{"id":"62d59a1e-08be-42e9-99b9-1e22e856c64c","arxiv_id":"2505.09571","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"SAGE edits real images with text by using self-attention maps from inverse DDIM as guidance, which preserves unedited regions without per-image optimization.","lead":"SAGE is a new image-editing method that edits real photos from a text prompt using a pre-trained diffusion model, without per-image optimization. It keeps unchanged areas intact by comparing internal attention maps from the inverse DDIM process, making editing faster and lighter than current optimization-based approaches.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Self-attention guidance may not be the load-bearing fidelity mechanism: Supp. Table VI shows SAGE reconstruction alone is far worse than DI/NT, while the final pipeline adds local blending (Eq. 5) that directly copies original latents.","rationale":"The reader's weakest assumption focuses on the unverified choice of attention layers and scales for the self-attention guidance. My concern is more direct and is supported by the paper's own supplementary results: the self-attention guidance term, in isolation, does not achieve high-fidelity reconstruction (Supp. Table VI), while the full SAGE pipeline includes local blending that explicitly copies inversion latents outside the edit mask. This makes the central attribution of SAGE's fidelity gains to self-attention guidance uncertain. The concern is not that the method fails, but that the paper has not isolated which component carries the reconstruction load. The internal inconsistency about whether S_out is taken from the Pout branch (Sec. III-A) or the Pin branch (Fig. 3 caption, Conclusion) further complicates interpretation: the two choices imply different mechanisms, and a reader cannot tell which one is actually implemented. This does not require rejecting the paper; the main evaluation could still support SAGE as a practical editing method. However, the central scientific claim should be conditionally interpreted until a clean ablation and code inspection resolve the attribution. I therefore keep the reader's conditional verdict, noting that the required conditions should include this ablation and the branch-consistency check.","tokens_in":13109,"tokens_out":6339,"duration_ms":61617,"concrete_test":"Using the released code, run the same PieBench background-preservation and CLIP-T evaluation in four variants on the same seed set: (a) full SAGE; (b) full SAGE with self-attention guidance disabled (λ=0, keeping CFG, cross-attention replacement, and local blending); (c) full SAGE without local blending but with self-attention guidance; (d) self-attention guidance only, as in Supp. Table VI. If (b) matches (a) on background LPIPS/SSIM, the fidelity gain comes from local blending, not self-attention guidance; if (c) degrades badly, self-attention guidance is necessary but not sufficient. Separately, inspect the released code to determine whether Eq. 3 compares against maps from the Pout branch or the Pin branch, resolving the Sec. III-A versus Fig. 3/Conclusion contradiction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that self-attention maps from DDIM inversion are sufficient to guide high-fidelity reconstruction during editing (Sec. III-A, Eqs. 3-4). The paper's own reconstruction-only experiment (Supp. Sec. VII-C, Table VI) directly tests this claim: with Pin=Pout and no local blending/CA replacement, SAGE achieves structure distance 12.0, LPIPS 65.8, and MSE 65.1, versus Direct Inversion's 3.0, 51.7, and 28.9, and Null-Text Inversion's 3.3, 54.8, and 31.1. Thus self-attention guidance alone reconstructs substantially worse than the optimization-based baselines it is claimed to replace. Yet the headline results in Tables II-IV come from the full pipeline, which includes Local Blending (Eq. 5) that replaces z_{t-1} with the inversion latent z_in_{t-1} outside the cross-attention mask. That is explicit reconstruction in unedited regions and can explain the background-preservation gains without invoking self-attention guidance. The paper also contradicts itself about which branch supplies S_out: Sec. III-A says ε_θ(z_t,Pout), while Fig. 3's caption and the Conclusion state that guidance is applied to the Pin branch. Which branch is used changes the mechanism entirely. Until these are resolved, the advertised attribution of SAGE's fidelity to self-attention guidance is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SAGE, a prompt-based image editing method that combines DDIM inversion with a self-attention guidance loss and cross-attention manipulation (local blending, cross-attention replacement, reweighting) to preserve unedited regions while applying text-driven edits. The central claim is that the self-attention maps recorded during DDIM inversion contain sufficient information to steer generation toward the input without per-image optimization, yielding high-fidelity editing with lower memory and compute than optimization-based baselines. The authors provide quantitative benchmarks on PieBench and MagicBrush, a user study, ablation studies, and public code.","tokens_in":13433,"tokens_out":2642,"duration_ms":26864,"significance":"If the central claim is established, SAGE would be a valuable practical contribution: it avoids per-image text-embedding optimization, runs in comparable time to Direct Inversion, and uses less memory in FP16. The paper ships code, provides ablations, and evaluates on two benchmark datasets, which are clear strengths. However, the current evidence does not yet isolate the effect of self-attention guidance from the local-blending component, and the manuscript contains internal inconsistencies about which CFG branch is guided. These issues are load-bearing for the claimed attribution of SAGE's fidelity to self-attention guidance, so the contribution, while promising, needs strengthening before publication.","major_comments":[{"comment":"There is a direct contradiction about which branch provides the self-attention maps S_out in Eq. (3). §III-A states that \"during synthesis, corresponding maps S_out_i,t are collected from ε_θ(z_t, P_out)\", whereas the caption of Fig. 3 and the Conclusion state that guidance is applied to the branch conditioned on the initial prompt P_in. The mechanism changes entirely depending on which branch is used: guiding toward P_in inversion maps is a reconstruction mechanism, while guiding toward P_out maps would push the edited image away from the input. The authors must clarify the exact branch and justify it, because the method's definition and the interpretation of all ablations depend on this choice.","section":"§III-A, Fig. 3, §V"},{"comment":"The reconstruction-only experiment in Supp. Table VI tests the load-bearing claim that self-attention guidance alone preserves the input. With P_in = P_out and no local blending or cross-attention replacement, SAGE achieves structure distance 12.0, LPIPS 65.8, and MSE 65.1, markedly worse than Direct Inversion (3.0, 51.7, 28.9) and Null-Text Inversion (3.3, 54.8, 31.1). Yet the headline results in Tables II and IV come from the full pipeline, which includes local blending (Eq. (5)) that directly replaces z_{t-1} with the inversion latent z_in_{t-1} outside the edit mask. That direct latent copying is an explicit reconstruction mechanism that can explain the background-preservation gains independently of self-attention guidance. To support the paper's attribution, the authors should add an ablation that runs local blending without self-attention guidance and self-attention guidance without local blending on the same benchmark, and report both structure and background metrics.","section":"Supp. §VII-C, Table VI; §III-B, Eq. (5)"},{"comment":"The abstract's claim that \"all 47 surveyed users preferred SAGE over competing methods\" is contradicted by Table IV. The user study recruited 22 participants for the PieBench comparisons and 25 additional participants for the MagicBrush DI comparison (total 47), but SAGE's global preference over DI is 52.7% on both, and its global preference over NT is 54.5%. Many of the 47 users did not prefer SAGE in those comparisons. The same internal contradiction appears between the abstract and the Introduction's \"preferred in over 60% of cases.\" The claims in the abstract and Sec. IV-D must be corrected to report preference rates per comparison, not a universal-preference statement.","section":"Abstract; §IV-D, Table IV"}],"minor_comments":[{"comment":"Typo: \"An appropiate balance\" should be \"An appropriate balance\".","section":"Supp. §VII-B"},{"comment":"The dataset name is written inconsistently as \"PIE-Bench\" in the supplementary material and \"PieBench\" elsewhere; please unify.","section":"Supp. §VI"},{"comment":"The footnote states \"FP32 best results are marked with *\", but in the table the asterisks mark Plug-n-Play time and Direct Inversion memory, which are not clearly the best in their columns (e.g., SAGE FP16 ties at 7.4 GB memory). Please clarify what the asterisk denotes.","section":"Table V"},{"comment":"The summation index N and the subscript i are used without defining the set of blocks or the spatial resolution of the maps; please specify that the sum ranges over selected U-Net blocks and state the map sizes (e.g., 32x32 from blocks 2 and 3, as in Supp. VI-A) in the main text.","section":"Eq. (3)"},{"comment":"The phrase \"'a cat tiger sitting next to a mirror\" contains an extra leading apostrophe; please fix the typo in the prompt string.","section":"Fig. 9 caption"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the engineering effort is evident, but the attribution of the method's success to self-attention guidance is not currently established because the full pipeline includes local blending, and the guidance branch is defined inconsistently. The reconstruction-only result in Supp. Table VI makes this concern concrete. I would ask for a clarifying revision plus the suggested ablation before reconsidering; the user-study claim in the abstract should also be corrected as a matter of factual accuracy."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"New here: a reconstruction guidance loss applied to self-attention maps recorded during DDIM inversion. The idea is simple and the paper's own ablations give it some support—among the guidance variants tested (cross-attention, z_t, self-attention), self-attention guidance has the best LPIPS. The efficiency story also checks out: 7.4 GB memory and about 12.6 s per 512×512 image, roughly on par with Direct Inversion in FP16.\n\nThe central claim, however, is not supported by the paper's own data. Supp. Table VI, the reconstruction-only experiment with P_in=P_out and no local blending, shows SAGE at structure distance 12.0 and LPIPS 65.8, versus Direct Inversion's 3.0/51.7 and Null-Text's 3.3/54.8. The full pipeline then adds Local Blending, which directly copies original latents into the unedited region. That alone can explain most of the background-preservation wins. The paper never isolates how much self-attention guidance contributes once Local Blending is in the loop—Table I only shows the cumulative pipeline. So the advertised attribution of fidelity to self-attention guidance is not established.\n\nThere is also a genuine internal contradiction: Section III-A describes S_out as coming from ε_θ(z_t, P_out), while Fig. 3's caption and the Conclusion state the guidance is applied to the P_in branch. That changes the mechanism entirely, and it must be resolved.\n\nThe abstract's 'all 47 surveyed users preferred SAGE' is contradicted by Table IV, where SAGE vs DI is 52.7/52.7 on global preference. The body is more careful (average preference 60.7%), but the abstract is false as written.\n\nFixable issues: no error bars in the quantitative tables; hyperparameters were selected by a 'preliminary analysis' without a disclosed validation protocol; MasaCtrl, an earlier self-attention-based editing method, is missing from the references and comparisons; and the qualitative comparison tunes baselines to their best while SAGE uses a fixed set.\n\nNone of this kills the underlying idea. The method is simple, cheap, and the ablation study is a reasonable attempt to compare guidance variants. But the contribution is oversold and the mechanism under-tested. I'd send this to a careful referee with a request for major revision, not a desk reject. The authors discuss limitations honestly, promise code and data, and the failure modes are documented. That is work worth engaging with.","headline":"A simple, efficient self-attention guidance idea that is unfortunately oversold by an unsupported central claim and a false abstract claim.","tokens_in":13971,"tokens_out":5240,"would_cite":true,"duration_ms":45477,"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":"Inversion-time self-attention maps alone can guide high-fidelity text edits in pre-trained diffusion models.","keywords":["diffusion models","image editing","DDIM inversion","self-attention guidance","classifier-free guidance","cross-attention control","prompt-based editing","attention maps"],"falsifier":"A direct test is to run SAGE twice on the same image and prompt, once with the real inversion-time self-attention maps and once with maps recorded from a different image's inversion; if structure and background preservation scores (LPIPS, structure distance) stay the same, the maps are not the carrier of the reconstruction signal and the method's explanation is wrong.","tokens_in":12919,"feed_emoji":"🎨","tokens_out":6447,"duration_ms":61040,"temperature":0.7,"pith_summary":"This paper introduces SAGE, a method for prompt-based image editing that uses a pre-trained text-to-image diffusion model without any per-image optimization or fine-tuning. The method's central claim is that deterministic DDIM inversion already records enough information to reconstruct the unedited parts of an image: the self-attention maps produced during inversion serve as fixed targets that pull generation back toward the input while the target prompt drives the edit. SAGE stores those maps, then at each sampling step adds a gradient term that minimizes the L1 distance between the generation-time self-attention maps and the stored inversion-time maps. This gives high-fidelity background and structure preservation with lower memory use than optimization-based baselines, and the paper reports top or near-top scores across quantitative benchmarks and a user study.","feed_headline":"Inversion attention maps alone guide faithful image edits","feed_subtitle":"SAGE stores the U-Net's self-attention maps during DDIM inversion, preserving unedited regions without per-image optimization.","key_machinery":"The central object is the collection of self-attention maps the diffusion U-Net computes during DDIM inversion; for 512x512 images the paper uses the 32x32 maps from the second and third encoder blocks, and 16x16 cross-attention maps for blending. These stored maps are the reference signal: a reconstruction loss compares them with the maps produced during DDIM sampling under the target prompt, and its gradient steers the latent trajectory back to the input's structure. The mechanism works because self-attention captures global interactions among all image tokens, so matching these maps preserves layout and identity without requiring the latent itself to match the input. A decreasing schedule for $\\lambda$ lets early steps perform the edit and later steps refine reconstruction.","core_discovery":"The discovery is that reconstruction guidance for editing does not need to operate in the pixel or latent space, nor require optimizing a null embedding: the self-attention maps from a single deterministic DDIM inversion are a sufficient and stable target. During generation, SAGE applies classifier-free guidance between the input prompt and the target prompt, and simultaneously minimizes $L_{\\text{self}}^t = \\sum_i \\|S_{\\text{in}, t,i} - S_{\\text{out}, t,i}\\|_1$, where $S_{\\text{in},t,i}$ are the self-attention maps recorded during inversion and $S_{\\text{out},t,i}$ are the maps from the U-Net when conditioned on the target prompt at the same timestep. The gradient, scaled by a decreasing weight $\\lambda$, is added to the noise update. The paper argues this stabilizes the reverse process and preserves unedited regions because self-attention encodes global spatial relationships, and it combines this with cross-attention local blending and replacement to refine where edits occur.","pith_inferences":["Editorial inference: because the reference maps come from an arbitrary inversion, the same guidance could be applied on top of exact or coupled inversion schemes by simply replacing the source of the recorded maps; the mechanism itself does not depend on DDIM's update rule.","Editorial inference: the reported failure mode in object removal, where removed objects are replaced by structurally similar content, suggests that masking self-attention guidance in the removed region is a concrete next test; the paper itself flags this direction.","Editorial inference: the fixed layer selection (32x32 maps from encoder blocks 2 and 3) is an empirical choice, so a systematic sweep over layers and resolutions could either harden the method or reveal that only a subset carries the reconstruction signal."],"forward_implications":["SAGE completes an edit in about 12.6 seconds on an A100 with 7.4 GB peak memory, versus over a minute for null-text inversion, while keeping reconstruction metrics competitive.","Because reconstruction is anchored in attention-map space, the method transfers across PieBench and MagicBrush with the same hyperparameter set.","Prompt-based operations such as word swaps, object removal, and style transfer all reduce to the same pipeline: store inversion maps, then apply self-attention guidance plus cross-attention local blending.","The ablation shows that guidance from self-attention maps outperforms guidance from cross-attention maps and from latent-space ($z_t$) targeting on structure distance and LPIPS."],"supporting_citations":[{"why":"Supplies the deterministic DDIM inversion that produces the recorded noise latents and self-attention maps.","marker":"[7]"},{"why":"Defines the prompt-based editing task and provides the null-text inversion baseline that SAGE avoids.","marker":"[3]"},{"why":"Provides the negative-prompt CFG strategy and the NPI baseline compared in experiments.","marker":"[4]"},{"why":"Provides the P2P-Zero baseline, which also uses guidance for reconstruction and is the comparison point for the attention-guidance ablation.","marker":"[5]"},{"why":"Supplies the classifier-free guidance formulation that SAGE combines with self-attention guidance.","marker":"[6]"},{"why":"Provides the Plug-and-Play baseline for feature-based editing and is compared on structure and fidelity metrics.","marker":"[14]"},{"why":"Provides the Direct Inversion baseline, the PieBench benchmark, and the comparison images used in quantitative and user studies.","marker":"[15]"},{"why":"Supplies the cross-attention manipulation mechanisms (local blending, replacement, reweighting) that SAGE adapts.","marker":"[16]"},{"why":"Provides the proximal guidance baseline (ProxNPI) used in quantitative comparison and the user study.","marker":"[23]"},{"why":"Supplies the decreasing-guidance schedule for $\\lambda$ that SAGE adopts for balancing early editing and late reconstruction.","marker":"[24]"}],"fun_headline_variants":["Using only inversion attention maps for editing","Self-attention from DDIM inversion powers edits","Attention guidance beats null-text optimization","SAGE edits with inversion self-attention only","Store attention maps, not null embeddings"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing empirical premise is that the self-attention maps recorded during DDIM inversion remain the correct targets while the model generates under the edited prompt, and that minimizing their L1 distance pulls the latent toward the input without suppressing the intended edit.","fun_headline_variants_meta":{"raw":{"variants":["Using only inversion attention maps for editing","Self-attention from DDIM inversion powers edits","Attention guidance beats null-text optimization","SAGE edits with inversion self-attention only","Store attention maps, not null embeddings"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000234,"raw_usage":{"total_tokens":1492,"prompt_tokens":939,"completion_tokens":553,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":489}},"tokens_in":555,"tokens_out":553,"duration_ms":5449,"temperature":1.0,"reasoning_tokens":489,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:28:25.685568+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test is to run SAGE twice on the same image and prompt, once with the real inversion-time self-attention maps and once with maps recorded from a different image's inversion; if structure and background preservation scores (LPIPS, structure distance) stay the same, the maps are not the carrier of the reconstruction signal and the method's explanation is wrong.","supporting_citations":[{"cited_title":"Denoising diffusion implicit models,","cited_arxiv_id":null,"evidence_quote":"Supplies the deterministic DDIM inversion that produces the recorded noise latents and self-attention maps."},{"cited_title":"Null- text inversion for editing real images using guided diffusion models,","cited_arxiv_id":null,"evidence_quote":"Defines the prompt-based editing task and provides the null-text inversion baseline that SAGE avoids."},{"cited_title":"Zero-shot image-to-image translation,","cited_arxiv_id":null,"evidence_quote":"Provides the P2P-Zero baseline, which also uses guidance for reconstruction and is the comparison point for the attention-guidance ablation."},{"cited_title":"Classifier-free diffusion guidance,","cited_arxiv_id":null,"evidence_quote":"Supplies the classifier-free guidance formulation that SAGE combines with self-attention guidance."},{"cited_title":"Plug-and-play diffusion features for text-driven image-to-image translation,","cited_arxiv_id":null,"evidence_quote":"Provides the Plug-and-Play baseline for feature-based editing and is compared on structure and fidelity metrics."},{"cited_title":"Prompt-to-prompt image editing with cross-attention control,","cited_arxiv_id":null,"evidence_quote":"Supplies the cross-attention manipulation mechanisms (local blending, replacement, reweighting) that SAGE adapts."},{"cited_title":"Improving tuning-free real image editing with proximal guidance,","cited_arxiv_id":null,"evidence_quote":"Provides the proximal guidance baseline (ProxNPI) used in quantitative comparison and the user study."},{"cited_title":"Zero- shot spatial layout conditioning for text-to-image diffusion models,","cited_arxiv_id":null,"evidence_quote":"Supplies the decreasing-guidance schedule for $\\lambda$ that SAGE adopts for balancing early editing and late reconstruction."}],"review_version":1}