{"id":"4ee4ef6f-ab93-4861-be02-9d8524d13d9a","arxiv_id":"2512.01382","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Exemplar-guided editing is reformulated as forward-only two-stage generation from noise—source-guided first, reference-guided second—with mask-based background preservation, reporting SOTA quality and fewer NFEs.","lead":"This paper introduces a training-free way to edit images using an example picture: instead of inverting the source into noise and then generating, it starts from random noise, denoises briefly with the source image as guide, then switches to the reference photo as guide. A mask option keeps the background unchanged, and tests on an image-editing benchmark report higher quality and lower compute than several inversion-based baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 11's substitution of a fresh Gaussian for the inverted noise is not justified by Eq. 8; ReInversion's theoretical equivalence claim lacks support.","rationale":"The reader's weakest_assumption correctly identifies the unsupported substitution in Eq. 11. My analysis confirms that Eq. 8 only relates X̃0 to the specific X0 used in the forward reconstruction, not to any independent Gaussian. Thus the theoretical justification for ReInversion's first stage is incomplete. However, the method is empirically validated on COCOEE (Table 1 and Table 3), and the reformulation may still hold as an empirical property of Flux-Kontext. The reader's CONDITIONAL verdict is appropriate: the method is plausible and likely works, but the central theoretical claim needs stronger evidence. I therefore recommend no change to the verdict. The concrete test I propose would directly evaluate the distributional equivalence of transition states, which is the crux of the concern. I agree with the reader's identification of the same weakest assumption.","tokens_in":14891,"tokens_out":3585,"duration_ms":36384,"concrete_test":"For a fixed set of at least 500 source images, compute the transition state from Recon-Inv: T_inv = X_s − Σ_{i=τ}^{n−1}(t_{i+1}−t_i) v_θ(X̂_ti, ti; X_s) (i.e., the actual inverted state at tτ). For each source, sample multiple fresh Gaussian noises X0' and compute the transition state from ReInversion's first stage: T_fresh = X0' + Σ_{i=0}^{τ−1}(t_{i+1}−t_i) v_θ(X_ti, ti; X_s), where the trajectory is generated from X0'. Compare the distributions of T_inv and T_fresh using FID or MMD on latent states. If the distributions are significantly different (e.g., FID > 5), the reformulation is not equivalent; if they are close, the concern is resolved. Additionally, run both pipelines to completion with a fixed second stage and compare the final edited outputs across seeds for statistical equivalence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim is that ReInversion is equivalent to Recon-Inv under negligible reconstruction error (Sec. 4.2). This rests on Eq. 8, which bounds ||X̃0 − X0|| = ||X_s − X̂1||, where X0 is the specific noise used in the forward reconstruction that produced X̂1. Eq. 11, however, replaces X̃0 with an independently sampled Gaussian X0'. The distance ||X0' − X̃0|| is not controlled by reconstruction error; it is a fresh sample from the prior. Therefore, the reformulated first stage does not follow from Eq. 8. It instead assumes that denoising from any fresh noise under X_s conditioning for t∈[0,tτ] yields a transition state statistically equivalent to the inverted one. This is an empirical property of the backbone, not a consequence of flow-matching theory. Table 3 provides only indirect validation—a single CLIP-FG comparison on COCOEE†, without error bars or a distributional comparison of transition states. The paper's own Supp. C concedes that the training-free approach 'may encounter reduced realism when handling complex semantic transformations,' but this does not address the equivalence gap. The load-bearing assumption is unverified, and the central claim that the two-stage process 'can be safely reformulated' is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ReInversion, a training-free exemplar-guided image editing method built on the Flux-Kontext flow-matching backbone. The method first defines Reconstruction-Based Inversion (Recon-Inv), which uses the velocity fields extracted from a forward reconstruction of the source image to invert the source with an error that is exactly equal to the reconstruction error (Eq. 8). It then reformulates Recon-Inv into a two-stage forward denoising process: a first stage conditioned on the source image, followed by a second stage conditioned on the reference image. A Mask-Guided Selective Denoising (MSD) module further constrains edits to foreground regions. The authors report state-of-the-art FID, CLIP-FG, CLIP-BG, and low NFE counts on a curated COCOEE subset, and provide ablations on the transition timestep and MSD coefficient.","tokens_in":15244,"tokens_out":5019,"duration_ms":55142,"significance":"If the reformulation in Eq. 11 were rigorously justified, this would be a valuable contribution: ReInversion achieves 14–18 NFEs with strong FID (4.90–5.01) and CLIP scores, substantially outperforming prior flow-inversion methods on the reported benchmark, and it is training-free. The identity in Eq. 8 is a clean, non-circular derivation linking inversion error to reconstruction error. The MSD strategy is simple and appears effective in the ablations. However, the central theoretical claim that the two-stage process 'can be safely reformulated' from Recon-Inv rests on an unproven distributional assumption about fresh Gaussian noise paths. The empirical evidence in Table 3 is a single aggregate metric without error bars or a distributional comparison of transition states. Thus the method may be practically useful, but the paper as written overstates the strength of its theoretical support.","major_comments":[{"comment":"The transition from Eq. (10) to Eq. (11) is not a logical consequence of Eq. (8). Eq. (8) bounds ‖X̃₀ − X₀‖ for the specific noise X₀ used in the forward reconstruction that produced X̂₁. However, the final algorithm (Supp. Listing J) initializes X = torch.randn_like(X_src), i.e., a fresh prior sample. The distance between that fresh sample and X̃₀ is not controlled by the reconstruction error; it is of order the prior scale. Additionally, Eq. (11) changes the velocity arguments from X̂_tᵢ (the reconstruction states) to X̃_tᵢ (the new trajectory states), which is a further approximation. The paper therefore assumes, without proof, that denoising any fresh Gaussian under X_s-conditioning for t∈[0,t_τ] lands at a transition state statistically equivalent to the one from Recon-Inv. This is an empirical property of the backbone, not a consequence of flow-matching theory. Please either provid","section":"Sec. 4.2, Eqs. (10)–(11) and Supp. Listing J"},{"comment":"The only quantitative support for the reformulation is a single CLIP-FG value (82.10 vs. 81.80) on one backbone and one benchmark, with no error bars or seed variance. This is insufficient to establish distributional equivalence, especially because the gap could be masked by metric insensitivity. Please report means and standard deviations over multiple seeds, and ideally compare distributions of the intermediate states (e.g., via MMD or FID of X_t_τ) between Recon-Inv and ReInversion. Without this, the statement that the reformulation 'introduces no observable degradation' is not well supported.","section":"Table 3 (Validation of the reformulation)"},{"comment":"The headline results are reported on COCOEE†, a manually curated subset of 2,079 samples after filtering out low-quality references. While the supplementary material includes results on the original COCOEE (Table E), the main-text claim of state-of-the-art performance rests on the curated subset, whose filtering procedure may introduce selection bias that cannot be fully assessed without a detailed protocol. Please move the original-COCOEE results to the main table or at least state explicitly in the main text that both curated and original results are available, and explain how the filtering criterion is model-agnostic.","section":"Sec. 5.1 and Table 1 (COCOEE† benchmark)"}],"minor_comments":[{"comment":"The summation upper limit in Eq. (9) appears to be a typo: it reads 'n−τ' but should likely be 'n−1' (or the notation should be defined consistently).","section":"Sec. 4.2, Eq. (9)"},{"comment":"'denoising from a gaussian noise' should read 'Gaussian noise' for capitalization consistency.","section":"Sec. 4.2, after Eq. (11)"},{"comment":"The figure caption says 'Top: source images; bottom: reconstructions,' but the axes of the images are not labeled. Adding labels or a small legend would improve clarity.","section":"Fig. 3"},{"comment":"No statistical significance tests or confidence intervals are reported for FID, QS, or CLIP metrics. For a stochastic generative method, reporting runs over multiple seeds or at least bootstrap confidence intervals is advisable.","section":"Sec. 5.2, quantitative comparisons"}],"recommendation":"major_revision","confidential_remarks":"The paper presents an attractive and efficient method, and the Eq. 8 identity is sound. The main risk is that the central reformulation is presented as theoretically justified when it actually relies on an unverified empirical property of the backbone. I recommend major revision with a request to either prove or thoroughly empirically validate the equivalence, and to strengthen the benchmarking transparency by presenting original-COCOEE results in the main text. I do not see a fundamental flaw that would require rejection, but the current level of support does not justify the strength of the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper has a genuinely useful observation: for a flow model that reconstructs its conditioning image well, you can skip the backward inversion entirely and just do forward denoising — first towards the source, then towards the exemplar. The ordering makes intuitive sense, and the numbers on COCOEE† are a large jump over FireFlow and RF-Solver. The identity in Eq. 8 is a nice, correct result: inversion error equals reconstruction error, which is the kind of clean relation that suggests the method is not just tuned noise. The MSD masking is simple and effective; the ablation shows a big CLIP-BG improvement.\n\nBut the central theoretical claim in Sec. 4.2 — that replacing the recovered noise X̃0 with a fresh Gaussian X0 is safe because reconstruction error is small — does not follow from Eq. 8. That bound controls the distance between the specific noise used in the forward reconstruction and the recovered noise; it says nothing about the distance to a fresh prior sample. The stress-test note is right. Table 3 is the only direct evidence for the reformulation, and it is a single CLIP-FG number on the same curated benchmark, with no error bars and no distributional check of the transition states. So the claim that the two-stage process 'can be safely reformulated' is an empirical observation in one setting, not a proven equivalence.\n\nOther soft spots: the 'lowest computational overhead' claim is contradicted by their own Table 1, where the base ReInversion at 18 NFEs takes 9.17s versus FireFlow's 7.47s; only the deterministic variant ReInversion* is faster. The 'first work' claim is defensible only under a narrow reading of 'exemplar-guided,' and the category is crowded enough that it will invite a challenge. The code link is a placeholder, hyperparameters appear tuned on the same benchmark, and no error bars are given anywhere.\n\nThat said, the empirical core looks solid. The qualitative results are convincing, the comparisons are against strong recent baselines, and the Eq. 8 identity is a real conceptual contribution. This is not a paper with a load-bearing implementation flaw; it is a paper with an over-claimed theoretical framing. The authors could revise the equivalence claim to a sufficient-condition argument or an empirical finding, add error bars, and fix the efficiency wording. I would cite it if I worked on flow-based editing, and the reading group could have a good discussion about exactly where the theory overreaches.\n\nRecommendation: send to peer review. It deserves a serious referee, and a good one will push on the Eq. 10–11 gap.","headline":"A clean two-stage forward-edit trick that likely works in practice, but the theoretical justification for replacing the inverted noise with a fresh sample (Eq. 11) does not follow from Eq. 8, and the efficiency claim is partly overstated.","tokens_in":15676,"tokens_out":3768,"would_cite":true,"duration_ms":42388,"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":"This paper claims that exemplar-guided image editing can be done without any training by replacing backward inversion with a two-stage forward denoising process — first conditioned on the source image, then on the reference — and that this","keywords":["exemplar-guided image editing","training-free editing","inversion","flow matching","rectified flow","mask-guided denoising","two-stage denoising","diffusion editing"],"falsifier":"Take the same two-stage procedure with a backbone whose reconstruction error is substantially larger than ~0.016 (e.g., a model that reproduces the source with visible artifacts). If editing quality and background preservation collapse as reconstruction error grows, the inversion-error equals reconstruction-error identity is confirmed as the mechanism; if they do not, the method's success is not explained by that identity. A second target: find any source image where the backbone's reconstruction is poor and show ReInversion produces a visibly wrong edit.","tokens_in":14772,"feed_emoji":"🖼️","tokens_out":5692,"duration_ms":53639,"temperature":0.7,"pith_summary":"Standard inversion-based editing drifts because the backward process must guess the velocity field at unknown earlier states. This paper argues the drift is unnecessary: if the editing model is a good reconstructor, the forward reconstruction pass itself yields reliable velocities, and the inversion error equals the reconstruction error. On that basis it reformulates editing as a single forward pass split into two stages — denoise toward the source for a few steps, then switch to denoising toward the reference — and adds a mask that forces the second stage to leave background pixels on a deterministic path toward the source. The result is training-free exemplar-guided editing that runs in 14–18 function evaluations instead of the 36–122 used by prior inversion methods, with reported state-of-the-art fidelity and background preservation. The load-bearing premise: the equivalence between random noise and the recovered source noise holds only because the underlying model reconstructs the source almost perfectly.","feed_headline":"Two-stage inversion edits images in one pass, no training","feed_subtitle":"A mask keeps the background fixed while the foreground adopts the reference, at 14–18 steps.","key_machinery":"The central object is a two-stage Euler integration of a flow-matching ODE, with the conditioning input swapped from the source image to the reference at a transition timestep tτ. The load-bearing identity is Eq. (8): the L2 error of the recovered noise equals the L2 error of the model's reconstruction of the source. Reusing the forward reconstruction velocities makes the inversion 'reversible' in the sense that no backward trajectory needs to be estimated; the mask term blends in a linear velocity (Xs − Xt)/(1−t) outside the target region so the background is pinned to the source.","core_discovery":"ReInversion is a training-free procedure for exemplar-guided editing built on the identity ∥X̃0 − X0∥ = ∥Xs − X̂1∥: the error of the inversion equals the reconstruction error of the model. Because the pretrained flow-matching editor they use reconstructs source images with average per-pixel L2 error of about 0.016, the paper treats the recovered noise as indistinguishable from a fresh Gaussian and reformulates the whole edit as one forward Euler integration that is conditioned on the source for t < tτ and on the reference for t ≥ tτ. A mask-guided selective denoising term then forces the background region to follow a deterministic velocity that points at the source, keeping unedited areas in","pith_inferences":["If the identity holds, inversion becomes unnecessary: the method is effectively 'conditional editing by trajectory stitching,' and the bottleneck shifts to the model's ability to edit from a partially denoised state under reference conditioning.","The deterministic stage-1 variant implies the source-conditioned stage can be replaced by linear interpolation toward the source; a testable extension is whether any flow-matching editor with high reconstruction fidelity can be turned into an EIE tool with this recipe, without retraining.","A sharper test: the method should degrade as reconstruction error grows. Measuring edit quality versus reconstruction L2 across several backbones and source images would either confirm the mechanism or expose that the two-stage switch works for other reasons.","The mask coefficient η is swept to 1.0 in the paper; an open question is whether η can be scheduled over timesteps (strong background pinning early, relaxed late) to allow softer edits while preserving the background."],"forward_implications":["Exemplar-guided editing can be done entirely without training: a two-stage conditioning switch on an off-the-shelf flow-matching editor is sufficient; no learned source–reference mapping is required.","Replacing the backward inversion with forward reconstruction velocities removes accumulated drift, so the edit starts from a reliable latent and preserves source structure.","Adding the mask term sharply improves background preservation (CLIP-BG from ~69 to ~83.5) while also slightly improving foreground alignment.","The reformulation halves the cost of reconstruction-based inversion (18 vs 36 NFEs, a 1.82× speedup) without a measurable drop in editing quality, and it transfers to different flow-based backbones and low step counts."],"fun_headline_variants":["No training, two-stage inversion for exemplar-guided editing","Reversible inversion edits with reference, no training required","Mask-guided selective denoising for training-free exemplar edits","One-pass edit: condition on source, then on reference","ReInversion: edit in 14-18 steps, no training, with mask lock"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Every claim rests on the assumption that a fresh random noise, denoised for a few steps under source conditioning, lands at the same intermediate latent as the source image's own reconstruction — which holds only because the chosen backbone reconstructs the source almost perfectly (L2 ≈ 0.016) and is not demonstrated for out-of-distribution sources or weaker backbones.","fun_headline_variants_meta":{"raw":{"variants":["No training, two-stage inversion for exemplar-guided editing","Reversible inversion edits with reference, no training required","Mask-guided selective denoising for training-free exemplar edits","One-pass edit: condition on source, then on reference","ReInversion: edit in 14-18 steps, no training, with mask lock"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000338,"raw_usage":{"total_tokens":1688,"prompt_tokens":713,"completion_tokens":975,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":457,"completion_tokens_details":{"reasoning_tokens":887}},"tokens_in":457,"tokens_out":975,"duration_ms":10635,"temperature":1.0,"reasoning_tokens":887,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T19:12:28.112394+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the same two-stage procedure with a backbone whose reconstruction error is substantially larger than ~0.016 (e.g., a model that reproduces the source with visible artifacts). If editing quality and background preservation collapse as reconstruction error grows, the inversion-error equals reconstruction-error identity is confirmed as the mechanism; if they do not, the method's success is not explained by that identity. A second target: find any source image where the backbone's reconstruction is poor and show ReInversion produces a visibly wrong edit.","supporting_citations":[],"review_version":1}