{"id":"55335215-602c-47a5-b8d1-764f923e9f10","arxiv_id":"2608.00537","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"HDPS alternates pixel-space Langevin refinement with latent decoder-inversion alignment inside flow matching, recovering high-frequency details that pure latent solvers miss.","lead":"This paper identifies why latent-space generative models struggle with image restoration: compressed autoencoders hide high-frequency errors from gradient updates. It proposes HDPS, which alternates pixel-space measurement corrections with latent-space semantic projection, and reports state-of-the-art results on five inverse problems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.1's claimed resolution term vanishes exactly when the latent gradient is zero, so the theory does not cover the blindness regime; the exact projection of a purely normal pixel correction returns to z0.","rationale":"The reader's weakest assumption was that early-stopped decoder inversion (N_z=15) may fail to preserve the orthogonal correction of the exact argmin. I agree that this is a serious gap. My stress-test goes further: even the exact-argmin theorem, as stated, does not resolve blindness when the latent gradient is zero, because the second-order term is proportional to g_parallel and vanishes at the stagnation point. This is directly testable on a toy decoder. The proof's additional claim that pure latent optimization has zero orthogonal correction 'at all orders' is mathematically false, since a finite latent step produces O(eta^2) curvature terms. These issues make the central theoretical claim substantially weaker than presented. However, the empirical results are extensive and the algorithm is plausible, so I would not reject the paper outright; it should be conditional on correcting the theorem, supplying the missing step sizes (eta_x, eta_z) and validation split, and reframing the 'resolution' claim to match what is actually proven. This is why I keep the verdict at CONDITIONAL rather than moving to ACCEPT or REJECT, but the revision required is substantive, not cosmetic.","tokens_in":19947,"tokens_out":17232,"duration_ms":225417,"concrete_test":"Analytical check: use the simplest rank-deficient decoder with curvature, D(z) = (z, z^2) at z0 = 0, so x0 = (0,0), J_D = (1,0)^T, and R(J_D) is the x-axis. Choose g = (0,1), so g_parallel = 0 and g_perp = g, exactly the blindness regime. For any eta_x > 0, solve zhat(eta_x) = argmin_z ||(0,-eta_x) - (z,z^2)||^2 exactly and compute <D(zhat)-x0, g_perp>. The expansion predicts 0 for all eta_x; this should be verified explicitly. If true, Theorem 4.1's resolution mechanism is inactive when the latent gradient vanishes, and the paper must either exhibit a nonzero g_parallel regime where the second-order term is actually large enough to matter, or drop the 'resolves manifold blindness' claim.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central theoretical claim is that HDPS resolves First-Order Manifold Blindness via the second-order term in Theorem 4.1. But the theorem's own expansion shows the term is proportional to v = -(J_D^T J_D)^\\dagger J_D^T g_\\parallel. In the stagnation regime that motivates the paper, the latent gradient vanishes: J_D^T g = 0, which implies g_\\parallel = 0 (for full-column-rank J_D). Then v = 0, so the claimed second-order correction in Eq. (10) is identically zero. The exact projection of xhat = x0 - eta_x g_\\perp onto the decoder manifold then has no O(eta_x) or O(eta_x^2) component: the expansion's optimality conditions give v = 0 and w = 0, so D(zhat) - x0 = O(eta_x^3). Thus the orthogonal correction is discarded precisely when the latent optimizer is blind. The proof's closing assertion that pure latent optimization has <Delta x, g_\\perp> = 0 'at all orders' is also incorrect: a finite latent step D(z0 + eta v) - x0 contains the curvature term (1/2) eta^2 v^T H_D v, whose projection onto g_\\perp is generally nonzero. So the contrast with latent-only optimization is overstated. The theorem may demonstrate a genuine second-order effect when g_\\parallel is nonzero, but it does not establish resolution in the regime it claims to fix, and the algorithm's early-stopped N_z=15 alignment (Section D.2's 'anchored compromise') is even further from the exact-argmin assumption. The reported gains must therefore be explained by something other than the stated mechanism, unless eta_x, eta_z, and the alignment trajectory are analyzed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper identifies a geometric limitation of latent-space inverse solvers, termed First-Order Manifold Blindness: because the decoder Jacobian is rank-deficient, pixel-space residuals orthogonal to its column space are invisible to first-order latent gradient updates. The authors propose Hybrid-Domain Posterior Sampling (HDPS), which alternates between pixel-space Langevin correction, optimization-based latent alignment by decoder inversion, and flow matching state updates. A theoretical result (Theorem 4.1) claims that this decoupled procedure resolves manifold blindness through a second-order curvature effect. Experiments are reported on FFHQ, AFHQ, and DIV2K at 768x768 resolution for inpainting, Gaussian/motion deblurring, and 12x super-resolution, with consistent large gains over several latent flow baselines when measured by PSNR/SSIM/LPIPS. Code is provided.","tokens_in":20413,"tokens_out":7193,"duration_ms":78109,"significance":"If the claimed mechanism is established, the decoupled framework is valuable: it offers a training-free way to inject pixel-space likelihood information into latent flow models, which is a timely problem. The paper's strengths include extensive benchmarking across five inverse problems and three datasets, public code, ablations of the main design choices, and a mostly correct local Taylor analysis. However, the central theoretical claim does not apply in the stagnation regime it sets out to fix, and the 'state-of-the-art' claim is stronger than Table 1 supports. The empirical results may still stand, but the theoretical narrative needs substantial correction or reframing.","major_comments":[{"comment":"The theorem's conclusion does not cover the regime that motivates it. The proof's first-order optimality condition gives v = -(J_D^T J_D)^\\dagger J_D^T g_parallel. In the stagnation regime described in Sec. 3.2, the latent gradient J_D^T g = 0; because g_perp is orthogonal to R(J_D), this implies J_D^T g_parallel = 0, hence g_parallel = 0 and v = 0. Then Eq. (10) gives <Δx, g_perp> = O(eta_x^3), not a second-order correction. Thus Theorem 4.1 demonstrates at most a second-order effect when g_parallel is nonzero, which is not the blindness case. Additionally, the Appendix A claim that pure latent optimization has <Δx, g_perp> = 0 'at all orders' is incorrect: a finite latent step z0 + eta v produces the curvature term (eta^2/2) v^T H_D v, whose projection onto g_perp is generally nonzero. The contrast with latent-only optimization is therefore overstated.","section":"Theorem 4.1 / Appendix A"},{"comment":"Theorem 4.1 assumes that zhat is the exact argmin of ||xhat - D(z)||^2, while Algorithm 1 and Eq. (9) perform only N_z = 15 gradient steps initialized at z_{0|t}. Appendix D.2 itself describes the finite alignment as 'an anchored compromise.' The early-stopped projection may not satisfy the first-order optimality condition used in the theorem, so the claimed second-order correction is not a rigorous explanation for the behavior of the implemented algorithm. The empirical gains in Table 1 could plausibly arise from the anchoring/regularization effect of early stopping rather than from the exact-argmin curvature mechanism. Please either analyze the finite-N_z case explicitly or present the theoretical result as a heuristic motivation rather than as the resolution of manifold blindness.","section":"Sec. 4.3, Eq. (9); Sec. 5.6; Appendix D.2"},{"comment":"The text states that HDPS 'consistently achieves state-of-the-art performance' and reports 'best or second-best scores across all metrics,' but Table 1 contains counterexamples. On FFHQ SR x12 (Avgpool), FlowDPS achieves PSNR 27.11 / SSIM 0.770 versus HDPS 26.99 / SSIM 0.719; on DIV2K SR x12 (Bicubic), FlowDPS has LPIPS 0.246 versus HDPS 0.250; on DIV2K SR x12 (Avgpool), FlowDPS has SSIM 0.525 versus HDPS 0.504. The overall superiority should therefore be qualified by task and metric. The broad 'state-of-the-art' claim in the abstract and Sec. 5.2 is too strong in its present form.","section":"Table 1 / Sec. 5.2"}],"minor_comments":[{"comment":"The schedule tau_t = sigma_t / sqrt(1 + sigma_t^2) is used but sigma_t is not explicitly defined as the same noise schedule appearing in Eq. (6) and Eq. (11). Please clarify the notation.","section":"Sec. 4.2, Eq. (8)"},{"comment":"Several numeric entries run together (e.g., '0.7700.158', '26.90 0.697 0.165'), making the table hard to read. Please fix the formatting.","section":"Table 1"},{"comment":"The labels g_pixel and g_latent are difficult to read at the printed size. Consider using separate panels with larger fonts.","section":"Figure 1(b)"},{"comment":"The generation anchoring step uses classifier-free guidance with a condition c, but the unconditional forward pass is not defined explicitly. Please give the exact formula used for v_theta(z_t, t, empty).","section":"Algorithm 1 / Sec. 4.1"},{"comment":"FlowDPS is reported with 'step size 15' in the baseline configuration. This seems inconsistent with the original FlowDPS settings and should be verified; also report the search range if it was tuned.","section":"Appendix B.2"}],"recommendation":"major_revision","confidential_remarks":"The paper's empirical contribution is potentially solid, but the theoretical narrative overreaches: Theorem 4.1 vanishes in the exact blindness regime, and the finite-N_z alignment used in practice is far from the theorem's exact-argmin assumption. The 'state-of-the-art' claim is also contradicted by some rows of Table 1. I would not reject on the experiments alone, but the manuscript needs a major revision to either prove a finite-step result or clearly reposition HDPS as a strong heuristic with empirical support, and to qualify the SOTA claim by task and metric."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper is worth a serious referee, but the main theoretical claim does not survive contact with its own equations. The algorithmic recipe is new and the experiments are strong. Proposition 3.1 is correct and cleanly explains why latent-only solvers over-smooth. The HDPS loop—pixel-space Langevin correction followed by early-stopped decoder inversion inside a latent flow reverse process—is genuinely different from prior decoupled or latent-only approaches, and the reported gains over FLAIR, FlowDPS, ReSample, and FlowChef are consistent across five tasks and three datasets at 768x768. The ablations support the design: pixel+alignment beats pixel+encoder and latent-only by large margins, and the low-resolution comparison against pixel-space DPS/DAPS is honest about backbone differences. Code is public.\n\nThe soft spot is Theorem 4.1. The stress-test note is right: at the stagnation point the latent gradient is J^T g = 0, and with a full-column-rank Jacobian that implies g_parallel = 0, so v = 0 and the second-order term in Eq. (10) vanishes. The theorem therefore does not prove resolution of the blindness it claims to fix. The assertion that pure latent optimization has zero orthogonal component 'at all orders' is also false, because the curvature term from a finite latent step has a nonzero projection onto g_perp. What likely makes the algorithm work is the early stopping in the alignment step (N_z = 15), which leaves the latent anchored near the flow prediction so the next decode retains part of the pixel correction. That is a plausible mechanism, but it is not the one Theorem 4.1 describes; Section D.2 hints at it without making it the main explanation.\n\nOther issues are minor but real: eta_x and eta_z are absent from the hyperparameter table, there are no error bars, and the operating point is tuned on the same benchmarks used for the final numbers. None of these sink the empirical contribution.\n\nVerdict: the paper deserves peer review. The theory needs a major revision or a retreat to a weaker statement, and the missing hyperparameters should be reported. With those changes it would be a solid, useful paper for anyone building latent flow or diffusion solvers for image restoration. I would cite it for the decoupling recipe and the experiments, not for the theorem.","headline":"Strong empirical decoupling recipe for latent inverse solvers, but Theorem 4.1's claimed 'resolution of blindness' vanishes exactly at the stagnation point it is supposed to fix.","tokens_in":20852,"tokens_out":7234,"would_cite":true,"duration_ms":80985,"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":"Latent flow inverse solvers lose high-frequency detail because the decoder Jacobian is rank-deficient; HDPS fixes this by alternating pixel-space Langevin correction with decoder-inversion latent alignment.","keywords":["inverse problems","latent flow matching","posterior sampling","first-order manifold blindness","decoder inversion","Langevin dynamics","image restoration","hybrid-domain inference"],"falsifier":"Measure the orthogonal component of the pixel gradient before and after the decoder-inversion loop: for a fixed measurement, compute g⊥ (the part of ∇_x loss orthogonal to the decoder Jacobian's column space) at the decoded anchor and again after N_z=15 alignment steps. If the post-alignment image has no greater g⊥ component than a latent-only update, or if direct encoder projection matches decoder inversion in PSNR, the claimed mechanism resolving manifold blindness is not operating.","tokens_in":19862,"feed_emoji":"🖼️","tokens_out":3301,"duration_ms":42114,"temperature":0.7,"pith_summary":"The paper argues that latent-space inverse problem solvers built on flow matching fail to recover high-frequency details because the decoder's Jacobian is severely rank-deficient: any measurement residual orthogonal to the decoder's tangent space is mathematically invisible to latent gradient updates. This 'first-order manifold blindness' is offered as the reason latent-only solvers stagnate and over-smooth. To bypass it, the paper proposes HDPS, which alternates between pixel-space Langevin steps that absorb precise measurement corrections and an optimization-based decoder inversion that projects the corrected image back onto the generative manifold. The paper claims this loop resolves manifold blindness through a second-order curvature effect and reports consistent gains over latent-only baselines, up to about 2.27 dB PSNR on motion deblurring. If correct, latent flow models can achieve high-fidelity restoration without retraining the decoder or abandoning the compressed prior.","feed_headline":"Pixel-space Langevin loop recovers detail latent solvers lose","feed_subtitle":"A hybrid sampler beats latent-only flow matching by up to 2.27 dB PSNR on deblurring and inpainting.","key_machinery":"First-Order Manifold Blindness: the decoder's Jacobian maps a low-dimensional latent space (about 2% of pixel degrees of freedom) into pixel space, so latent gradients only see the projected component of the pixel gradient. HDPS replaces the composite backpropagation through A∘D with a decoupled loop: pixel-space Langevin dynamics absorb the full-rank measurement gradient, then optimization-based latent alignment inverts the decoder to return the corrected image to the manifold. Theorem 4.1 shows this loop yields a second-order correction along the orthogonal residual by leveraging the decoder's Hessian, without ever computing it explicitly.","core_discovery":"The central claim is that the bottleneck in latent flow inverse solvers is geometric, not representational: even when the decoder can render the target image, first-order gradient updates on the latent code cannot reach the high-frequency residuals that lie in the orthogonal complement of the decoder Jacobian's column space. The paper proves that a decoupled procedure—Langevin dynamics in pixel space followed by projecting the corrected image back through decoder inversion—produces a nonzero update along those previously invisible directions, via the decoder's curvature, at second order in the pixel step size. The method, HDPS, is reported to outperform latent-only and baseline decoupled sol","pith_inferences":["The same decoupling likely applies to latent diffusion models, not just flow matching, since the first-order blindness argument depends only on the decoder Jacobian, not on the generative ODE.","The early-stopped decoder inversion acts as an implicit regularizer; its step count N_z may need to be scheduled with noise level rather than fixed, and the alignment residual could serve as a data-dependent stopping criterion.","A testable consequence is that the improvement over latent-only solvers should vanish for decoders that are locally linear, or when the residual happens to lie inside the Jacobian's column space.","Replacing direct encoding with decoder inversion should matter most when pixel-space Langevin artifacts are non-Gaussian; measuring the artifact distribution could predict when direct encoding fails."],"forward_implications":["Latent flow inverse solvers can recover high-frequency structure without retraining or altering the decoder, as long as measurement consistency is enforced in pixel space.","The pixel-space correction plus decoder-inversion pattern is a general template that can be layered onto any latent generative model with a fixed decoder.","Because the composite gradient through A∘D is replaced by two simpler specialized operations, the decoupled loop can be cheaper per step than composite backpropagation at matched latent iterations.","The method's ceiling is set by the decoder's capacity to represent the corrected image, so gains should be largest when the decoder is expressive but its Jacobian is rank-deficient.","The framework extends in principle to nonlinear differentiable forward operators by replacing the linear adjoint in the Langevin step."],"supporting_citations":[{"why":"Supplies the flow matching formulation that the latent generative prior is based on.","marker":"[22]"},{"why":"Supplies the pre-trained latent autoencoder whose rank-deficient decoder causes manifold blindness.","marker":"[29]"},{"why":"FlowDPS is the key latent-only flow posterior sampling baseline that HDPS compares against and outperforms.","marker":"[19]"},{"why":"FLAIR is the strongest latent-only baseline, with HDPS reporting gains over it in all tasks.","marker":"[11]"},{"why":"DAPS provides the annealed Langevin strategy that the pixel-space correction stage adapts.","marker":"[45]"},{"why":"ReSample is a latent diffusion inverse solver baseline that the method is compared with.","marker":"[32]"},{"why":"FlowChef is a rectified-flow inverse solver baseline used for comparison.","marker":"[27]"},{"why":"Stable Diffusion 3 is the pre-trained backbone used to instantiate all flow-based solvers in the experiments.","marker":"[12]"}],"fun_headline_variants":["Hybrid sampler sees around decoder's blind spot","Pixel-space Langevin recovers detail latent flow loses","HDPS: merging pixel-space and latent steps for sharper recovery","Latent flow meets Langevin: solving the geometric blind spot","Langevin loop adds high-frequency detail latent solvers miss"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The early-stopped decoder inversion (about 15 steps) must map the pixel-corrected image back onto the decoder manifold without erasing the high-frequency corrections the Langevin steps just added; the paper itself calls this an 'anchored compromise'.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid sampler sees around decoder's blind spot","Pixel-space Langevin recovers detail latent flow loses","HDPS: merging pixel-space and latent steps for sharper recovery","Latent flow meets Langevin: solving the geometric blind spot","Langevin loop adds high-frequency detail latent solvers miss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000259,"raw_usage":{"total_tokens":1428,"prompt_tokens":757,"completion_tokens":671,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":590}},"tokens_in":501,"tokens_out":671,"duration_ms":8310,"temperature":1.0,"reasoning_tokens":590,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T00:41:44.594641+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the orthogonal component of the pixel gradient before and after the decoder-inversion loop: for a fixed measurement, compute g⊥ (the part of ∇_x loss orthogonal to the decoder Jacobian's column space) at the decoded anchor and again after N_z=15 alignment steps. If the post-alignment image has no greater g⊥ component than a latent-only update, or if direct encoder projection matches decoder inversion in PSNR, the claimed mechanism resolving manifold blindness is not operating.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"FlowDPS is the key latent-only flow posterior sampling baseline that HDPS compares against and outperforms."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DAPS provides the annealed Langevin strategy that the pixel-space correction stage adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ReSample is a latent diffusion inverse solver baseline that the method is compared with."},{"cited_title":"Metaxas, and Yezhou Yang","cited_arxiv_id":null,"evidence_quote":"FlowChef is a rectified-flow inverse solver baseline used for comparison."}],"review_version":1}