{"id":"d37a57c5-f9dc-4996-9407-1ee19560a7ae","arxiv_id":"2608.08664","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"Fixed-noise refinement with one-step Pixel Mean Flow prediction produces visual counterfactuals about 3x faster than the strongest diffusion baseline while keeping image quality and localization competitive or better.","lead":"This paper presents a new way to generate visual counterfactual explanations, images that show what would need to change for an AI classifier to change its decision. It edits an image at a single fixed noise level instead of the usual long denoising path, cutting inference time by about three times while preserving edit quality.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing premise is that Eq. (8)'s PMF-backpropagated gradients are reliable descent directions in the fixed-noise state; the paper provides no direct test of this, and the k=0 computation path in Algorithm 1 is ambiguous.","rationale":"The reader's weakest assumption is also the one I would stress-test. The paper's empirical scaffolding is genuinely useful: it reruns external baselines, provides a progressive ablation in Table 4, includes sensitivity tables for K and t, and explicitly acknowledges the COUT-versus-sparsity trade-off. That is why I do not move to REJECT: the problem is not an observed contradiction but an unvalidated critical mechanism. However, the method's stated advantage depends on classifier gradients being transported through PMF in a way that improves the target probability at a fixed noise level. A one-step denoiser's Jacobian can be poorly conditioned or can amplify unrealistic modes; because the loss is evaluated on a blended image and the update is applied under a union hard mask, the effective descent direction could be dominated by mask accumulation or early stopping rather than by the geometry of PMF. The paper contains no diagnostic for this. The k=0 graph issue adds a small but concrete ambiguity: if the code follows Algorithm 1 as written, the first refinement step is vacuous. This does not necessarily break the method, but it means the reported implementation may differ from the equations. A focused gradient-validity experiment on one task would settle whether fixed-noise refinement is actually doing the work or whether the controls alone carry the result. If the check passes, CONDITIONAL can be upgraded; if it fails, the central claim would need substantial revision. Thus the verdict stays CONDITIONAL.","tokens_in":15175,"tokens_out":8880,"duration_ms":101236,"concrete_test":"Run literal Algorithm 1 on 200 CelebA-HQ Smile images with instrumentation. At k=1..3, before applying Eq. (10), compute (a) a finite-difference approximation of the gradient of p_target(X^k) with respect to Z_t^k restricted to M_hard^k, using random projections with small epsilon, and measure its cosine similarity to the normalized Eq. (8) gradient; and (b) the one-step change in p_target(X^{k+1}) for the Eq. (8) update versus a same-norm random masked update. If the median cosine is near zero or negative, or if the Eq. (8) update does not raise p_target substantially above the random control, the fixed-noise descent premise fails and the reported quality must be attributed to masking and early stopping rather than the claimed refinement.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"FiRe's central claim is that refining a single noisy state with gradients guided by a clean-image classifier produces efficient, localized counterfactuals. The key link is Eq. (8): the clean-image loss gradient is transported back through the PMF prediction and soft-mask update path via ∂X^k/∂Z_t^k. Everything downstream—the K=15 updates, hard-mask accumulation, and early stopping—is only as good as this transported gradient. PMF is trained as an unconditional one-step clean-image predictor, not as an editor, and the paper gives no conditioning, bias, Lipschitz, or finite-difference check for this composite Jacobian. Since X^k is constructed by blending PMF output with original pixels outside M_soft^{k-1}, the backprop path mixes PMF denoising with identity; whether that mixture points reliably toward higher p_target at fixed t=0.4 is unverified. The ablations show the full system works, but they do not isolate whether the fixed-noise descent itself, rather than the dual mask plus early stopping, is responsible for the improvements. There is also a concrete graph ambiguity at k=0: Algorithm 1 sets X^0=X0, so by Eq. (8) ∂X^0/∂Z_t^0=0 and the first update is vacuous. This makes the intended computation graph not fully specified and reinforces that the gradient mechanism is the least secured load-bearing link between the method and the headline efficiency and quality numbers.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FiRe, a fixed-noise refinement framework for visual counterfactual explanations. Instead of following a DDPM reverse denoising trajectory with changing noise levels, FiRe maps the input to a fixed intermediate noise level and iteratively refines the noisy state at that level. A Pixel Mean Flow (PMF) predictor produces clean classifier-facing images in one step, and the clean-image classifier loss is backpropagated through the PMF and soft-mask blending path to obtain noisy-state gradients. Three controls are introduced for fixed-noise refinement: a dynamic dual-mask strategy (a union hard mask for noisy-state edit memory and a feathered soft mask for clean-image visibility), adaptive guidance that strengthens or weakens updates based on target confidence, and early stopping once the target probability reaches a threshold. The method is evaluated on five tasks across CelebA, CelebA-HQ, and CheXpert, reporting about 3x faster online inference and 8x fewer FLOPs than the strongest recent baseline (MaskDiME) while achieving comparable or better counterfactual quality on most metrics.","tokens_in":15475,"tokens_out":8612,"duration_ms":91821,"significance":"If the claims hold, FiRe is a meaningful step for efficient classifier-guided counterfactual editing: it avoids both recursive denoising and low-quality one-step estimates by combining a fixed-noise optimization state with a one-step clean-image predictor, and it is the first to formulate counterfactual generation as fixed-noise refinement rather than trajectory editing. The paper is transparent: it reruns all baselines under its own CheXpert protocol, reports the full ablation progression from a basic PMF-guided generator to the complete FiRe model, and explicitly discusses the lower COUT values as a design choice. The empirical evidence is substantial, with consistent improvements in FID/sFID and MNAC across four face tasks and a strong medical shortcut-removal result, alongside a clear efficiency comparison.","major_comments":[{"comment":"The computation graph at k=0 is underspecified. Algorithm 1 sets X^0 to be the input image and then, at k=0, computes ∇_{Z_t^0} L_joint^0 via Eq. (8). If X^0 is treated as a constant, then ∂X^0/∂Z_t^0 = 0 and the first update in Eq. (10) is vacuous, meaning the effective refinement budget is K-1 guided steps preceded by a mask-constrained one-step PMF prediction. Alternatively, the authors may intend to differentiate through the inverse of Eq. (2), X^0 = (Z_t^0 - t Z^0)/(1-t), which gives ∂X^0/∂Z_t^0 = I/(1-t) and a nontrivial first step. The text says that ∂X^k/∂Z_t^k follows the path 'including PMF prediction and soft-mask clean-image update', which does not apply to X^0. This ambiguity affects the meaning of the K=15 budget, the reported efficiency accounting, and reproducibility; please specify the intended graph explicitly and, if necessary, correct Algorithm 1.","section":"Algorithm 1 / §3.5, Eq. (8)"},{"comment":"The central mechanism of FiRe is the gradient of the clean-image loss transported through the PMF predictor and soft-mask blending back to the fixed-noise state. The paper provides no direct evidence that this composite Jacobian is a reliable descent direction at t=0.4. The ablations show that the full system works and that PMF-Fixed outperforms PMF-Reverse, but they do not isolate the transported gradient from the dual-mask and early-stopping controls; the mask makes a large contribution (FID drops from 19.13 to 3.97 when the FiRe mask is added on top of PMF-Fixed). Please add a diagnostic that directly tests the gradient path, for example by measuring the alignment between ∇_{Z_t} L_joint and the finite-difference change in p_target after a small masked update, or by comparing the fixed-noise refinement with a variant that uses a one-step DDPM Tweedie estimate instead of PMF while keeping all controls identical. This would substantiate the claim that fixed-noise refinement, rather than the spatial controls alone, is responsible for the quality and efficiency gains.","section":"§3.5, Eq. (8) / Table 4"}],"minor_comments":[{"comment":"The notation [·]_+ in Eq. (9) is not defined; please state that it clamps negative values to zero, i.e., [x]_+ = max(x, 0).","section":"§3.5, Eq. (9)"},{"comment":"The metric MAD is reported but never defined in the text; please define it explicitly in §4.1.","section":"Table 3"},{"comment":"The PMF-Reverse row is described as '15-step refinement along a DDPM-based reverse denoising trajectory', but the noise schedule is not specified. Please provide the schedule or a reference so that the comparison between reverse and fixed refinement is reproducible.","section":"Table 4"},{"comment":"The marker-size legend for peak GPU memory is described in the caption for panel (a) but panel (b) also uses marker sizes; please clarify whether the same legend applies to both panels.","section":"Figure 3"},{"comment":"Several entries for DiVE, DiVE100, and STEEX are shown as '–'; please state in the captions whether these values were not reported in the original papers or were not computed in this work.","section":"Tables 1 and 2"}],"recommendation":"major_revision","confidential_remarks":"MaskDiME [11], the main baseline and the basis for the headline '3x faster / 8x fewer FLOPs' comparison, shares two authors with this manuscript (Changlu Guo and Anders Nymark Christensen, among the FiRe authors). This is not problematic in itself, but the editor may wish to verify that the MaskDiME numbers used here come from the published implementation and that the comparison protocol is independently audited. Also, the k=0 ambiguity in Algorithm 1 should be resolved in revision; if the implementation differentiates through the inverse noising equation, that path should be stated explicitly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"FiRe is worth a serious look. The core idea—keep the editable state at a fixed noise level and use Pixel Mean Flow to get clean images for classifier guidance—is genuinely new relative to the DiME/ACE/FastDiME/MaskDiME line, and the paper supports it with a coherent set of experiments. I would send this to a strong referee.\n\nThe method is well designed: the union hard mask for edit memory and the soft mask for visible changes is a sensible separation that the ablations actually validate. The efficiency claim (roughly 3x faster, 8x fewer FLOPs than MaskDiME) is backed by a rerun on the same hardware, and the quality numbers on CelebA, CelebA-HQ, and CheXpert are consistently at or near the top. The ablation progression from one-step PMF to fixed-noise with masks and early stopping is the right way to show each component earns its keep. I also appreciate that the authors openly discuss the lower COUT values and do not overclaim medical generalization.\n\nWhere I would push back: there are no error bars on any headline metric, and the task-specific saliency ratio and loss weights are tuned without a disclosed validation protocol, so selection effects are hard to rule out. The strongest baseline MaskDiME is same-group work; the authors rerun external baselines, which mitigates it, but independent runs would be better. The stress-test note about Eq. (8) is fair but not fatal: the gradient path through PMF is a heuristic with no stability analysis, and Algorithm 1 has a genuine ambiguity at k=0—since X^0 is set to X0, the derivative ∂X^0/∂Z_t^0 is zero and the first update is a no-op. That is a minor implementation bug, but it shows the computation graph is not fully pinned down. None of this undermines the central empirical claim; it just means the paper needs a revision pass rather than immediate acceptance.\n\nBottom line: this is a solid contribution to the visual-counterfactual subfield, not a paradigm shift. It deserves a proper peer review with a request for code and data, error bars, and a clearer derivation of the gradient path. I would take it to our reading group if anyone is still tracking counterfactual explanations.","headline":"FiRe is a genuine new formulation for visual counterfactuals with real efficiency gains; the paper deserves peer review but needs error bars, a validation protocol, and code before the headline numbers are fully trustworthy.","tokens_in":16085,"tokens_out":3111,"would_cite":true,"duration_ms":32424,"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":"FiRe reframes visual counterfactual explanation as iterative refinement at one fixed noise level, using Pixel Mean Flow to predict clean images and three controls to keep edits localized, achieving about 3x faster inference and 8x fewer…","keywords":["visual counterfactual explanations","fixed-noise refinement","Pixel Mean Flow","diffusion models","classifier guidance","spatial masks","medical imaging","explainable AI"],"falsifier":"A decisive experiment would replace PMF with a Tweedie one-step estimate, keeping everything else fixed; the outcome separates whether the gain comes from fixed-noise refinement or from the one-step predictor.","tokens_in":14942,"feed_emoji":"🎯","tokens_out":5778,"duration_ms":56434,"temperature":0.7,"pith_summary":"This paper claims that classifier-guided visual counterfactual explanation does not need to follow a long reverse denoising trajectory. Instead, it maps the input image to a single intermediate noise level and refines that noisy state in place, using Pixel Mean Flow to predict a clean image for the classifier at every step. The authors argue this fixed-noise formulation keeps enough semantic editability while improving spatial locality, and that their three controls—a union hard mask for edit memory, a soft mask for visible changes, and adaptive guidance with early stopping—prevent over-editing and drift. On five tasks across CelebA, CelebA-HQ, and CheXpert, they report about 3x faster online inference and 8x fewer FLOPs than the strongest recent baseline, with comparable or better counterfactual quality. The point would matter because it suggests efficient, localized counterfactual explanations can be built without recursive denoising or low-quality one-step estimates.","feed_headline":"One fixed noise level gives faster, cleaner counterfactual edits","feed_subtitle":"FiRe edits classifier images at a single noise level, skipping the slow denoising trajectory while keeping edits localized.","key_machinery":"The central mechanism is Pixel Mean Flow used as a one-step clean-image predictor that turns any noisy state at a fixed noise level $t$ into a clean image the classifier can evaluate. Around that predictor, FiRe builds an update loop: a SmoothGrad attribution map selects the top-$\\rho$ salient region, a union hard mask accumulates edit regions across steps, a soft mask feathers the visible edit boundary, and an adaptive guidance scale $\\eta^k$ strengthens updates when the target probability is far from the flip threshold and weakens them as it approaches. The update is applied to the noisy state under the hard mask, PMF predicts a clean candidate, and early stopping returns the first clean image whose target probability reaches $P_{\\mathrm{flip}}$.","core_discovery":"FiRe's central claim is that fixed-noise refinement is a better formulation than DDPM reverse-trajectory editing for visual counterfactual explanation. The paper shows that after forward-noising an input to a fixed level $t=0.4$, iteratively updating the noisy state with classifier gradients backpropagated through a Pixel Mean Flow clean-image predictor produces edits that flip the target classifier with high validity. The method couples this update with a union hard mask that preserves previously selected edit regions, a feathered soft mask that controls visible blending, and a confidence-progress guidance scale with early stopping. The reported experiments support that this formulation reaches the target decision with fewer unintended changes and lower FID than trajectory-based baselines, while cutting compute drastically.","pith_inferences":["A direct test of how much of the gain is due to fixed-noise refinement versus the one-step predictor: replace Pixel Mean Flow with a Tweedie one-step estimate at the same noise level; if flip rate collapses, the gradient path through PMF is the load-bearing part, not the fixed noise level.","The union hard mask acts as persistent edit memory, which could be reused for interactive editing or multi-round counterfactual refinement where earlier edits must survive later updates.","Applying the same fixed-noise loop inside a latent diffusion or rectified-flow model, where one-step clean-image prediction is already available, is a natural next step; the mask-memory and adaptive-guidance controls are defined for pixel space but should transfer to latent coordinates.","Because guidance is computed on a clean image but applied through a learned predictor, the composite gradient may be biased; a calibration experiment comparing PMF-predicted clean-image gradients with ground-truth clean-image gradients at $t=0.4$ would test that bias directly."],"forward_implications":["Counterfactual generation can run at a single noise level without recursive denoising or low-quality one-step estimates, cutting online inference time roughly threefold and FLOPs roughly eightfold versus MaskDiME.","The union hard mask lets previously selected edit regions stay active across refinement steps, so repeated updates at the same noise level accumulate rather than interrupt the counterfactual change.","Early stopping at a target-probability threshold trades a modest drop in classifier-output transition for better realism and sparsity; the ablation shows FID rises from 1.98 to 3.47 when early stopping is disabled.","The same fixed-noise recipe transfers to a 512x512 medical shortcut-removal task with a different data distribution, indicating the formulation is not face-specific."],"supporting_citations":[{"why":"Supplies the Pixel Mean Flow one-step clean-image predictor that lets classifier guidance run on clean images at a fixed noise level.","marker":"[34]"},{"why":"The strongest recent baseline and the trajectory-based adaptive-mask method FiRe compares against and extends.","marker":"[11]"},{"why":"Introduces DDPM reverse-trajectory editing that FiRe replaces with fixed-noise refinement.","marker":"[19]"},{"why":"Provides the one-step Tweedie estimate baseline and the CheXpert shortcut-removal protocol FiRe adopts.","marker":"[44]"},{"why":"Provides the evaluation protocol and metrics used for the face counterfactual tasks.","marker":"[20]"},{"why":"Defines the reverse denoising process that fixed-noise refinement is designed to avoid.","marker":"[15]"},{"why":"SmoothGrad stabilizes the attribution maps used to build the dual masks.","marker":"[41]"}],"fun_headline_variants":["Fixed-noise refinement triples counterfactual edit speed","FiRe: one noise level, 8x fewer FLOPs for counterfactuals","Skip denoising trajectory: fixed-noise counterfactual edits","Single noise level accelerates counterfactual editing 3x","Fixed-noise refiner localizes edits and cuts compute 8x"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that gradients of the target-classifier loss, sent backward through PMF's one-step clean-image prediction and the soft-mask blend, stay reliable as a descent direction in the fixed-noise state for up to 15 accumulated updates.","fun_headline_variants_meta":{"raw":{"variants":["Fixed-noise refinement triples counterfactual edit speed","FiRe: one noise level, 8x fewer FLOPs for counterfactuals","Skip denoising trajectory: fixed-noise counterfactual edits","Single noise level accelerates counterfactual editing 3x","Fixed-noise refiner localizes edits and cuts compute 8x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1558,"prompt_tokens":951,"completion_tokens":607,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":517}},"tokens_in":567,"tokens_out":607,"duration_ms":6260,"temperature":1.0,"reasoning_tokens":517,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:28:10.675981+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive experiment would replace PMF with a Tweedie one-step estimate, keeping everything else fixed; the outcome separates whether the gain comes from fixed-noise refinement or from the one-step predictor.","supporting_citations":[{"cited_title":"MaskDiME: Adaptive Masked Diffusion for Precise and Efficient Visual Counterfactual Explanations","cited_arxiv_id":"2602.18792","evidence_quote":"The strongest recent baseline and the trajectory-based adaptive-mask method FiRe compares against and extends."},{"cited_title":"Diffusion models for counterfac- tual explanations","cited_arxiv_id":null,"evidence_quote":"Introduces DDPM reverse-trajectory editing that FiRe replaces with fixed-noise refinement."},{"cited_title":"Fast diffusion-based counterfactuals for shortcut removal and generation","cited_arxiv_id":null,"evidence_quote":"Provides the one-step Tweedie estimate baseline and the CheXpert shortcut-removal protocol FiRe adopts."},{"cited_title":"Adversarial counterfactual visual explanations","cited_arxiv_id":null,"evidence_quote":"Provides the evaluation protocol and metrics used for the face counterfactual tasks."}],"review_version":1}