{"id":"08474864-ddfa-4d52-99dd-f603a0cf6c83","arxiv_id":"2505.12935","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"LatentINDIGO guides latent diffusion sampling with wavelet-inspired invertible networks, achieving state-of-the-art blind image restoration without retraining the diffusion model.","lead":"The paper combines latent diffusion models with wavelet-inspired invertible neural networks to restore degraded photos without knowing the exact degradation. The method improves restoration quality on face and natural-image benchmarks, and a latent-space variant cuts runtime nearly in half.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-time refinement of the INN is forward-loss-only, shares parameters with the inverse, and is never quantitatively ablated; it nonetheless underpins the unknown-degradation claim.","rationale":"The reader identified the same load-bearing assumption: on-the-fly refinement of the INN parameters against the single observed measurement is unanalyzed and unablated. This is genuinely the weakest point in the argument. The rest of the framework has meaningful support: the INN is invertible by construction, the guidance losses are ablated, and the regularization study in Table IV shows a real effect. But the paper's headline distinguishing capability, handling unknown degradations without retraining, is specifically delegated to the refinement step, and that step is never tested in isolation. The correct response is not rejection, because the concern could be resolved by a simple ablation or by a quantitative unseen-degradation experiment; it is confirmation that the manuscript should be conditional on such evidence, exactly as the reader concluded. I therefore agree with the reader's weakest assumption and do not recommend changing the verdict.","tokens_in":21404,"tokens_out":3268,"duration_ms":37648,"concrete_test":"Disable refinement by freezing Θ in Algorithm 1 and Algorithm 2, i.e., remove line 11 and line 9, and rerun the medium-degradation CelebA-Test experiment, the WebPhoto/CelebChild no-reference experiments, and the unseen JPEG q=5 DIV2K experiment with quantitative PSNR/LPIPS. If the frozen-Θ runs match or beat the refined runs, the refinement mechanism is not load-bearing; if they are worse, report the gap and also plot ||x_{c,t} − y|| against LPIPS between x_inv and D(z0,t) across the refinement window to check whether forward fitting degrades the inverse reconstruction. Additionally, report the value of the refinement step size l.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central unknown-degradation claim rests on the per-image refinement update in Algorithm 1, line 11, and Algorithm 2, line 9: Θ ← Θ − l ∇_Θ ||x_{c,t} − y||² (or the latent analogue). This is the only mechanism proposed for adapting to degradations outside the INN training distribution, exemplified by the unseen JPEG q=5 result in Fig. 9. The implicit assumption is that a few gradient steps on the forward coarse-term error make g_Θ a better simulator of the true degradation, while the inverse g_Θ^{-1} remains a faithful reconstruction operator. But the parameters are shared between forward and inverse passes, so a forward-only gradient step can alter the inverse mapping in an uncontrolled way. Nothing prevents the optimizer from folding the noise in y into the detail component or changing the coarse/detail separation so that g_Θ^{-1}(y, x_d, γ) becomes worse. No convergence, step-count, or overfitting analysis is provided, and the step size l is not reported in the implementation details. Critically, no ablation isolates this mechanism: Table III ablates the guidance losses and regularization but never removes the refinement update, and the only claimed demonstration of refinement on an unseen degradation is a single qualitative example without quantitative PSNR/LPIPS. The reported synthetic gains on CelebA-Test could be driven entirely by the INN guidance and regularization, while the specific mechanism claimed to enable generalization to unknown degradations remains unvalidated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LatentINDIGO, a latent diffusion framework for blind image restoration that uses a wavelet-inspired invertible neural network (INN) as a learned degradation simulator and inverse reconstructor. Two variants are introduced: LatentINDIGO-PixelINN, which performs guidance in the pixel domain, and LatentINDIGO-LatentINN, which operates entirely in the latent space to avoid per-iteration decoding. At each sampling step the method alternates between updating the latent variable under INN guidance and refining the INN parameters on the observed measurement, followed by a re-encoding regularisation step intended to keep latents on the natural image manifold. Experiments on synthetic face restoration and real-world benchmarks report consistent gains over DiffBIR and several other baselines, and a single qualitative example on unseen JPEG degradation is used to illustrate the refinement mechanism.","tokens_in":21684,"tokens_out":4687,"duration_ms":45687,"significance":"The central idea of using an invertible lifting-scheme network to factor a clean image into a degraded coarse component and a detail component, then fusing LDM-sampled details with the observed measurement through the inverse transform, is elegant and the reported gains over DiffBIR are consistent across degradation levels. The LatentINN variant is a practically useful contribution because it avoids decoding at every iteration, and the paper includes careful ablations of the guidance losses and the regularisation schedule. The quantitative claims are, however, only as strong as the support for the per-image refinement mechanism, which is the component that addresses unknown degradations and is not directly ablated. If the refinement mechanism is validated with additional experiments, the paper would be a solid contribution to blind image restoration with latent diffusion models.","major_comments":[{"comment":"The on-the-fly refinement update Θ ← Θ − l∇Θ∥x_{c,t} − y∥² (also Algorithm 2 line 9 and Eq. (16)) is the only mechanism claimed to adapt the INN to unknown degradations, yet the paper never isolates it experimentally. Table III ablates ℓ_forw, ℓ_inv, and the regularisation, but the refinement update is present in every case, and the only out-of-distribution demonstration (Fig. 9, JPEG q=5) is a single qualitative example. Please add a quantitative ablation that disables the refinement update, report the learning rate l and the number of gradient steps per sampling iteration, and evaluate the effect of the first-half-only schedule. Without this, the 'handle unknown degradations' contribution is unsupported.","section":"Section III-B3, Algorithm 1 line 11"},{"comment":"The claim that the method generalises to degradations outside the INN training distribution rests entirely on one qualitative example on DIV2K with JPEG q=5. There are no PSNR/LPIPS numbers for this setting and no comparison against the same method with refinement disabled. Please provide quantitative results over a set of images for at least one unseen degradation, including a no-refinement control, to demonstrate that the per-image fitting does not overfit the single measurement.","section":"Section IV-B2, Fig. 9"},{"comment":"Several hyperparameters (α_forw, α_inv, α, and the regularisation schedule) are selected by evaluating on the test set (CelebA-HQ medium degradation), and all reported metrics are point estimates. This raises the risk that the comparisons in Tables I and II overstate the improvement. Please select hyperparameters on a validation set and report mean ± std over multiple independent restorations or seeds.","section":"Section IV-C1, Figs. 12 and 13, Table IV"}],"minor_comments":[{"comment":"There is a typo in the text after Eq. (5): 'ϵ is samped from N(0,I)' should read 'sampled'.","section":"Section II-A"},{"comment":"In Algorithm 3, line 11 the reverse step is computed from \\tilde{z}_{0,t}, but line 10 defines \\hat{z}_{0,t} = E(D(\\tilde{z}_{0,t})) and the surrounding text says the regularisation update is preserved. This is inconsistent with Algorithm 2, where z_{t-1} is computed from \\hat{z}_{0,t}. Please correct the algorithm or explicitly state that the regularisation step is skipped in the arbitrary-resolution variant and justify that choice.","section":"Section III-D, Algorithm 3"},{"comment":"The introduction describes two proposed approaches, but the experiments evaluate three variants (StableSR-PixelINN, DiffBIR-PixelINN, DiffBIR-LatentINN). Please clarify the relationship between the conceptual approaches and the evaluated variants.","section":"Section IV-B"},{"comment":"The text in Section III-B3 states that the refinement mechanism is applied only during the first half of the sampling process, but the pseudocode shows the refinement inside the full loop without such a condition. Please make the schedule explicit in the pseudocode.","section":"Algorithms 1 and 2"},{"comment":"The relationship to the authors' prior INDIGO+ work (Ref. [62]) is not discussed in the introduction or related work; the incremental novelty of the two approaches over [62] should be stated explicitly.","section":"Related Work / References"}],"recommendation":"major_revision","confidential_remarks":"The paper's relationship to the authors' own INDIGO+ (Ref. [62]) should be clarified in the revision, and the novelty overlap needs to be explicit. The selection of hyperparameters on the test set is a more serious concern and should be addressed in the main text; the reviewer should request a validation-based selection or an explicit statement about the lack of a separate validation set."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper extends the authors' earlier INDIGO+ pixel-domain framework into latent diffusion, and the core idea is sound: a wavelet-inspired invertible network whose forward transform mimics degradation and whose inverse merges the LDM's detail with the measurement. Two variants, PixelINN and LatentINN, both plug into DiffBIR/StableSR without retraining the LDM, and the latent variant runs about half the time. The quantitative story is consistent: DiffBIR-PixelINN gains up to 1.22 dB PSNR on CelebA-Test and better LPIPS, and both variants beat DiffBIR on no-reference metrics on real web photos. I also think the ablation in Table III and the guidance-strategy comparison in Fig. 11 are genuinely informative; they isolate the contribution of the forward/inverse losses and show the INN guidance beating PGDiff/LDPS/HGS on a shared baseline.\n\nThe soft spots are real but not fatal. The biggest is the test-time refinement mechanism (Algorithm 1, line 11; Algorithm 2, line 9). It is the only thing in the paper that specifically addresses \"unknown degradations,\" and it is a forward-loss-only gradient step on shared parameters, with no ablation that removes it, no convergence or overfitting analysis, and no report of the step size l. The single unseen-JPEG demonstration (q=5, Fig. 9) is qualitative only. The stress-test note is right: if that refinement collapses the detail channel or overfits to noise, the whole generalization story weakens. That is a fixable omission, but it is currently load-bearing.\n\nAlso, several hyperparameters (alpha_forw, alpha_inv, alpha, regularization schedule) are chosen by sweeping on the CelebA-HQ test set, and the main tables have no error bars. The paper says code will be released upon acceptance, but the experiments are not reproducible from the text alone.\n\nFor all that, the central contribution — a latent-space INN that guides LDM sampling with a stable, invertible data-consistency step — holds up for the degradations tested. This is a subfield improvement, not a paradigm shift, but it is an honest one. Worth a serious referee, and I would engage with it. If I were reviewing it, my first request would be an ablation that turns the refinement off, plus a quantitative JPEG q=5 result and either error bars or code.\n\nRecommendation: accept for peer review, expect revision.","headline":"A genuinely useful extension of INDIGO+ to latent diffusion, with consistent gains and a clever latent INN, but the test-time refinement that carries the unknown-degradation claim is never ablated.","tokens_in":22222,"tokens_out":2203,"would_cite":true,"duration_ms":22499,"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":"A wavelet-inspired invertible network lets latent diffusion models restore images with unknown degradations, without retraining, by alternating guided latent updates with on-the-fly refinement of the degradation model.","keywords":["image restoration","blind image restoration","latent diffusion models","invertible neural networks","wavelet transform","lifting scheme","diffusion posterior sampling"],"falsifier":"Train the INN only on synthetic Gaussian blur plus noise, then run both algorithm variants on a JPEG-only or motion-blur-only degradation with refinement enabled and disabled, and record the forward loss $\\|x_{c,t}-y\\|_2^2$ at each sampling step. If refinement does not reduce this loss relative to the frozen INN, or if disabling refinement already matches enabling it, the claim that refinement adapts the forward model to unknown degradations is falsified; equivalently, a degradation on which PSNR or LPIPS worsens as refinement steps increase would show the per-image fit overfitting to the measurement.","tokens_in":21161,"feed_emoji":"🖼️","tokens_out":9610,"duration_ms":89896,"temperature":0.7,"pith_summary":"The paper argues that blind image restoration, recovering a clean image from a degraded one without knowing the degradation, can be handled by latent diffusion models without retraining them, as long as sampling is steered by a wavelet-inspired invertible neural network. The INN's forward transform is trained to mimic the degradation, splitting a clean image into a coarse component and lost details, while the inverse transform merges the observed measurement with detail information from the current diffusion estimate to reconstruct the image. During sampling, the algorithm alternates between guiding the latent variable toward consistency with the measurement and refining the INN parameters on the fly so the forward model adapts to degradations it has never seen. If this works as claimed, existing latent diffusion pipelines could gain a plug-in restoration ability covering super-resolution, JPEG compression, and real-world web photos, offered in both pixel-domain and fully latent variants.","feed_headline":"One invertible network steers latent diffusion past unknown degradations","feed_subtitle":"Pixel and latent variants add up to 1.22 dB over DiffBIR on synthetic faces and lift real-photo quality scores, with no LDM fine-tuning","key_machinery":"The central object is the wavelet-inspired invertible neural network, whose predict and update modules are trainable networks arranged in the lifting scheme, making the whole transform exactly invertible by construction. In the forward direction the network factors an image into a coarse component that should match the degraded measurement and a detail component that holds the lost information; in the inverse direction it fuses the actual measurement with the detail component produced by the current diffusion estimate. Because forward and inverse parameters are shared, refining the INN against the observed measurement also improves reconstruction, which is the mechanism the paper relies on to handle unknown degradations. A secondary mechanism is the regularization step that re-encodes the guided latent through the autoencoder, pulling intermediate latents back toward the natural image manifold.","core_discovery":"The central claim is that an invertible neural network built on the lifting scheme of the wavelet transform can serve as a learned, trainable degradation simulator inside latent diffusion sampling, and that by alternating latent guidance with parameter refinement the pipeline can restore images under unknown degradations without fine-tuning the diffusion model. The paper proposes two instantiations: LatentINDIGO-PixelINN, where the INN acts in pixel space and guidance uses a forward data-consistency loss plus a perceptual backprojection loss, and LatentINDIGO-LatentINN, where the INN acts on VAE latents and guidance is a simple interpolation between the denoised latent and the INN-reconstructed latent. Both share a per-step regularization that re-encodes the guided latent to keep it near the natural image manifold, and both refine the INN parameters at inference by gradient descent on the squared error between the predicted coarse component and the observed measurement. The paper reports that DiffBIR-PixelINN improves PSNR by up to 1.22 dB over DiffBIR on synthetic face degradation with consistent LPIPS gains, that both variants improve no-reference quality scores on real web-photo benchmarks, and that the latent variant roughly halves the runtime of the pixel variant.","pith_inferences":["The paper leaves open whether the gain comes from invertibility itself or from the per-image refinement; an experiment that freezes the INN would separate the two.","A testable extension is to apply the refinement-only idea to non-invertible guidance such as DPS or LDPS; if those methods then match LatentINDIGO on unseen degradations, the invertible architecture is not the essential ingredient.","The observation that a simple latent-space blend works suggests that in latent space, data consistency can be enforced with a cheap coupling; the same design might transfer to other latent inverse problems such as compressed sensing or inpainting.","The regularization schedule implies a tuning knob: applying re-encoding only in early sampling steps worked best for faces, and the same schedule may need re-tuning per task and noise level."],"forward_implications":["Existing pretrained latent diffusion models can be turned into blind restoration solvers by adding a small INN guidance module, with no retraining or fine-tuning of the denoiser.","The pixel variant and the latent variant trade cost against texture fidelity: the latent variant roughly halves runtime by staying in latent space, while the pixel variant gives globally consistent reconstructions.","Because the INN is conditioned on a degradation embedding estimated from the measurement, one training recipe covers super-resolution, JPEG compression, and real-world low-quality photos.","On-the-fly refinement extends the framework to degradations outside the INN's synthetic training distribution, as shown by the JPEG quality-factor-5 experiment on DIV2K without retraining.","The patch-based extension supports arbitrary output resolutions by processing latent tiles while applying guidance on the whole latent, reducing memory overhead and tile-boundary artifacts."],"supporting_citations":[{"why":"Supplies the DiffBIR baseline and its pretrained LDM denoiser that both LatentINDIGO variants are tested on and compared against.","marker":"[1]"},{"why":"Defines the latent diffusion framework whose pretrained denoiser and VAE encoder/decoder the algorithm uses without retraining.","marker":"[3]"},{"why":"Provides the DPS gradient-based data-consistency update that the PixelINN forward-loss guidance extends to a learned forward model.","marker":"[17]"},{"why":"Introduces the LDPS/PSLD latent-space posterior sampling and the manifold regularization strategy that the proposed re-encoding step resembles.","marker":"[44]"},{"why":"Establishes the prior INN-guided probabilistic diffusion algorithm for image restoration that this work extends to latent diffusion and on-the-fly refinement.","marker":"[62]"},{"why":"Supplies the wavelet-inspired lifting-scheme INN construction whose predict/update modules are the backbone of PixelINN and LatentINN.","marker":"[64]"},{"why":"Defines the LPIPS perceptual feature extractor used in the inverse guidance loss and as an evaluation metric.","marker":"[67]"},{"why":"Provides the pretrained implicit degradation estimator that yields the conditioning vector for the INN.","marker":"[68]"},{"why":"Supplies the WebPhoto-Test and CelebChild real-world face datasets used to test generalization to unknown degradations.","marker":"[71]"}],"fun_headline_variants":["INN-guided latent diffusion handles unknown degradations","Wavelet-INN guides latent diffusion past unknown degradations","Pixel and latent INN variants improve restoration by up to 1.22 dB","No fine-tuning: INN steers latent diffusion for image restoration","LatentINDIGO: INN-guided diffusion for arbitrary degradations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a few gradient updates to the INN parameters against the squared error on the single observed image make the INN's forward transform faithfully simulate the true degradation, while the shared inverse transform still reconstructs the clean image; if this per-image fitting overfits to noise or collapses the detail channel, the guided sampling degrades even though the pretrained diffusion model is strong.","fun_headline_variants_meta":{"raw":{"variants":["INN-guided latent diffusion handles unknown degradations","Wavelet-INN guides latent diffusion past unknown degradations","Pixel and latent INN variants improve restoration by up to 1.22 dB","No fine-tuning: INN steers latent diffusion for image restoration","LatentINDIGO: INN-guided diffusion for arbitrary degradations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000374,"raw_usage":{"total_tokens":2063,"prompt_tokens":1079,"completion_tokens":984,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":695,"completion_tokens_details":{"reasoning_tokens":895}},"tokens_in":695,"tokens_out":984,"duration_ms":8855,"temperature":1.0,"reasoning_tokens":895,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:23:02.175536+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the INN only on synthetic Gaussian blur plus noise, then run both algorithm variants on a JPEG-only or motion-blur-only degradation with refinement enabled and disabled, and record the forward loss $\\|x_{c,t}-y\\|_2^2$ at each sampling step. If refinement does not reduce this loss relative to the frozen INN, or if disabling refinement already matches enabling it, the claim that refinement adapts the forward model to unknown degradations is falsified; equivalently, a degradation on which PSNR or LPIPS worsens as refinement steps increase would show the per-image fit overfitting to the measurement.","supporting_citations":[{"cited_title":"Solving linear inverse problems provably via posterior sampling with latent diffusion models,","cited_arxiv_id":null,"evidence_quote":"Introduces the LDPS/PSLD latent-space posterior sampling and the manifold regularization strategy that the proposed re-encoding step resembles."},{"cited_title":"Indigo+: A unified inn-guided probabilistic diffusion algorithm for blind and non-blind image restoration,","cited_arxiv_id":null,"evidence_quote":"Establishes the prior INN-guided probabilistic diffusion algorithm for image restoration that this work extends to latent diffusion and on-the-fly refinement."},{"cited_title":"WINNet: Wavelet-inspired invertible network for image denoising,","cited_arxiv_id":null,"evidence_quote":"Supplies the wavelet-inspired lifting-scheme INN construction whose predict/update modules are the backbone of PixelINN and LatentINN."},{"cited_title":"The unreasonable effectiveness of deep features as a perceptual metric,","cited_arxiv_id":null,"evidence_quote":"Defines the LPIPS perceptual feature extractor used in the inverse guidance loss and as an evaluation metric."},{"cited_title":"Knowledge distillation based degradation estimation for blind super-resolution,","cited_arxiv_id":null,"evidence_quote":"Provides the pretrained implicit degradation estimator that yields the conditioning vector for the INN."}],"review_version":1}