{"id":"335cd840-b679-427a-891e-992d08893a44","arxiv_id":"2509.09365","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A linear average of GAP and HQS data-consistency updates during DDIM sampling is proposed, but it reduces to a single scaled projection when the sensing matrix is orthogonal.","lead":"Diffusion models for image reconstruction can be guided by two standard data-consistency rules, GAP and HQS. This paper averages the two rules during DDIM sampling for single-pixel imaging, but under the paper's own orthogonality assumption the average is just one rule with a rescaled step.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The fused GAP/HQS update degenerates to a single projected-gradient step under the paper's own HH^T=I assumption; the claimed 'hybrid' of soft and hard constraints reduces to a step-size reparameterization, and the orthogonality condition is never verified.","rationale":"The reader identified the orthogonality assumption as the weakest point. My concern sharpens it: even under the assumption, the proposed hybrid does not combine two mechanisms; it is equivalent to a single projected-gradient step with a tunable step size. If the assumption fails, the paper's Eq. 19 and its interpretation are invalid, and the algorithm's true behavior is uncharacterized. Either way, the central conceptual claim is not established. The experimental support is also insufficient for the abstract's strong claim of consistent improvement: Table 1 only reports CR=5%, lacks direct baseline entries for DDNM/DiffPIR, and provides no error bars. Therefore the paper's main assertion is over-reaching, and the REJECT verdict is justified.","tokens_in":7788,"tokens_out":7565,"duration_ms":72175,"concrete_test":"Re-derive Eq. 16 under HH^T=I using the SVD of H, or the identity (H^T H+λI)^{-1}=λ^{-1}(I-H^T(HH^T+λI)^{-1}H), and verify symbolically that Eq. 16 equals Eq. 19 with ρ=1-λδ/(1+λ), and that HQS equals GAP with step size 1/(1+λ). Then, in the authors' experimental setup, compute UU^T and VV^T for the binary masks used in Algorithm 1; if either is not exactly identity, Eq. 19 is not the correct description of the algorithm, and the paper must state whether the claimed equivalence holds. A quick numerical check: for any test image, run Algorithm 1 with δ=0 and with ρ=1; outputs should be identical if HH^T=I.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's central claim (Sec. 3.2) is that linearly combining PnP-GAP and PnP-HQS updates in Eq. 16 yields a hybrid that 'combines strict and soft constraints' and 'improves measurement consistency without disrupting the diffusion sampling trajectory.' This is not supported by the paper's own algebra. Just before Eq. 19 the paper assumes HH^T=I (i.e., UU^T=I and VV^T=I for the separable SPI model). Under this assumption, the Moore-Penrose inverse simplifies to H†=H^T, and the HQS update (H^T H+λI)^{-1}(H^T y+λx) is exactly the same projected-gradient direction as GAP, x+H^T(y-Hx), but with step size 1/(1+λ) instead of 1. Both updates leave the null-space component of x unchanged and differ only in how strongly they replace the measurement-subspace component by the projected measurement. Consequently Eq. 16 equals x+ρ H^T(y-Hx) with ρ=1-λδ_t/(1+λ): a single projected gradient step. The fusion weight δ_t is therefore just a reparameterization of the step size, not a combination of two distinct mechanisms. Any δ_t can be reproduced by a suitably chosen ρ in DDNM, so the reported gains relative to DDNM/DiffPIR (if real) are step-size effects, not evidence for a new hybrid mechanism. If, alternatively, the experimental masks do not satisfy HH^T=I, Eq. 19 and the projected-gradient interpretation are invalid, and the paper never reports whether its random binary masks satisfy UU^T=I, VV^T=I. The manuscript's own text (Sec. 3.3) flags this as a supposition but never verifies it in Sec. 4. Thus the conceptual contribution is either vacuous or unjustified. Additionally, Table 1 reports a single 5% CR with no error bars and no direct DDNM/DiffPIR entries, so the abstract's 'consistently outperforms' is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a unified view of plug-and-play (PnP) optimization and DDIM sampling for image compressive sensing. It decouples the DDIM update into denoising, data-consistency, and sampling stages, and then introduces a 'hybrid' data-consistency module that linearly combines the PnP-GAP and PnP-HQS updates during diffusion sampling. The method is tested on simulated single-pixel imaging at compression ratios of 1--20%, with reported PSNR/SSIM/LPIPS improvements over a small set of baselines.","tokens_in":8232,"tokens_out":8376,"duration_ms":89544,"significance":"If the central claim were valid, the paper would offer a simple and intuitive way to combine hard and soft data-consistency constraints within diffusion-based restoration, with practical value for single-pixel imaging. The paper usefully emphasizes the PnP/diffusion connection, and the proposed modular decomposition of DDIM is a reasonable expository framework. However, the key technical claim is not supported by the paper's own algebra: under the stated orthogonality assumption the 'fusion' degenerates into a single projected-gradient step, so the alleged combination of two distinct mechanisms reduces to a step-size reparameterization. In addition, the experimental section does not actually compare against the DDNM and DiffPIR baselines invoked in the conclusions. Because the main novelty is the hybrid mechanism, these issues are load-bearing rather than cosmetic.","major_comments":[{"comment":"Under the paper's own assumption HH^T=I stated before Eq. (19), the fused update in Eq. (16) is not a genuine hybrid. Let P=H^T H, which is an orthogonal projector when HH^T=I. Then the GAP update is x + H^T(y-Hx) = (I-P)x + H^T y, while the HQS update is (P+λI)^{-1}(H^T y + λx) = (I-P)x + (1/(1+λ))H^T y + (λ/(1+λ))Px = x + (1/(1+λ))H^T(y-Hx). Substituting both into Eq. (16) gives x + ρ H^T(y-Hx) with ρ = 1 - λδ_t/(1+λ), which is exactly Eq. (19). Thus the two 'mechanisms' differ only in their range-space step size; the fusion weight δ_t is a reparameterization of that step size. The abstract and Sec. 3.2 claim that the hybrid 'combines strict and soft constraints' and 'improves measurement consistency without disrupting the diffusion trajectory'; this is not supported by the derivation. Any gain from tuning δ_t could be reproduced by choosing a suitable step size in DDNM/GAP alone.","section":"Sec. 3.2, Eqs. (14)-(19)"},{"comment":"Algorithm 1 computes the exact GAP term H†(y-Hx) and the exact HQS term (H^T H+λI)^{-1}(H^T y+λx), whereas Eq. (19) is derived under the orthogonality condition HH^T=I. The manuscript never reports whether the experimental binary masks satisfy UU^T=I and VV^T=I (the condition just before Eq. (19) is merely stated as a supposition, and the text even contains a typo in the condition). If the masks do not satisfy the condition, Eq. (19) does not describe Algorithm 1 and the projected-gradient interpretation is invalid. If they do satisfy it, the fusion collapses to a step size as explained above. Either way, the paper does not establish that Algorithm 1 combines two distinct data-consistency mechanisms.","section":"Sec. 3.3, Algorithm 1, Eq. (19)"},{"comment":"The abstract and conclusion claim the method 'consistently outperforms DDNM and DiffPIR baselines', but the experiments do not contain DDNM or DiffPIR results. Table 1 reports only H†y, DDIM+GAP, DDIM+HQS, and the proposed fusion; Table 2 reports only the proposed method at different compression ratios, without baseline columns. Moreover, at the reported 5% ratio the gain over DDIM+HQS is 0.12 dB PSNR, with no variance or statistical significance information. The claim of consistent improvement over DDNM/DiffPIR is therefore unsupported by the presented evidence.","section":"Sec. 4.2, Tables 1-2; Sec. 5"}],"minor_comments":[{"comment":"For Y=UXV^T, the vectorized forward model should be H=V⊗U, not U⊗V (with column-major vectorization). Equation (19) is consistent with H=V⊗U, so Eq. (18) appears to be a typo that should be corrected.","section":"Eq. (18)"},{"comment":"The line 'which implies UU^T ⊗ HH^T = I' appears to contain a typo: it should be UU^T ⊗ VV^T = I.","section":"Sec. 3.3"},{"comment":"The text says 'five compression rates: 1%, 5%, 10% and 20%' but lists four rates. Also, the values of λ and the exact δ_t schedule are not specified; the reader cannot reproduce the experiments from the text.","section":"Sec. 4.1"},{"comment":"Line 9 introduces √¯α_{t-1}, w_t, and ζ without defining them in the algorithm or the surrounding text. This notation should be reconciled with Eq. (13).","section":"Algorithm 1"},{"comment":"The caption contains a typo: 'Visulization' should be 'Visualization'.","section":"Fig. 4"}],"recommendation":"reject","confidential_remarks":"The central contribution of the manuscript is the claimed hybrid GAP/HQS data-consistency mechanism, but under the paper's own orthogonality assumption the fused update reduces to a single projected-gradient step. The experimental section also omits the actual DDNM/DiffPIR comparisons used in the conclusions. These are not local presentation issues; they invalidate the paper's main claims as stated. A future version that reframes the contribution as a step-size schedule with careful comparisons against matched DDNM and DiffPIR baselines, and that verifies the orthogonality condition or avoids Eq. (19), could potentially be viable, but that would be a substantially revised manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the central conceptual claim doesn't survive the paper's own algebra. Under HH^T=I, Eq. 16 reduces to Eq. 19, a single projected gradient step with step size ρ. HQS is just rescaled GAP. So the “hybrid” of hard and soft constraints is a reparameterization of the step size, not a new mechanism. Any δ_t schedule can be reproduced by an appropriate ρ for DDNM.\n\nWhat's good: the three-stage decoupling of DDIM (denoising, data consistency, sampling) is cleanly stated and does make the connection to PnP explicit. That framing is not new—DiffPIR and DDNM already do this—but the paper lays it out clearly and uses it consistently. The algebra through Eq. 19 is correct. If the fused update is actually implemented with exact GAP and HQS (Algorithm 1 lines 4-6), then the orthogonality condition doesn't hold and Eq. 19 is not a valid description; the paper never says which one is used or whether the binary masks satisfy UU^T=I and VV^T=I. That's a load-bearing omission.\n\nThe experimental support is thin. Table 1 gives no error bars and only one CR; Table 2 has no baselines; there is no code or data. The abstract's claim that the method “consistently outperforms DDNM and DiffPIR” is contradicted by the fact that DDNM/DiffPIR entries are absent from the tables. The gains might be real, but as reported they are not supported. The step-size interpretation means the gains could just be tuning ρ.\n\nBottom line: the paper is a short empirical report with a clear mathematical structure but an overstated contribution. A careful referee could make it useful by forcing the authors to state which update is actually implemented, verify the orthogonality condition, and report a proper baseline comparison with error bars. I would send it to review rather than desk reject, but I would expect either rejection or a major revision. It's not a paper I'd cite for the hybrid claim; the step-size equivalence is the takeaway.","headline":"The fused GAP/HQS update collapses to a single projected gradient step under the paper's own orthogonality assumption, so the claimed hybrid mechanism is really a step-size knob; the three-stage DDIM decoupling is a useful framing, but the experiments are too thin to rescue it.","tokens_in":8800,"tokens_out":2254,"would_cite":false,"duration_ms":24956,"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 fusing GAP and HQS data-consistency updates inside DDIM sampling improves single-pixel compressive-sensing reconstruction, and that the fused update beats either update alone at 1–20% compression ratios.","keywords":["Diffusion models","Plug-and-play","Single-pixel imaging","Compressive sensing","Data consistency","Denoising diffusion implicit models","Inverse problems","Zero-shot image restoration"],"falsifier":"Run the same zero-shot pipeline with the fusion weight δ_t fixed to 0 and to 1 over all timesteps, and report PSNR, SSIM, LPIPS, and the measurement residual ‖y−Hx̂‖ at each compression ratio. The paper's claim that fusion beats either update alone fails if the fused schedule does not dominate both endpoints; the 'improves measurement consistency' language also needs qualification if the fused residual is not below the HQS endpoint.","tokens_in":7665,"feed_emoji":"📷","tokens_out":11770,"duration_ms":120066,"temperature":0.7,"pith_summary":"This paper tries to establish that a diffusion model and a plug-and-play (PnP) solver are two views of the same loop—denoise, enforce consistency with the measurements, and sample—and that the consistency step can be improved by fusing two PnP updates. The proposed fused update takes the diffusion model's denoised estimate and replaces it with a convex combination of a hard projection onto the measurements (GAP) and a soft regularized least-squares correction (HQS). The claim is that this correction makes reconstructions more measurement-consistent without derailing the diffusion sampling trajectory, and experiments on single-pixel imaging at 1–20% compression ratios show better PSNR, SSIM, and LPIPS than using either correction alone. A sympathetic reader would care because the fusion is a modular, zero-shot change to existing diffusion samplers: it needs no retraining and works with any linear forward model.","feed_headline":"Mixing two data-consistency updates in diffusion lifts image recovery","feed_subtitle":"Fusing hard and soft measurement updates inside diffusion sampling beats either update alone on single-pixel imaging","key_machinery":"The fused data-consistency operator in Equation 16 is the key mechanism: it takes the denoised estimate x_{0|t} from the diffusion model and replaces it with a weighted average of two plug-and-play corrections—GAP's hard pseudoinverse back-projection and HQS's regularized least-squares solve. The scalar δ_t controls the blend. When the sensing matrix is H=U⊗V with HH^T=I, the fused operator collapses to the single projected-gradient step x'_{0|t}=x_{0|t}+ρ U^T(y−Ux_{0|t}V^T)V, with ρ=1−λδ_t/(1+λ). This closed-form collapse is what makes the fusion cheap enough to run inside every DDIM iteration while preserving the paper's three-stage PnP–diffusion interpretation.","core_discovery":"The paper's central claim is that the diffusion sampling loop and PnP optimization are structurally the same iterated denoise-and-correct loop, and that the data-consistency step inside that loop can be strengthened by a linear fusion of two existing PnP update rules. Re-writing DDIM sampling as denoise (estimate x_{0|t}), correct (map x_{0|t} to a measurement-consistent x'_{0|t}), and resample (draw x_{t-1}), it inserts the fused correction x'_{0|t} = (1−δ_t)(x_{0|t}+H†(y−Hx_{0|t})) + δ_t(H^T H+λI)^{-1}(H^T y+λx_{0|t}), where the first term is the hard GAP back-projection and the second the soft HQS least-squares solve. Under the separable single-pixel sensing model H=U⊗V with HH^T=I, this","pith_inferences":["The paper's three-stage decoupling suggests that the same fused correction could be attached to other diffusion or score-based samplers that produce a clean estimate at each timestep, not just DDIM; the fusion itself is defined only through H, y, and the denoised estimate.","The paper notes that its fused method's metrics initially drop before recovering during sampling; this transient behavior points to a fixed δ_t schedule that is not adapted to the early diffusion steps, and adapting δ_t to the current measurement residual is a direct testable extension.","Because Equation 19 relies on HH^T=I, a practical implementation should either verify that the sensing masks are row-orthonormal or use the exact GAP/HQS terms from Algorithm 1; otherwise the claimed interpretation as a single projected-gradient step may not describe the actual computation."],"forward_implications":["The paper's decoupling gives a direct recipe: in any diffusion sampler, apply the PnP-style data-consistency correction to the denoised estimate x_{0|t} rather than to the noisy state x_t, then continue the ordinary sampling update.","Under the orthogonal separable sensing model, the fused update is a single projected-gradient descent step, so the added computation per diffusion iteration is small and the method remains zero-shot.","The method maintains reconstruction quality even at 1% compression (PSNR 21.23 dB in the paper), indicating it can operate with very few measurements.","The paper's identification of GAP with DDNM-style guidance and HQS with DiffPIR-style splitting implies that a positive result transfers to those two families of diffusion restoration methods."],"fun_headline_variants":["Fusing two data-consistency rules in diffusion lifts single-pixel imaging","Hard+soft data consistency in diffusion: better single-pixel reconstruction","Hybrid data-consistency in diffusion beats single updates on single-pixel","Diffusion + fused correction recovers sharper single-pixel images"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing assumption is that the sensing matrix has orthonormal rows (HH^T=I); the paper does not report whether the binary single-pixel masks used in the experiments satisfy it, and Equation 19—the compact projected-gradient reading of the fused update—depends on it.","fun_headline_variants_meta":{"raw":{"variants":["Fusing two data-consistency rules in diffusion lifts single-pixel imaging","Hard+soft data consistency in diffusion: better single-pixel reconstruction","Hybrid data-consistency in diffusion beats single updates on single-pixel","Diffusion + fused correction recovers sharper single-pixel images"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000259,"raw_usage":{"total_tokens":1425,"prompt_tokens":747,"completion_tokens":678,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":600}},"tokens_in":491,"tokens_out":678,"duration_ms":7593,"temperature":1.0,"reasoning_tokens":600,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T19:12:36.567338+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same zero-shot pipeline with the fusion weight δ_t fixed to 0 and to 1 over all timesteps, and report PSNR, SSIM, LPIPS, and the measurement residual ‖y−Hx̂‖ at each compression ratio. The paper's claim that fusion beats either update alone fails if the fused schedule does not dominate both endpoints; the 'improves measurement consistency' language also needs qualification if the fused residual is not below the HQS endpoint.","supporting_citations":[],"review_version":1}