto a new photo without any training or optimization. It is about four times faster than the previous optimization-based approach and achieves competitive scores on several quality metrics.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The CLIP edit-direction vector in Section 2a is assumed to transfer linearly across content images, but this is never validated or ablated; without a test, the claimed benefit of the image-space conditioning is unsupported.","rationale":"The reader and I identify the same soft spot: the image-space edit vector is central to the proposed method but entirely unvalidated. I considered whether the evaluation weaknesses (self-curated dataset, no significance tests, no code release) are more load-bearing, but those affect the strength of the empirical claim, not the mechanism itself; the paper's scientific contribution would still be interesting if the mechanism were sound and the evaluation merely incomplete. Conversely, if Δ_img is not transferable, the proposed image-space capture is inert or misleading, the method effectively reduces to text-plus-feature injection, and the claimed advantage over text-only baselines is unexplained. The concern is falsifiable with data already in the paper's pipeline, so it is the right target for a stress-test. The absence of an ablation or alignment analysis is not proof of failure, but it is a genuine gap in the central argument; the appropriate disposition remains a conditional acceptance pending the proposed check, consistent with the reader's verdict.","tokens_in":5862,"tokens_out":8130,"duration_ms":90467,"concrete_test":"On the paper's 170 curated triples, compute the cosine similarity between Δ_img = pool(E_img(x_edit) − E_img(x)) and Δ_target = pool(E_img(y_edit_gt) − E_img(y)) after L2-normalizing each direction. Report the mean and distribution, and compare against a control formed by pairing each y with a random exemplar's Δ_img. If the source direction is not significantly more aligned with the true target direction than the random control, the transferability premise of Section 2a is empirically false; if it is significantly aligned, the premise is supported and the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2a defines Δ_img = pool(E_img(x_edit) − E_img(x)) and feeds g = concat(Δ_img, E_text(g_caption)) into Stable Diffusion's cross-attention. Two unstated premises are load-bearing. First, that a CLIP embedding difference computed on one exemplar pair is a portable edit direction: the same semantic edit applied to a different content image y should induce an approximately collinear shift in CLIP space. CLIP embeddings are pooled, global representations, so the difference between x and x_edit also encodes identity, pose, background, and lighting changes; there is no a priori reason this direction transfers to y. Second, that a pooled image difference concatenated with a text embedding can be interpreted by SD's cross-attention as conditioning; the paper does not specify how this vector is shaped into the expected 77×768 token sequence or whether a projection is learned. Neither premise is tested: there is no ablation with Δ_img removed or randomized, no analysis comparing Δ_img to the target shift E_img(y_edit_gt) − E_img(y), and no release of code or data. If the image branch is inert, the reported LPIPS/SSIM gains over VISII and IP2P are attributable to the LLaVA caption plus Plug-and-Play feature injection, and the paper's claimed image-space contribution is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes ReEdit, an inference-time, optimization-free pipeline for exemplar-based image editing. Given an exemplar pair (x, x_edit) and a content image y, ReEdit uses a multimodal VLM (LLaVA) to verbalize the edit as a text description g_caption, computes a CLIP image-space edit vector Δ_img = pool(E_img(x_edit) - E_img(x)), concatenates it with the text embedding to form g, and conditions Stable Diffusion with g while injecting features and modifying self-attention via Plug-and-Play [32] to preserve structure. The authors curate a 170-example dataset and report quantitative results on seven metrics against VISII and Instruct-pix2pix baselines, claiming an approximately 4x speedup and superior LPIPS/SSIM.","tokens_in":6200,"tokens_out":6796,"duration_ms":64293,"significance":"If the claims hold, ReEdit would be a fast, training-free alternative to optimization-based exemplar editing, and the curated dataset could support standardized evaluation. However, the significance is currently limited by the untested image-space conditioning, the narrow baseline set, and the statistical reporting; the paper's central contribution is plausible but not yet established.","major_comments":[{"comment":"The image-space edit vector Δ_img is the central novelty of the method, but its construction and usage are underspecified and unvalidated. The definition Δ_img = pool(E_img(x_edit) - E_img(x)) and the final embedding g = concat(Δ_img, E_text(g_caption)) do not specify the dimension of Δ_img, the nature of the pool operator, or how the concatenated vector is shaped into the token sequence (typically 77×768) expected by Stable Diffusion's cross-attention. More importantly, no ablation removes or randomizes Δ_img, so the contribution of the image branch is unknown. Without such an ablation, the claim in Section 2a that 'Both the image and text conditioning in g work in tandem' is unsupported; the reported gains could be attributable entirely to the LLaVA caption plus Plug-and-Play feature injection.","section":"Section 2a"},{"comment":"The uncertainty representation in Table 1 is uninterpretable: the table reports 'mean ± coefficient of variance', which is not a valid way to express dispersion for these metrics. Coefficient of variation is a dimensionless ratio, so e.g. LPIPS 0.16±0.4812 does not communicate a standard deviation or confidence interval. No significance tests are reported, yet Section 1 claims 'significant improvements in quantitative scores'; this claim is not supported. In addition, FID is reported as a single number without variance, which is unreliable for a dataset of only 170 images.","section":"Section 3, Table 1"},{"comment":"The baseline set is too narrow to support the paper's claim of outperforming exemplar-based editing methods. The introduction lists ImageBrush [36], Paint by Example [35], Diffusion Image Analogies [31], and Visual Style Prompting [14], but the experiments compare only against VISII and Instruct-pix2pix. Including at least one recent exemplar-based baseline is necessary to contextualize the reported gains and to demonstrate that the improvements are not specific to the chosen comparisons.","section":"Section 3"},{"comment":"The curated dataset of 170 examples, presented as a tangible contribution, is described only by its sources (Instruct-Pix2Pix, HQ-Edit, Imagic). No details are provided on the distribution of edit types, the selection criteria, or the release plan, and the ground-truth images come from heterogeneous datasets with different edit definitions. This prevents reproduction and makes the combined evaluation difficult to interpret. If the dataset is to be a contribution, its composition and access should be specified.","section":"Section 3"}],"minor_comments":[{"comment":"The prompts p1 and p2 used to query LLaVA are not provided, and the 20-word limit for g_caption is introduced without justification.","section":"Section 2a"},{"comment":"The description of feature injection and self-attention modification is vague; specifying 'fourth layer' and 'layers 4 to 11' in terms of the U-Net architecture would improve reproducibility.","section":"Section 2b"},{"comment":"The claim that the method is 'independent of the base diffusion model' is not demonstrated, as experiments use only Stable Diffusion.","section":"Section 1"},{"comment":"The qualitative analysis would benefit from failure cases and a clearer explanation of what each row shows.","section":"Figure 2"},{"comment":"The claim '~4x faster' is based on a single runtime comparison; reporting the hardware and full pipeline timing would strengthen it.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely problem and the proposed approach is plausible, but the current experimental validation is insufficient for the claims made. The missing ablation of the image-space edit vector is the most serious concern, as it leaves the central contribution unsupported. I recommend major revision; the authors should be asked to add ablations, improve statistical reporting, and broaden the baseline comparisons."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nReEdit combines three known pieces—LLaVA captioning, CLIP edit-direction arithmetic, and Plug-and-Play attention/feature injection—into an end-to-end recipe that is optimization-free. That specific integration is new against the cited baselines, and the speed claim (about 140s vs 9 minutes for VISII) is plausible and practically useful. The paper also collates a 170-example dataset, which would be a real community resource if released.\n\nThe soft spots are real, though. The central conditioning vector Δ_img = pool(E_img(x_edit) − E_img(x)) is assumed to transfer linearly to a different content image y, and this is never tested. The paper doesn't ablate the image branch, doesn't randomize it, doesn't compare it to the actual target shift on y, and doesn't even say how the pooled vector is reshaped into the expected 77×768 token sequence for cross-attention. If that branch is inert, the reported gains over VISII and IP2P come from the LLaVA caption plus Plug-and-Play, and the claimed image-space contribution is unsupported.\n\nEvaluation is also thin: one table, no significance tests, an unusual 'mean ± coefficient of variance' presentation, only two real baselines, and no code or data release. The Dir. Similarity metric is particularly awkward because the method explicitly conditions on the same CLIP edit direction it is then scored against—so that metric is partly circular. The paper also claims to be 'independent of the base diffusion model' but tests only Stable Diffusion.\n\nThat said, the paper is honest about prior art, the idea is not crazy, and the qualitative examples look decent. The central claim is unproven but not obviously false. I'd send it to peer review with a strong request for an ablation of the image branch, full implementation details of the conditioning, proper error bars, and code/data release. As it stands, I wouldn't cite it in my own work yet.\n\nRecommendation: engage, but require the missing evidence before trusting the claims.","headline":"ReEdit is a reasonable, mostly-incremental combination of LLaVA captioning, CLIP edit-vector conditioning, and Plug-and-Play injection, with a plausible speed advantage over VISII, but the load-bearing image-space conditioning is never validated or ablated.","tokens_in":6690,"tokens_out":1680,"would_cite":false,"duration_ms":19288,"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":"ReEdit claims that exemplar-based image editing can be made optimization-free by combining a VLM's textual description with a CLIP-space edit vector, and that this is both higher quality and about 4x faster than the prior…","keywords":["exemplar-based image editing","diffusion models","multimodal VLMs","CLIP embedding space","attention and feature injection","LLaVA","Stable Diffusion","optimization-free inference"],"falsifier":"One decisive experiment is to ablate the image-space branch by setting $\\Delta_{\\mathrm{img}}$ to zero and keeping only the LLaVA text caption; if the edited outputs do not degrade on LPIPS and human preference, the CLIP delta is not doing the claimed work. A complementary check computes $\\Delta_{\\mathrm{img}}$ for the same semantic edit using several different exemplar pairs; if those vectors have low cosine similarity, the transferability assumption behind the method fails.","tokens_in":5714,"feed_emoji":"🎨","tokens_out":14817,"duration_ms":133330,"temperature":0.7,"pith_summary":"The paper tries to establish that the edit shown in a before/after image pair can be transferred to a new photograph without any per-example optimization. Its proposed pipeline, ReEdit, captures the edit twice: as a text caption generated by a multimodal VLM (a model that reads images and writes natural language), and as a difference vector in CLIP embedding space, the shared image-and-text representation used by the model. A frozen text-to-image diffusion model is then conditioned on this combined signal while attention and feature injection preserve the content image's structure. On a 170-example benchmark the method beats optimization-based baselines on several structural and perceptual metrics and runs about four times faster. If correct, this makes exemplar editing cheap, interactive, and automatically upgraded whenever better off-the-shelf diffusion models appear.","feed_headline":"Before/after pair becomes an edit: no training, 4x faster","feed_subtitle":"The method transfers an edit to new photos in about 140 seconds using off-the-shelf models, beating optimization baselines.","key_machinery":"The load-bearing object is the edit embedding $g$ formed by concatenating a text embedding and an image-space edit direction. The image direction $\\Delta_{\\mathrm{img}} = \\mathrm{pool}(E_{\\mathrm{img}}(x_{\\mathrm{edit}})-E_{\\mathrm{img}}(x))$ is the difference of CLIP embeddings of the edited and original exemplars after pooling to the required size. Complementary to that, the VLM verbalization $g_{\\mathrm{caption}}$ is encoded by the CLIP text encoder. The second mechanism is structural conditioning: DDIM inversion of the content image yields a noise latent, and the denoiser runs twice—once to record feature maps and self-attention entries, and once conditioned on $g$ while injecting those features at the fourth upsampling layer and replacing keys and queries in self-attention layers 4 to 11, following the Plug-and-Play features recipe.","core_discovery":"The central claim is that a pretrained text-to-image diffusion model can apply the edit of an exemplar pair to a new image with no training or optimization. The edit is represented by $g = \\mathrm{concat}(\\Delta_{\\mathrm{img}}, E_{\\mathrm{text}}(g_{\\mathrm{caption}}))$, where $\\Delta_{\\mathrm{img}} = \\mathrm{pool}(E_{\\mathrm{img}}(x_{\\mathrm{edit}}) - E_{\\mathrm{img}}(x))$ is the pooled CLIP image-embedding difference and $E_{\\mathrm{text}}(g_{\\mathrm{caption}})$ is the CLIP embedding of the VLM-generated caption. The content image is DDIM-inverted, and a parallel denoising run conditioned on $g$ injects the inverted features and swapped self-attention from the content image, so the edit lands while the surrounding structure stays intact. The authors report that this pipeline outperforms the VISII optimization baseline and Instruct-Pix2Pix variants on LPIPS, SSIM, human preference score, CLIP score, and directional similarity, and is roughly 4x faster than VISII.","pith_inferences":["A natural extension the paper does not pursue is edit composition: if the CLIP-space delta is a true transferable direction, adding or interpolating deltas from several exemplar pairs would perform multiple edits on one image in a single denoising pass.","The mechanism suggests a direct control: scaling $\\Delta_{\\mathrm{img}}$ should scale how strongly the edit is applied, giving an intensity slider that the paper does not report.","Because the pipeline reuses off-the-shelf components, its ceiling is set by the VLM's ability to verbalize fine visual changes and the diffusion model's semantic alignment, so upgrading either component should directly improve edit fidelity.","The 170-example dataset is assembled from instruction-editing corpora; whether the method retains its advantage on purpose-built exemplar datasets with more diverse edit types remains an open empirical question."],"forward_implications":["Exemplar editing becomes an inference-time operation, taking about 140 seconds per image instead of nine-plus minutes, so interactive use and batch processing are practical.","The method is independent of the base diffusion model, so it can be ported to newer text-to-image backbones without retraining an editing module.","The combined text-plus-CLIP-delta conditioning covers both global edits (style, scene) and localized semantic edits (object identity, attributes), while structure preservation keeps non-edited regions stable.","The curated dataset of 170 exemplar pairs with ground-truth edited images offers a standardized evaluation target for future work on exemplar-based editing.","Removing the optimization step eliminates per-example training loops and the hyperparameters that come with them, leaving the edit capture dependent only on the fixed base models."],"supporting_citations":[{"why":"The optimization-based VISII baseline the method must beat; supplies the central comparison and the S-Visual metric.","marker":"[26]"},{"why":"Instruct-Pix2Pix provides a strong text-instruction baseline and contributes examples to the evaluation dataset.","marker":"[3]"},{"why":"Supplies the attention and feature injection mechanism used to preserve the content image's structure during editing.","marker":"[32]"},{"why":"DDIM inversion gives the noise latent and feature maps used in the parallel denoising runs.","marker":"[30]"},{"why":"The multimodal VLM family that verbalizes the exemplar edit into the text caption used as conditioning.","marker":"[20–22]"},{"why":"Imagic contributes exemplar/ground-truth examples to the curated evaluation dataset.","marker":"[16]"},{"why":"HQ-Edit contributes additional examples to the curated evaluation dataset.","marker":"[13]"}],"fun_headline_variants":["Exemplar pair edit: no training, 4x faster than optimization","Edit transfer from before/after pairs using off-the-shelf models","VLMs read exemplar pairs, diffusion applies edit without training","Optimization-free edit transfer from pairs, 4x speedup","Teach edits via pairs, not text: 4x faster, no training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the CLIP-space difference between the two exemplar images is a transferable representation of the edit, so the same conceptual change applied to a different content image shifts CLIP embeddings in roughly the same direction and lets the diffusion model reproduce the edit from that vector alone.","fun_headline_variants_meta":{"raw":{"variants":["Exemplar pair edit: no training, 4x faster than optimization","Edit transfer from before/after pairs using off-the-shelf models","VLMs read exemplar pairs, diffusion applies edit without training","Optimization-free edit transfer from pairs, 4x speedup","Teach edits via pairs, not text: 4x faster, no training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000165,"raw_usage":{"total_tokens":1226,"prompt_tokens":894,"completion_tokens":332,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":238}},"tokens_in":510,"tokens_out":332,"duration_ms":3531,"temperature":1.0,"reasoning_tokens":238,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:55:06.397223+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One decisive experiment is to ablate the image-space branch by setting $\\Delta_{\\mathrm{img}}$ to zero and keeping only the LLaVA text caption; if the edited outputs do not degrade on LPIPS and human preference, the CLIP delta is not doing the claimed work. A complementary check computes $\\Delta_{\\mathrm{img}}$ for the same semantic edit using several different exemplar pairs; if those vectors have low cosine similarity, the transferability assumption behind the method fails.","supporting_citations":[{"cited_title":"Advances in Neural Information Processing Systems36 (2024)","cited_arxiv_id":null,"evidence_quote":"The optimization-based VISII baseline the method must beat; supplies the central comparison and the S-Visual metric."},{"cited_title":"In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"Supplies the attention and feature injection mechanism used to preserve the content image's structure during editing."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"Imagic contributes exemplar/ground-truth examples to the curated evaluation dataset."}],"review_version":1}