{"id":"8f21139d-305c-49bf-9302-9811cf4e5a50","arxiv_id":"2501.14014","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"An invertible neural network trained to mimic image degradations is used to steer a pretrained diffusion model in every sampling step, giving a blind and non-blind image restoration algorithm.","lead":"The paper presents INDIGO and BlindINDIGO, algorithms that restore corrupted images by pairing a pretrained AI image generator (a diffusion model) with a small invertible network that learns to mimic the corrupting process. The system guides the generator to produce images that match the observed blur, noise, or JPEG artifacts, yielding competitive results on face and natural image benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (12) does not by itself guarantee the coarse/detail separation that makes substituting y and inverting a valid posterior guide; a direct disentanglement test is needed before the central claim can be accepted.","rationale":"The reader's weakest-assumption analysis identified the same load-bearing point: the information-separation property asserted in Sec. III-A is not proven and is essential to the method. My stress-test agrees and proposes a concrete falsifiable check of that separation. The rest of the paper is a plausible engineering contribution: the non-blind INDIGO is an extension of prior work by the same authors, the blind variant adds a conditional INN with degradation embedding and test-time finetuning, and the experimental section includes ablations (Tables VII and VIII) and runtime comparisons. The main weaknesses are the unverified theoretical justification, missing code, and a few overclaims relative to the tables; none of these make the central construction internally inconsistent, but they do justify a conditional rather than accept verdict. Since the reader already set CONDITIONAL with moderate confidence, and my concern is the same one, the verdict should remain unchanged. The proposed swap test would provide the missing evidence: if it passes, the invertibility argument gains real support; if it fails, the method reduces to a heuristic projection whose success depends on incidental properties of the learned INN rather than on the stated structural guarantee.","tokens_in":20891,"tokens_out":8686,"duration_ms":81103,"concrete_test":"Train an INN with Eq. (12) on bicubic 4x super-resolution with FFHQ pairs. On held-out images x_i and x_j with low-resolution versions y_i and y_j, compute (c_i,d_i)=f_phi(x_i) and reconstruct xhat_{i,j}=f_phi^{-1}(y_j,d_i). If the Sec. III-A separation holds, xhat_{i,j} should be a natural HR image matching y_j and retaining the texture of x_i, with small ||H(xhat_{i,j}) - y_j|| under the true bicubic operator. If swapping y produces artifacts, ghosting, or large true-consistency error, the detail channel is not independently encoding the lost details, and the guidance in Algorithm 1 is not supported by the stated invertibility argument.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on the assertion in Sec. III-A that training f_phi with L(phi) = (1/N) sum_i ||f^c_phi(x_i) - y_i||^2 (Eq. 12) makes the coarse channel simulate the degradation and, 'due to invertibility, the other part will inevitably represent the detailed information lost.' Invertibility only guarantees that (c,d)=f_phi(x) is a bijective reparameterization of x; it does not guarantee that the detail channel d is semantically the lost high-frequency content, nor that f_phi^{-1}(y,d_t) lies on the natural-image manifold for the d_t produced by a diffusion estimate. Eq. (12) contains no term on d, so the decomposition is unconstrained beyond what is needed to match the training pairs. If the level sets of f^c are not aligned with the true posterior, replacing c_t with y can distort the sample or impose consistency with a surrogate degradation that differs from the true H. The blind variant (Algorithm 2) inherits this issue and adds test-time finetuning of phi against the current estimate x0,t, which can reinforce estimation errors rather than correct them. Figure 18 shows only that the coarse output tracks degradation level, not that the required detail-channel disentanglement holds.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two inference-time algorithms for image restoration with pretrained diffusion models. INDIGO trains a lifting-inspired invertible neural network so that the coarse output of its forward transform matches the degraded measurement; at sampling time it replaces the coarse channel of the intermediate estimate by the measurement and inverts the INN to produce a data-consistent estimate that guides the next reverse-diffusion step. BlindINDIGO extends this to unknown degradations with a conditional INN, a degradation-embedding extractor, an initialization from a pretrained restoration network, and test-time finetuning of the INN. Experiments report competitive or state-of-the-art numbers against ILVR, DDRM, DPS, PGDiff, DifFace, DR2, and StableSR on FFHQ, CelebA-HQ, ImageNet, DRealSR, and CelebChild, with ablations on step size, loss design, initialization, finetuning, sampling acceleration, and a consensus strategy.","tokens_in":21191,"tokens_out":4724,"duration_ms":46322,"significance":"If the central mechanism works, the contribution is practically significant: it removes the need for an analytic degradation model in non-blind restoration, treats nonlinear and real-world degradations through learned simulation, and leaves the pretrained diffusion model untouched. The INN is light (0.71M and 0.91M parameters), the inversion is exact by construction, and the experiments are broad, with component ablations and a DDIM variant that reduces runtime. The main conceptual risk is that the method's success depends on an information-separation property that is asserted rather than established. Since the paper is empirical, this can be addressed with targeted experiments, but the property is load-bearing for both INDIGO and BlindINDIGO.","major_comments":[{"comment":"The load-bearing claim that minimizing ||f^c_phi(x_i) - y_i||^2 forces the detail channel f^d_phi(x_i) to contain exactly the lost information is not justified by invertibility alone. Invertibility only makes (c,d)=f_phi(x) a bijective reparameterization; the loss contains no term on d, so training can allocate information between c and d arbitrarily as long as c matches y on the training pairs. Because the guidance step in Algorithm 1 replaces c_t by y and inverts with the diffusion-estimated d_t, an incorrect allocation could distort x0,t or impose consistency with a surrogate degradation different from the true H. The evidence in Fig. 18, showing that the coarse output tracks degradation level, does not test the required separation. Please add a direct disentanglement evaluation: measure the residual fidelity of f^{-1}(y, f^d(x)) on validation data, check whether d_t changes with y for fixed x, and report a final data-consistency metric such as ||H(f^{-1}(y,d_t)) - y||_2 for the restored images.","section":"Section III-A, Eq. (12), Algorithm 1"},{"comment":"The test-time finetuning update at the end of each iteration, phi <- phi - l * grad_phi ||c_t - y||^2, uses c_t = f^c_phi(x0,t, gamma_deg), where x0,t is itself produced by the previous guidance using the same phi. This self-referential loop has no convergence or stability guarantee and can reinforce estimation errors rather than correct the degradation model. The manuscript's own Fig. 17 shows that the pretrained INN can fail badly on complex real-world degradation, and the ablation in Table VII (cases 6 vs 7 and 9 vs 10) reports only aggregate improvement. Please add an analysis of the finetuning trajectory (e.g., loss and output quality over iterations), a check that phi does not drift arbitrarily far from the pretrained initialization, and a failure-mode discussion for the cases where finetuning does not help.","section":"Algorithm 2, Eq. (16), Table VII, Fig. 17"},{"comment":"The claims that INDIGO handles nonlinear JPEG-degradation and real-world DRealSR degradation are supported only by visual examples and no quantitative comparison or consistency metric. Since the non-blind contribution is specifically the ability to work without an analytic degradation model, the absence of any baseline (e.g., a diffusion method using a differentiable JPEG approximation, or a quantitative comparison against existing methods on DRealSR) leaves the practical advantage unquantified. Please add quantitative results for these settings and, where possible, compare with methods that can handle the same degradation or report a data-consistency measure.","section":"Section IV-A3 and IV-A4, Figs. 8 and 9"}],"minor_comments":[{"comment":"StableSR is cited as [1] in the table but as [4] in the text; the reference numbering should be corrected.","section":"Table IV"},{"comment":"The loss weights lambda_F and lambda_I and the definitions of LF and LI are introduced after Algorithm 2; define them before the algorithm is presented.","section":"Eq. (16) and Algorithm 2"},{"comment":"The caption says 'different conditions' but the experiment varies degradation level; the wording should match the actual setup.","section":"Fig. 18"},{"comment":"The consensus strategy triples the number of sampling trajectories, but the runtime cost is not reported for Table I; state clearly that 'Ours' in Table I uses the consensus strategy and report the corresponding runtime so the comparison is transparent.","section":"Appendix A and Table I"},{"comment":"No code is released; for reproducibility, release code and trained models or provide full training details for the INN (optimizer, epochs, batch size, data augmentation).","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper's central mechanism rests on an unproven information-separation property of the INN decomposition, and the blind finetuning loop is self-referential. I recommend a direct disentanglement experiment and a stability analysis of the finetuning update before acceptance. The paper is otherwise within scope and the experimental breadth is a strength."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new material here is the blind variant, BlindINDIGO: a conditional INN with a learned degradation embedding, test-time finetuning of the INN, SwinIR initialization, and DDIM acceleration. The non-blind core was already published by the same authors at MMSP 2023, and the paper is honest about that. The writing is clear, the method is simple to follow, and the ablations in Tables VII and VIII actually isolate the contribution of each component. That is more than many papers in this area do.\n\nThe central idea is to train the INN forward transform so its coarse output matches the degraded image, then at sampling time replace the coarse channel with the measurement y and invert. The invertibility makes the substitution exact in a mathematical sense, and the reported gains over ILVR, DDRM, and DPS on FFHQ 4x SR are plausible. The blind results on CelebA-HQ and ImageNet are competitive with PGDiff, DifFace, DR2, and StableSR, and the runtime comparison is useful.\n\nThe soft spot is the one the stress-test note flags: Eq. (12) constrains only the coarse channel. Invertibility guarantees that (c,d) is a bijective reparameterization of x, but it does not guarantee that d is semantically the lost detail, nor that f^{-1}(y,d_t) stays on the natural-image manifold when d_t comes from a diffusion estimate. The paper asserts this separation in Section III-A without proof. That is a real gap, but it is not fatal: the method works empirically on several benchmarks, and the assertion is testable. A direct disentanglement experiment (e.g., swapping detail channels between different images and checking that the inverse keeps the coarse content) would settle it. A second, more minor issue is that the non-linear degradation and real-degradation sections show only the authors' own results, with no baselines, which makes the 'competitive' claim hard to verify in those settings. There is also no code release, which is a practical barrier for a method that has several hyperparameters (zeta, lambda weights, starting timestep, finetuning rate).\n\nOverall this is a solid, well-organized empirical paper with an honest limitations section and a plausible new combination. The main theoretical concern is addressable and does not undermine the experimental findings as presented. I would send it to review; the referees should push for code and for the disentanglement test.\n\nRecommendation: engage with it, conditional on the authors providing code and a direct check of the information-separation assumption.","headline":"A plausible and well-tested extension of the authors' MMSP 2023 INDIGO to blind restoration, with the main open risk being the unproven coarse/detail disentanglement that the whole guidance step relies on.","tokens_in":21709,"tokens_out":1074,"would_cite":true,"duration_ms":11960,"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":"The paper proposes INDIGO and BlindINDIGO, which train a lightweight invertible neural network to simulate any degradation and then use its inverse, with the measurement substituted into the coarse channel, to guide pretrained diffusion…","keywords":["image restoration","blind image restoration","diffusion models","invertible neural networks","posterior sampling","degradation estimation","super-resolution","DDIM acceleration"],"falsifier":"Train the INN on a single synthetic degradation such as 4x bicubic downsampling with Eq. (12). For a set of test images, compute $(c,d) = f_\\phi(x)$ and also take a second degraded image $y'$ from a different clean image under the same degradation. If the information-separation claim is correct, $f_\\phi^{-1}(y', d)$ should look like the second image in coarse structure while retaining fine details from the first; if it instead shows blur, color bleed, or texture from the wrong image, the coarse/detail split has not isolated the degradation. Quantify this with PSNR and LPIPS against the ideal second image and against a no-guidance diffusion baseline.","tokens_in":20673,"feed_emoji":"🖼️","tokens_out":8780,"duration_ms":72294,"temperature":0.7,"pith_summary":"Most diffusion-based restoration methods need a closed-form expression of the degradation in the non-blind setting, or a fixed family of predefined degradations in the blind setting. This paper proposes INDIGO and BlindINDIGO to remove both restrictions: a small invertible neural network (INN) is trained, purely from paired images, so that its forward pass simulates whatever degradation produced the measurement. At sampling time the network's coarse channel is replaced by the actual measurement and the inverse transform yields an intermediate image that guides the pretrained diffusion model through a gradient step. The authors' central claim is that this yields competitive restorations on synthetic and real-world low-quality images, including non-linear degradations such as downsampling combined with JPEG compression, without retraining the diffusion model.","feed_headline":"Diffusion restoration works without knowing the degradation model","feed_subtitle":"A learned invertible network supplies data consistency, so no analytic degradation model is needed for blind restoration.","key_machinery":"The load-bearing object is a lifting-inspired invertible neural network, i.e., a wavelet-lifting construction in which the predict and update steps are neural networks, so that the mapping $[c,d] = f_\\phi(x)$ is exactly invertible. Its job is to separate each image into a coarse part that matches the degraded measurement and a detail part that holds the lost information. At inference the coarse part is replaced by the measurement $y$, the inverse $f_\\phi^{-1}(y, d_t)$ produces a data-consistent intermediate image, and the distance to the diffusion estimate $x_{0,t}$ is back-propagated into the current noisy sample $x_t$ via a gradient step. In the blind version the same blocks are conditioned on a degradation embedding $\\gamma_{\\rm deg}$ and finetuned during sampling.","core_discovery":"The central claim is that the perfect-reconstruction property of an invertible neural network can be turned into a degradation simulator and, at the same time, into an implicit inverse operator for guiding diffusion sampling. Writing the forward transform as $[c,d] = f_\\phi(x)$ and training it with $\\mathcal{L}(\\phi) = \\frac{1}{N}\\sum_i \\|f^c_\\phi(x_i) - y_i\\|^2$ makes the coarse output $c$ approximate the degraded measurement; the authors argue that, because the transform is invertible, the detail output $d$ then necessarily carries the information lost in degradation. At each reverse step they form $\\hat{x}_{0,t} = f_\\phi^{-1}(y, d_t)$ and update the sample by a gradient step on $\\|\\hat{x}_{0,t} - x_{0,t}\\|^2$. The blind variant conditions the INN on an implicit degradation embedding extracted by a pretrained estimator, refines the INN parameters during sampling against the observed measurement, and starts from a diffusion-initialized clean estimate to cut the number of timesteps.","pith_inferences":["An implication the authors leave implicit is that any inverse problem with paired data, including medical or scientific imaging where the forward model is non-linear and not analytically available, could use the same INN-as-degradation-simulator mechanism.","The information-separation assumption could be probed directly: on a known synthetic degradation, compare the detail channel $d$ with the actual lost high-frequency content and check whether inverting with a different measurement preserves the expected structure.","The consensus strategy, which averages guidance over parallel diffusion chains, suggests that variance reduction across chains is a cheap robustness lever at high noise; a theory of why three chains outperform two or four would be a natural follow-up.","A testable extension is to replace the fixed gradient scale $\\zeta$ with a noise-level-dependent schedule, since the paper's own ablation shows that too-small $\\zeta$ weakens data consistency while too-large $\\zeta$ amplifies noise."],"forward_implications":["Non-blind restoration no longer requires an analytic expression of the degradation operator: any degradation that can be represented by paired training data can be simulated by the INN.","Blind restoration can adapt to an unseen degradation at test time by finetuning the INN against the observed measurement during sampling.","The pretrained diffusion model is left unchanged, so its generative prior is preserved and only the lightweight INN (about 0.71M parameters for INDIGO, 0.91M for BlindINDIGO) is trained.","Combining a diffusion-initialized starting point with DDIM sampling reduces the required number of function evaluations from 1000 to 250 with only a modest perceptual-quality loss.","Because the guidance is a gradient step on the distance between the INN-refined image and the diffusion estimate, the procedure inherits the posterior-sampling interpretation of earlier gradient-based diffusion solvers."],"supporting_citations":[{"why":"Supplies the lifting-inspired invertible block architecture (LINN) that the INN is built from.","marker":"[37]"},{"why":"Provides the DDPM forward and reverse processes and the pretrained noise estimator that acts as the generative prior.","marker":"[8]"},{"why":"Supplies the gradient-guidance formulation for posterior sampling that INDIGO adapts by using the INN-refined image in place of a known forward model.","marker":"[22]"},{"why":"Introduces the initialization strategy with a pretrained restoration network that BlindINDIGO uses to start sampling from a diffused clean estimate.","marker":"[2]"},{"why":"Provides DDIM, used to accelerate inference by skipping reverse steps.","marker":"[9]"},{"why":"Provides the pretrained degradation estimation module (DEM) that yields the implicit degradation embedding $\\gamma_{\\rm deg}$.","marker":"[53]"},{"why":"Provides the SwinIR network used as the initialization prediction network $g_\\omega(y)$ in BlindINDIGO.","marker":"[46]"},{"why":"Is the earlier conference version of INDIGO for non-blind inverse problems that this paper extends to the blind setting.","marker":"[24]"}],"fun_headline_variants":["INN-guided diffusion eliminates the need for degradation models","Unified INN-guided diffusion for blind and non-blind restoration","Restoring images without knowing how they were degraded","Diffusion gets a hand from an invertible network for blind restoration","One network guides diffusion for both blind and non-blind restoration"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that forcing the INN's coarse channel to equal the degraded image automatically forces its detail channel to hold exactly the information lost in degradation, so that substituting the measurement and inverting yields a trustworthy intermediate restoration rather than a distorted or off-manifold image.","fun_headline_variants_meta":{"raw":{"variants":["INN-guided diffusion eliminates the need for degradation models","Unified INN-guided diffusion for blind and non-blind restoration","Restoring images without knowing how they were degraded","Diffusion gets a hand from an invertible network for blind restoration","One network guides diffusion for both blind and non-blind restoration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000587,"raw_usage":{"total_tokens":2786,"prompt_tokens":1005,"completion_tokens":1781,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":1700}},"tokens_in":621,"tokens_out":1781,"duration_ms":12575,"temperature":1.0,"reasoning_tokens":1700,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:28:42.848236+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the INN on a single synthetic degradation such as 4x bicubic downsampling with Eq. (12). For a set of test images, compute $(c,d) = f_\\phi(x)$ and also take a second degraded image $y'$ from a different clean image under the same degradation. If the information-separation claim is correct, $f_\\phi^{-1}(y', d)$ should look like the second image in coarse structure while retaining fine details from the first; if it instead shows blur, color bleed, or texture from the wrong image, the coarse/detail split has not isolated the degradation. Quantify this with PSNR and LPIPS against the ideal second image and against a no-guidance diffusion baseline.","supporting_citations":[{"cited_title":"WINNet: Wavelet-inspired invertible network for image denoising,","cited_arxiv_id":null,"evidence_quote":"Supplies the lifting-inspired invertible block architecture (LINN) that the INN is built from."},{"cited_title":"Diffusion posterior sampling for general noisy inverse problems,","cited_arxiv_id":null,"evidence_quote":"Supplies the gradient-guidance formulation for posterior sampling that INDIGO adapts by using the INN-refined image in place of a known forward model."},{"cited_title":"Swinir: Image restoration using swin transformer,","cited_arxiv_id":null,"evidence_quote":"Provides the SwinIR network used as the initialization prediction network $g_\\omega(y)$ in BlindINDIGO."},{"cited_title":"Indigo: An inn-guided proba- bilistic diffusion algorithm for inverse problems,","cited_arxiv_id":null,"evidence_quote":"Is the earlier conference version of INDIGO for non-blind inverse problems that this paper extends to the blind setting."}],"review_version":1}