{"id":"4f3fdd7a-8904-4584-854e-cd23872f0cae","arxiv_id":"2603.21129","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A rotation-equivariant diffusion U-Net for multi-focus fusion achieves near-zero discrete equivariance error under 90° groups and modest metric gains on four MFIF benchmarks.","lead":"ReDiffuse rebuilds a diffusion U-Net for multi-focus image fusion so the whole network is approximately rotation-equivariant, with error bounds for pooling, upsampling, and the full stack. It cuts parameters and improves standard fusion metrics by reducing orientation-warping artifacts under defocus.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The equivariance guarantee is weakest exactly where the paper claims the biggest benefit—strong defocus—because the regularity assumptions of Theorems 3.1–3.4 fail there.","rationale":"The reader correctly isolates the load-bearing soft spot: the continuous-map regularity (Eq. 20) that underpins the equivariance-error theorems is precisely what strong defocus violates, and the paper itself flags this in §4.6. That does not invalidate the architecture or the multi-dataset gains, but it does make the “theoretical zero error + helpful under defocus” package conditional on regimes where the assumptions hold. No stronger internal inconsistency appears; ablations, generalization swaps, and public code give independent support. Hence the verdict stays CONDITIONAL—accept-shaped contribution once the practical scope of the equivariance guarantee is delimited more carefully—rather than REJECT or UNCHANGED. The concrete residual test above would settle whether the concern is merely theoretical or empirically operative.","tokens_in":16086,"tokens_out":610,"duration_ms":5950,"concrete_test":"On the MFFW failure-case pair (and 5–10 other strong-defocus pairs), compute the empirical equivariance residual ||Φ(π_R I) − π_R Φ(I)|| for R ∈ {90°,180°,270°} after training, both for ReDiffuse and the non-equivariant ablation (Table 3, Config I). If residuals remain near zero while fusion still fails, the guarantee is not the operative mechanism; if residuals spike precisely on those pairs, the continuous-map assumptions fail where claimed most needed.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that end-to-end rotation equivariance (with theoretical zero error for m=4 via Corollary 3.5) yields better orientation/structural consistency under defocus. Theorems 3.1–3.4 and Corollaries 3.5–3.6 bound equivariance error by C1 δ (and C2 m^{-1} δ for continuous angles) only under the uniform bounds on |r_c|, ||∇r_c||, ||∇^{2}r_c|| and the analogous filter bounds in Eq. (20). Defocus blur is a low-pass operator that can inflate local gradient/Hessian estimates of the latent continuous maps (or make them unreliable), so the mesh-size control no longer guarantees near-zero error. The paper’s own §4.6 failure case on MFFW explicitly attributes incorrect structural selection to “severe destruction of local texture structures” and links it to Corollary 3.6 via larger G0/H0. Thus the theoretical support for “faithful preservation under defocus” is least secure on the hardest, most relevant regime, leaving open whether the reported gains come from true equivariance or from the lighter B-Conv parameterization and training recipe.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes ReDiffuse, a diffusion model for multi-focus image fusion (MFIF) whose U-Net backbone is redesigned for approximate end-to-end rotation equivariance via B-Conv (Rot-E) layers, equivariant residual blocks, and redesigned down/upsampling. The motivation is that defocus blur warps edges and textures, so an equivariant inductive bias should better preserve orientation consistency and local isotropic symmetry during conditional denoising. The authors give equivariance-error bounds for max-pooling, bilinear upsampling, and group normalization (Theorems 3.1–3.3), an O(δ) bound for the full network (Theorem 3.4), exact equivariance for the discrete m=4 group (Corollary 3.5), and a continuous-angle bound (Corollary 3.6). Empirically, ReDiffuse is trained on Real-MFF and compared on Lytro, MFFW, MFI-WHU, and Road-MF against 11 baselines across six no-reference metrics, with ablations, plug-in tests on three other diffusion fusion models, equivariance-error maps, and a failure-case discussion. Parameter count is reduced from 26.91M to 7.55M via group parameter sharing.","tokens_in":16433,"tokens_out":1827,"duration_ms":27624,"significance":"If the empirical gains hold under fair comparison, the work is a solid, practically useful contribution to MFIF: it couples a modern diffusion fusion pipeline with a carefully constructed rotation-equivariant U-Net, reports consistent improvements on four public benchmarks, and ships code. The lightweight parameter sharing is a concrete engineering benefit. The theoretical analysis (module-wise bounds plus full-network corollaries) is a genuine strength relative to most fusion papers, even though the regularity assumptions limit how far the guarantees extend under severe defocus. The plug-in generalization experiment (Table 2) and explicit failure-case analysis (§4.6) increase credibility. Overall significance is incremental-to-moderate for the MFIF/diffusion community rather than a foundational advance in equivariant deep learning, but it is a well-scoped and useful application paper.","major_comments":[{"comment":"§3.3 (Eq. 20), Theorems 3.1–3.4, and Corollaries 3.5–3.6 vs. the central claim in the abstract/§1: the equivariance-error theory controls discrete error by mesh size δ only under uniform bounds on values, gradients, and Hessians of latent continuous image/filter maps. Defocus is precisely a regime that can make those local derivatives unreliable. The paper’s own §4.6 failure case on MFFW attributes incorrect structural selection to strong defocus spread and explicitly links it to larger G0/H0 in Corollary 3.6. The theory therefore provides the weakest formal support exactly where the introduction claims the largest benefit. This does not invalidate the empirical results, but the abstract and §1 currently overstate what the theory guarantees under defocus. Please restate the theoretical claims as: (i) exact equivariance for m=4 under the stated regularity assumptions / discrete grid align","section":null},{"comment":"§3.2–3.3 and Figure 4: the analysis establishes (approximate) equivariance of the noise-predictor U-Net modules, but the MFIF pipeline is an iterative reverse process (Eqs. 4–6) conditioned on two source images IA, IB concatenated with Ft. For the end-to-end claim to be load-bearing, please state explicitly that (a) rotating (IA, IB, Ft) jointly induces the corresponding rotation of εθ, and (b) the deterministic reverse step (Eq. 6) preserves equivariance because it is an equivariant function of Ft and εθ. A short proposition or remark would close the gap between “Rot-E U-Net is equivariant” and “the full ReDiffuse fusion map is equivariant.” Without this, “end-to-end rotation equivariance” is slightly stronger than what is proved in the main text.","section":null},{"comment":"§4.1 Implementation / Table 1: fairness of the multi-method comparison needs a clearer protocol statement. The paper says competing methods use “official provided code and weights,” while ReDiffuse is trained 10,000 epochs on Real-MFF. If several baselines were trained on different data, different pair counts, or released only for other fusion settings, the 0.28–6.64% average margins over the second-best method are hard to interpret as pure architectural gains. Please add a short table or paragraph listing, for each baseline, training data, whether weights were used as released or retrained, and input resolution. If retraining all diffusion baselines on Real-MFF under the same schedule is infeasible, at least retrain the closest non-equivariant backbone (the “w/o rotation equivariance” model already in Table 3) and report that as the primary controlled comparison, with official-weight ba","section":null}],"minor_comments":[{"comment":"Figure 1 caption and §1: “non-local directional similarity” and “local isotropic symmetry” are illustrated but not operationalized. A one-sentence definition or pointer to how these are measured (even qualitatively) would help readers who are not already convinced by the red/blue boxes.","section":null},{"comment":"Table 1, Lytro: TC-MoA exceeds ReDiffuse on QMI (0.993 vs 0.982). The text says ReDiffuse is best “across nearly all” metrics; please acknowledge the few second-place entries explicitly when claiming average improvements.","section":null},{"comment":"§3.2: T=2000 is stated without justification relative to FusionDiff or other diffusion MFIF work. A brief note on sampling steps at inference (and any DDIM-style acceleration) belongs in §4.1 or Appendix B.","section":null},{"comment":"Related Work §2.2: the claim that “theoretical guarantees for enforcing rotation equivariance within diffusion models remain largely unexplored” should cite any concurrent equivariant-diffusion work outside MFIF (even if not fusion-specific) to avoid overclaiming novelty of the theory setting.","section":null},{"comment":"Notation: π^I_R̃ vs π^F_R̃ and the composition notation [·] in Eq. (9) are dense; a short example for m=4 on a single feature map in Appendix A.1 would improve accessibility.","section":null},{"comment":"Figure 8: “after 1,000 training epochs” is a mid-training snapshot. Please also report final-epoch equivariance error (mean/max over the test set) so the maps can be tied to the deployed model.","section":null},{"comment":"Typos / polish: “ReDiffusetheoretically” (Fig. 4 caption); “o90” labels in Fig. 1; inconsistent spacing around citations (e.g., “methods [2, 26, 38?]”); “MS-SSIM:0.970” style in Fig. 2 could use a space after the colon.","section":null},{"comment":"§4.2: runtime is deferred to Appendix C.1; a one-line statement in the main text (e.g., relative FPS vs FusionDiff) would help readers assess the cost of equivariance.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The skeptic’s concern about regularity assumptions under strong defocus is real but already partially owned by the authors in §4.6; I treat it as a scoping/writing issue rather than a reason to reject. Empirical package (four datasets, ablations, plug-in tests, code link) is above average for this venue track. I would not block on novelty of B-Conv itself (prior work of overlapping authors) so long as the MFIF diffusion integration and full-network error analysis are clearly positioned as the contribution. Fit is appropriate for a solid CV / information-fusion journal; borderline for a top general ML venue unless the theory appendices are unusually complete."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is simple: they built a full diffusion U-Net for multi-focus fusion that is rotation-equivariant end-to-end (B-Conv plus equivariant down/upsampling and GN), proved discrete equivariance error bounds for that stack, and show consistent wins on four benchmarks plus a plug-in transfer to three other diffusion fusion models. Code is public; parameter count drops from ~27M to ~7.5M via group sharing.\n\nWhat is actually new is the assembly and the analysis, not the individual pieces. B-Conv, steerable/group convs, and diffusion MFIF already exist; they cite them. The contribution is making the whole denoising U-Net (including the operators that usually break equivariance on grids) approximately equivariant, giving Theorems 3.1–3.4 and Corollaries 3.5–3.6 (zero error for m=4 under the stated regularity), and showing that this helps orientation/structural consistency under defocus. Experiments are the right shape: 11 baselines, ablations (strip equivariance; swap G-Conv/F-Conv), generalization table, equivariance-error maps, difference images, and an honest failure case. Circularity is low—the loss is ordinary noise prediction; equivariance is architectural.\n\nThe soft spot the stress-test flags is real but not fatal. The bounds need uniform gradient/Hessian control on latent continuous maps (Eq. 20). Strong defocus is exactly where those local derivatives become unreliable, and §4.6 says so. So the “theoretical zero error under defocus” claim is overstated for the hardest regime; gains there may partly come from the lighter parameterization and training recipe. That is a limit of the theory, not a contradiction of the tables. Other minor issues: MS-SSIM is a proxy average against sources (no GT), no uncertainty on the metric deltas, and significance stays inside MFIF/restoration.\n\nThis is for people who care about geometric inductive bias in generative fusion or who want a reusable equivariant diffusion backbone. Math and citations look careful; data package is above average for the area. I would send it to peer review. Engage if you work on equivariant denoisers or MFIF; otherwise skim the theory + Table 1.","headline":"Solid subfield methods paper: end-to-end rotation-equivariant diffusion U-Net for MFIF with real equivariance-error analysis and consistent gains; theory is weakest exactly on hard defocus, but the empirical package still deserves referees.","tokens_in":17074,"tokens_out":581,"would_cite":true,"duration_ms":5717,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Embedding rotation equivariance into a diffusion U-Net stops defocus from warping fused multi-focus images.","keywords":["Rotation Equivariant","Diffusion Model","Multi-focus Image Fusion","Equivariance Error Analysis","U-Net","B-Conv","Defocus Blur"],"falsifier":"Replace every Rot-E block with ordinary convolutions while keeping identical capacity and training; if the six fusion metrics on Lytro, MFFW, MFI-WHU and Road-MF no longer improve and the measured equivariance-error maps no longer collapse to near zero under 90/180/270 rotations, the central claim fails.","tokens_in":16951,"feed_emoji":"🖼️","tokens_out":998,"duration_ms":9561,"temperature":0.7,"pith_summary":"Multi-focus image fusion must combine several differently focused shots into one sharp all-in-focus picture. Diffusion models can generate high-fidelity results, yet defocus blur warps edges and textures, so the network often invents artifacts while denoising. The paper argues that the missing inductive bias is rotation equivariance: if a pattern is rotated, the network’s response must rotate with it, preserving orientation and local isotropic symmetry. ReDiffuse rebuilds every block of the diffusion U-Net (convolutions, down-sampling, up-sampling, normalization) so that the whole pipeline is approximately rotation-equivariant end-to-end. A sequence of theorems bounds the residual equivariance error by the mesh size and shows that the error is exactly zero for the common 90-degree rotation group. On four standard benchmarks the method improves six quality metrics by 0.28–6.64 percent while cutting parameters from 26.91 M to 7.55 M. A sympathetic reader cares because the same geometric prior can be dropped into other diffusion fusion models and immediately raises their scores, and because the theory supplies a concrete reason why the prior works.","feed_headline":"Rotation-equivariant diffusion fixes warped multi-focus fusion","feed_subtitle":"End-to-end geometric bias cuts artifacts and lifts six metrics while shrinking the model three-fold","key_machinery":"Rotation-equivariant (Rot-E) U-Net built from B-Conv filters, equivariant max-pooling / bilinear up-sampling, and group normalization; Theorems 3.1–3.4 and Corollaries 3.5–3.6 bound its intrinsic equivariance error by mesh size δ and prove exact equivariance when m = 4.","core_discovery":"Defocus blur destroys reliable orientation cues inside a diffusion denoiser; forcing the entire U-Net to be rotation-equivariant restores those cues, so fused images keep the original directional consistency of textures and edges. For the discrete 90-degree group the equivariance error is theoretically zero, and experiments confirm both smaller residual maps and higher fusion metrics.","pith_inferences":["Because the equivariance guarantee is discrete and mesh-size dependent, continuous or arbitrary-angle rotations still leave a residual error that may matter for camera-shake or free-form viewpoint changes.","The same Rot-E blocks could be inserted into multi-modal fusion pipelines (infrared-visible, medical) where orientation consistency is equally fragile under blur or sensor misalignment.","If non-local attention is later combined with the Rot-E U-Net, the failure case on strong defocus spread may disappear without sacrificing the zero-error corollary for m = 4."],"forward_implications":["Existing diffusion MFIF backbones can be upgraded by swapping in Rot-E modules and should inherit both better metrics and a roughly 3\times parameter reduction.","For the standard four-fold rotation group the theoretical equivariance error vanishes, so residual orientation artifacts can no longer be blamed on the architecture.","The same construction supplies a practical recipe for any diffusion image-restoration task whose ground-truth structures are approximately rotation-stable.","Failure under severe defocus spread points to non-local correlation modules as the next architectural addition needed to keep the equivariance guarantee useful."],"fun_headline_variants":["Rotation equivariance stops defocus warp in multi-focus diffusion","Equivariant diffusion restores edge orientation in MFIF","Zero-error 90° equivariance fixes artifacts in focus fusion","Rotation-equivariant U-Net keeps texture consistency in fusion","Embedding equivariance lifts six MFIF metrics without warp"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The error bounds assume that latent continuous image and filter maps have uniformly bounded gradients and Hessians; under strong defocus those local derivatives become unreliable, which is exactly where the method can still fail.","fun_headline_variants_meta":{"raw":{"variants":["Rotation equivariance stops defocus warp in multi-focus diffusion","Equivariant diffusion restores edge orientation in MFIF","Zero-error 90° equivariance fixes artifacts in focus fusion","Rotation-equivariant U-Net keeps texture consistency in fusion","Embedding equivariance lifts six MFIF metrics without warp"]},"model":"grok-4.5","effort":"low","cost_usd":0.006,"raw_usage":{"total_tokens":1576,"prompt_tokens":768,"num_sources_used":0,"completion_tokens":69,"cost_in_usd_ticks":60000000,"prompt_tokens_details":{"text_tokens":768,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":739,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":768,"tokens_out":69,"duration_ms":7276,"temperature":1.0,"reasoning_tokens":739,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T21:01:32.600057+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Replace every Rot-E block with ordinary convolutions while keeping identical capacity and training; if the six fusion metrics on Lytro, MFFW, MFI-WHU and Road-MF no longer improve and the measured equivariance-error maps no longer collapse to near zero under 90/180/270 rotations, the central claim fails.","supporting_citations":[],"review_version":1}