{"id":"992cf947-4350-419d-872e-dfc14f4fd216","arxiv_id":"2506.19103","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A cycle-consistency reconstruction loss on the forward consistency model improves 4-step image inversion and editing, closing most of the quality gap to full-step diffusion editing.","lead":"Researchers fine-tune a consistency model's forward pass with a reconstruction loss so that a 4-step inversion and regenerating loop returns the original image more faithfully. This makes fast text-guided image editing more accurate, approaching the quality of slow, full-step diffusion methods at a fraction of the compute.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline claim that the method \"matches or surpasses full-step diffusion models\" rests on a hand-picked inference configuration tuned during experiments on Pie-Bench, with no held-out validation or variance reporting, so the reported superiority is not yet established.","rationale":"I read the paper in good faith. The core idea is plausible: fine-tuning the forward consistency model with a cycle-consistency loss is tractable because the pipeline is only four steps, and Table 3 shows that the proposed fine-tuning does materially improve content preservation over the iCD baseline. However, the headline claim that the method \"matches or surpasses full-step diffusion models\" is not established by the evidence as presented. Table 2 shows Ours trails NTI, InfEdit, and Guide-and-Rescale on LPIPS, and trails Guide-and-Rescale and ReNoise on CLIP, so the conclusion depends on which metric is prioritized. The paper itself states in Section 5.2 that the feature and self-attention guider coefficients and the noise-rescaling bounds were varied during experiments, with no held-out validation split and no error bars. This is the weakest point in the argument: the selected operating point could be an optimistic fit to Pie-Bench rather than a configuration that generalizes. The proposed test—tuning on a validation split and freezing before evaluation—would directly settle whether the reported gains are reproducible under honest model selection. I do not see an internal inconsistency in Equation 4 or the guidance derivation, so the issue is evidentiary rather than logical. The reader's CONDITIONAL verdict remains appropriate; no change is needed.","tokens_in":11068,"tokens_out":8215,"duration_ms":90431,"concrete_test":"Hold out a random subset (e.g., 20%) of Pie-Bench or a second editing benchmark as a test set. Tune all inference hyperparameters (CFG schedule, self-attention guider weight, feature guider weight, r_lower, r_upper) on the remaining 80% using the same search procedure described in the paper, then freeze the selected configuration and evaluate on the held-out set. Repeat with at least three independent fine-tuning runs or editing passes to estimate variance, and report DINOv2, LPIPS, CLIP, and ImageReward. If the held-out numbers no longer beat or match the full-step baselines within variance, the state-of-the-art claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that four-step editing reaches or exceeds full-step diffusion quality. In Section 5.2 the authors state that \"During our experiments, we vary the feature and self-attention guider coefficients, as well as the lower and upper bounds for noise rescaling,\" and the dynamic CFG schedule (0, 7, 11, 19) is also adopted after experimentation. None of these choices is made on a held-out validation split, and no seed-to-seed or run-to-run variance is reported for Table 2. Because the reported operating point (self-attention guider weight 20000, feature guider weight 0.5, r_lower=0, r_upper=1.0, and the CFG schedule) is the one that produces the favorable DINOv2/LPIPS trade-off, the comparison against baselines run at their defaults cannot establish a general efficiency-quality advantage. This is load-bearing for the abstract's \"state-of-the-art\" claim because Table 2 already shows that Ours is not uniformly better than full-step methods on LPIPS or CLIP; the \"match or surpass\" conclusion depends on which metric is emphasized at the selected operating point. The concern is not that the numbers are fabricated, but that the evaluation protocol does not support the claim that the selected configuration generalizes beyond the tuning set.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Inverse-and-Edit, a framework for four-step text-guided image editing built on invertible consistency distillation (iCD). The main idea is to fine-tune only the forward consistency model by backpropagating a cycle-consistency LPIPS reconstruction loss through the full four-step inversion-and-generation pipeline, while freezing the backward model. The authors also adapt the Guide-and-Rescale self-attention and feature guidance mechanism to guidance-distilled consistency models, together with a dynamic CFG schedule and noise rescaling bounds. The paper reports improved reconstruction quality on MS-COCO and improved content-preservation metrics on Pie-Bench relative to several few-step baselines, claiming performance comparable to or better than full-step diffusion methods while being faster.","tokens_in":11277,"tokens_out":3133,"duration_ms":32822,"significance":"If the central claims are supported, the main contribution is useful: the paper demonstrates that the short inference trajectory of distilled consistency models makes full-process cycle-consistency optimization feasible, and that this can improve inversion fidelity and editing quality without retraining the backward model. The method is clearly described, the code is promised to be public, and the ablation in Table 3 at least supports the qualitative direction that fine-tuning plus guidance changes the editability/preservation trade-off. The practical value is that the approach requires only four sampling steps and does not rely on blend words or Prompt-to-Prompt post-processing. However, the strength of the headline claims is not yet commensurate with the evidence: the evaluation protocol mixes multiple metrics without an explicit decision rule, the hyperparameters for the final operating point are described as chosen during experiments rather than on a held-out validation set, and some reported comparisons contradict the abstract's state-of-the-art wording.","major_comments":[{"comment":"The abstract and introduction claim that the method 'matches or surpasses full-step diffusion models' and achieves 'state-of-the-art performance', but Table 2 does not support a uniform superiority claim. For example, TurboEdit achieves ImageReward 0.536 versus 0.279 for Ours, and InfEdit achieves DINOv2 0.781 and LPIPS 0.236 versus 0.747 and 0.296 for Ours. The conclusion that the method is 'comparable' to full-step methods may be defensible, but the conclusion that it outperforms other accelerated models is contradicted by at least one baseline on one of the four reported metrics. The authors should either restrict the claim to specific metrics or datasets, or provide a statistically grounded multi-metric comparison with error bars and a stated aggregation rule.","section":"§5.2, Table 2"},{"comment":"The final inference configuration is selected from choices that the authors explicitly report varying during experiments: the dynamic CFG schedule (0, 7, 11, 19), self-attention guider weight 20000, feature guider weight 0.5, and noise rescale bounds r_lower=0 and r_upper=1.0. No held-out validation split is used to select these values, and no ablation or sensitivity analysis is reported for them. Since all baselines are run at their default settings, the comparison in Table 2 risks reflecting favorable tuning of the proposed method rather than a general efficiency-quality advantage. The authors should report the selection procedure, the range of values explored, and the performance variation over that range, or validate the chosen configuration on an independent split of Pie-Bench.","section":"§5.2 and Appendix A.2"},{"comment":"The cycle-consistency loss in Eq. (4) directly optimizes LPIPS, and the reconstruction evaluation in Table 1 uses LPIPS as a headline metric, on the same domain used for fine-tuning (COCO train versus COCO validation). The reported LPIPS improvement over iCD is therefore partly by construction and does not by itself establish that inversion quality is generally better. The authors should complement the LPIPS comparison with held-out, non-latent-aware distortion measures (e.g., human preference ratings, or metrics not aligned with the training loss) and should report the behavior of MSE and ImageReward more carefully, since those metrics do not show a uniform improvement over iCD or DDIM.","section":"§4.1, Eq. (4); §5.1, Table 1"},{"comment":"Table 3 shows that adding guidance to the proposed method improves DINOv2 and LPIPS but degrades ImageReward (0.313 to 0.279) and CLIPScore (0.304 to 0.302). The text states that guidance improves content preservation and calls this 'key to better visual quality', but the reported metrics move in opposite directions, so the claimed benefit is metric-selective. The authors should either present a user study or a clear preference-based evaluation to support the claim that the guided operating point is visually better, or explicitly frame the contribution as a trade-off controller rather than an unambiguous improvement.","section":"§5.3, Table 3"}],"minor_comments":[{"comment":"The sentence 'The consistency distillation loss from Equation (4) is adapted for both the fCM and CM training objectives' appears to refer to Eq. (3), since Eq. (4) is the reconstruction loss introduced later; please update the cross-reference.","section":"§3, 'Inversion in consistency models'"},{"comment":"The phrase 'for fCD and the forward preservation loss' appears to contain a typo; it should likely read 'fCM'.","section":"Appendix A.1"},{"comment":"Several rows in Table 1 are typeset without separators between the metric value and the model name (e.g., '0.424iCD (4 steps)', '0.372ReNoise SDXL-Turbo (4 steps)', '0.444Ours (4 steps)'), making the table difficult to read.","section":"Table 1"},{"comment":"The statement that 'most of the error attributed to approximation mismatches between adjacent timesteps' is presented without supporting evidence; a quantitative attribution or at least a diagnostic experiment would be needed to justify this claim.","section":"§5.1"},{"comment":"The paper does not state the number of seeds or runs used to produce Table 2, and with 420 Pie-Bench images some reported metric differences may be within noise; please add a statement on variance or provide per-run results.","section":"§5.2"},{"comment":"The Null-text Inversion reference is duplicated in the bibliography (Mokady et al., 2022 and Mokady et al., 2023); the duplicate entry should be removed or cross-referenced properly.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The core technical idea is reasonable and the ablation suggests a real effect, but the paper's evaluation protocol is not yet strong enough for the claimed 'state-of-the-art' and 'matches or surpasses full-step models' statements. The main risks are selective reporting of metrics and tuning of inference hyperparameters without a validation split. These are fixable with additional experiments and a more disciplined evaluation narrative, so I do not recommend rejection, but the revision needs to be substantial."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a decent, useful paper. The central idea is simple—fine-tune only the forward consistency model so that noising followed by denoising reconstructs the input image, keeping the backward model frozen—and it works in their ablations. The gains in reconstruction and content preservation over the iCD baseline are clear, and the adaptation of Guide-and-Rescale to consistency models is a reasonable port. The code is public, which earns credit.\n\nWhat's new: cycle-consistency fine-tuning of a forward CM for inversion is not in the prior work they cite. The contribution is incremental but not trivial: it makes 4-step editing much more usable.\n\nWhere it gets soft: the abstract says 'state-of-the-art' and 'matches or surpasses full-step diffusion models,' but Table 2 doesn't support that. TurboEdit beats them on ImageReward by a wide margin, and they don't uniformly win on LPIPS or CLIP either. The conclusion depends on which metric you emphasize. The bigger problem is the evaluation protocol. The CFG schedule (0, 7, 11, 19), guider weights (20000, 0.5), and rescale bounds are chosen during experiments on Pie-Bench, with no held-out validation split and no variance reporting. That means the comparison is against baselines at their defaults, at an operating point selected on the test set. That's a real threat to the generalization claim, and it's load-bearing for 'state-of-the-art.' Also, the reconstruction loss is LPIPS, and LPIPS is a headline metric, evaluated on COCO val after training on COCO train—so part of that improvement is by construction. The editing gains on Pie-Bench are not directly optimizing the editing metrics, so the circularity is partial, not fatal.\n\nNone of this is a dealbreaker. The method is plausible, the ablations support the core effect, and the limitations section is honest. But the paper needs a held-out validation procedure for the hyperparameters, error bars or seed variance, and a more careful claim. As is, I'd accept it for review, but I'd expect revision.\n\nWho it's for: people working on efficient image editing with distilled diffusion models. I'd bring it to a reading group as a useful case study in how evaluation choices shape claims. I wouldn't cite it as a definitive result until the evaluation is tightened. Send it to a serious referee.","headline":"The cycle-consistency fine-tuning is a real improvement over iCD, but the 'state-of-the-art' claim and hand-tuned evaluation protocol need substantial revision before the results are trustworthy.","tokens_in":11870,"tokens_out":2699,"would_cite":false,"duration_ms":27822,"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":"A cycle-consistency loss lets four-step consistency models edit images as well as 50-step diffusion models.","keywords":["image editing","consistency models","diffusion inversion","cycle-consistency loss","distilled diffusion models","zero-shot editing","self-guidance","LPIPS"],"falsifier":"Run the released method on a new, out-of-distribution set of 300 images and prompts using only the hyperparameters printed in the paper; if reconstruction LPIPS or DINOv2 similarity no longer beats the iCD baseline by the reported margin, the cycle-consistency gain is configuration-dependent rather than a property of the loss itself.","tokens_in":10802,"feed_emoji":"🖼️","tokens_out":4661,"duration_ms":40908,"temperature":0.7,"pith_summary":"The paper tries to close the gap between fast and full-step diffusion editing. It argues that the bottleneck is inversion quality, not generation speed: distilled four-step models reconstruct source images poorly, so edits drift. The fix is a cycle-consistency fine-tuning of the forward consistency model, backpropagating a perceptual loss through the full noising-then-denoising loop. With that inversion improvement plus a self-guidance mechanism, the method claims editing results that match or surpass full-step diffusion models while using only four steps.","feed_headline":"Four-step image editing now rivals full diffusion quality","feed_subtitle":"A cycle-consistency loss sharpens inversion in consistency models, cutting compute while matching 50-step editors.","key_machinery":"The central mechanism is the cycle-consistency loss, a perceptual reconstruction objective evaluated after the complete inversion-and-generation cycle: encode the image, noise it through four forward consistency-model steps, denoise it through the frozen backward model, decode, and compare with LPIPS against the original. Because the loop is only eight model calls, end-to-end backpropagation is feasible, so the forward model is directly optimized for the reconstruction task, while consistency-distillation and forward-preservation losses keep its trajectory aligned with the original model.","core_discovery":"On its own terms, the paper claims that fine-tuning the forward consistency model with the cycle-consistency loss $L_{\\mathrm{rec}} = \\mathrm{LPIPS}(G_{\\theta_+}(F_{\\theta_-}(x_0)), x_0)$ substantially improves image inversion and content preservation, measured by LPIPS, MSE, and ImageReward. Combined with a dynamic classifier-free guidance schedule and noise-rescaled self-attention and feature guidance, the resulting four-step editing pipeline matches the editing quality of full-step diffusion baselines such as Guide-and-Rescale, while outperforming other few-step methods, without needing blend words or Prompt-to-Prompt.","pith_inferences":["The fine-tuning is done on MS-COCO only; a natural test is whether the inversion gain persists on out-of-domain images (e.g., paintings, medical scans, faces), where both the VAE and LPIPS behave differently.","A similar cycle-consistency objective could be applied to other distilled backbones (e.g., SDXL-Turbo or latent consistency models), where inversion quality is currently a bottleneck; the paper does not test this.","Because the guidance weights and CFG schedule are varied during experiments, a fair cross-method comparison would require fixing them on a validation set disjoint from the test set; the paper does not describe such a split.","The reported DINOv2-vs-LPIPS trade-off suggests the method favors semantic structure over pixel-level similarity; an interactive user study on target edits could confirm whether that bias matches human preference."],"forward_implications":["Four-step editing becomes a practical alternative to 50-step diffusion editing, cutting compute by roughly an order of magnitude without a visible quality drop.","Inversion quality in distilled consistency models can be improved without retraining the generation (backward) model, so existing fast-generation checkpoints can be reused.","Editing no longer depends on attention-manipulation tricks or hand-picked blend words; switching the prompt after a high-fidelity inversion is enough for many edits.","The trade-off between editability and content preservation can be controlled explicitly through guidance strength and noise-rescale bounds.","If the cycle-consistency gains transfer, other four-step editing methods built on the same distilled backbone could adopt the same fine-tuning and improve immediately."],"supporting_citations":[{"why":"Supplies the invertible consistency distillation baseline and the forward/backward consistency-model architecture that the method fine-tunes and extends.","marker":"[Starodubcev et al., 2024]"},{"why":"Provides the consistency-distillation objective and self-consistency property that underpin the frozen backward and fine-tuned forward models.","marker":"[Song et al., 2023]"},{"why":"Source of the Guide-and-Rescale self-guidance and noise-rescaling mechanism that the paper adapts to consistency models.","marker":"[Titov et al., 2024]"},{"why":"Defines the perceptual LPIPS loss used as the cycle-consistency reconstruction objective.","marker":"[Zhang et al., 2018]"},{"why":"Supplies the multi-step consistency sampling procedure used in the generation phase.","marker":"[Heek et al., 2024]"},{"why":"Provides the MS-COCO training and validation images for fine-tuning and quantitative reconstruction evaluation.","marker":"[Lin et al., 2015]"},{"why":"Contributes the Pie-Bench dataset used for editing and qualitative inversion evaluation, following the iCD protocol.","marker":"[Ju et al., 2023]"},{"why":"Defines Prompt-to-Prompt, the baseline technique the method shows it no longer needs.","marker":"[Hertz et al., 2022]"}],"fun_headline_variants":["4-step editing matches full diffusion quality","Cycle consistency enables 4-step image editing","4-step editing rivals full-step diffusion","Consistency models cut editing to four steps","Cycle-consistent inversion delivers 4-step editing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported editing advantage relies on a hand-picked inference configuration (dynamic CFG schedule, guider weights, and noise-rescale bounds) that the paper says is varied during experiments, with no held-out validation described; if those settings are tuned to the evaluation benchmark, the gains may not generalize to new images and prompts.","fun_headline_variants_meta":{"raw":{"variants":["4-step editing matches full diffusion quality","Cycle consistency enables 4-step image editing","4-step editing rivals full-step diffusion","Consistency models cut editing to four steps","Cycle-consistent inversion delivers 4-step editing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001191,"raw_usage":{"total_tokens":4860,"prompt_tokens":840,"completion_tokens":4020,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":456,"completion_tokens_details":{"reasoning_tokens":3955}},"tokens_in":456,"tokens_out":4020,"duration_ms":30788,"temperature":1.0,"reasoning_tokens":3955,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:36:42.940676+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released method on a new, out-of-distribution set of 300 images and prompts using only the hyperparameters printed in the paper; if reconstruction LPIPS or DINOv2 similarity no longer beats the iCD baseline by the reported margin, the cycle-consistency gain is configuration-dependent rather than a property of the loss itself.","supporting_citations":[],"review_version":2}