{"id":"b27ded10-dc2b-4048-8c37-8c9181f4baa4","arxiv_id":"2506.16960","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Defusion restores degraded images with one model by conditioning a residual-space diffusion model on visual instructions built from degradations applied to standard test charts.","lead":"Defusion is a one-model image restorer that builds visual hints by applying degradations to standard test charts, then uses those hints to drive a diffusion model that removes the difference between a degraded photo and its clean version. If it holds up, the approach could let a single system handle blur, noise, rain, haze, snow, and JPEG artifacts without per-task retraining.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No inference-time mechanism is provided to build the visual instruction from an arbitrary input; all evaluations use known degradation categories, so the all-in-one and real-world generalization claims rest on an unspecified oracle protocol.","rationale":"The reader's weakest assumption identifies the same load-bearing gap: the paper provides no mechanism to construct the visual instruction for an arbitrary input at inference. My reading confirms and sharpens this: the construction in Sec. 3.1 requires knowing the degradation category and its parameters in order to apply it to a visual ground, and the supplement explicitly says the instruction is generated for the 'same category of degradation' as the image. Every reported evaluation supplies this category implicitly through the benchmark or dataset. The diffusion equations themselves are coherent, and the ablations give real support for the role of the visual instruction and degradation-space diffusion under known-category settings. But the paper's title and abstract promise all-in-one handling of unknown and real-world mixed degradations, and that promise is exactly where the missing component sits. Because the missing protocol is testable by code release and controlled blind experiments, I do not see grounds to reject the paper outright; the conditional acceptance is appropriate, with the missing inference protocol and oracle-free evaluation as conditions.","tokens_in":24598,"tokens_out":2951,"duration_ms":34764,"concrete_test":"Release the inference code path and run Defusion on a held-out blind mixed-degradation set (e.g., WED with unseen rain+snow+noise compositions) where the visual instruction v is produced automatically from x_LQ alone by a proposed estimator, with no dataset category labels. Report PSNR/SSIM against the oracle-category numbers in Table 3. If the automatic-instruction results are not reported, or are substantially below the oracle numbers, then the all-in-one and real-world generalization claims are not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Defusion is an all-in-one restorer that handles 'mixed or unknown degradations' and 'complex and real-world degradations.' The method's conditioning signal v is a visual instruction constructed by applying the degradation to a TE42 visual ground. But the paper never explains how v is obtained at test time from x_LQ alone. Sec. 3.1 only describes construction when the degradation is known; the supplement, Sec. 6, states: 'the visual ground is augmented with the same category of degradation as a visual instruction, either individually or as a mixed degradation category.' Thus the reported benchmarks (Tables 1-4) all use ground-truth or benchmark-implied degradation categories. For a real unknown or mixed degradation, no estimator, prompt, or protocol is given that would select the correct category or parameters, so the required v cannot be produced. Since Eqs. (10)-(12) condition the diffusion model on v, and the model is trained to 'do nothing' for mismatched instructions (Sec. 3.2), an incorrect or absent instruction would leave the LQ image unprocessed or restored incorrectly. The ablation in Table 5 shows that removing the instruction drops PSNR from 33.83 to 30.86, confirming that the instruction is load-bearing. Without a defined inference-time instruction construction mechanism, Defusion is an oracle-guided restorer on known categories rather than an all-in-one solution for unknown degradations. This is not an internal inconsistency in the diffusion formulation, but it is a missing component that the central generalization claim depends on.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Defusion, an all-in-one image restoration framework that conditions a diffusion model on \"visual instructions\": standardized TE42 charts that are degraded by the same corruption process as the input low-quality image. The instructions are tokenized by a VQ-GAN encoder and injected into a U-Net via a cross-attention adapter; the diffusion model operates in a \"degradation space\" defined as the difference between LQ and HQ images. The authors report state-of-the-art results on eight restoration tasks, on real-world datasets, on a synthesized mixed-distortion dataset, and on underwater images, and they claim generalization to unknown and composite degradations.","tokens_in":24958,"tokens_out":5683,"duration_ms":56554,"significance":"The degradation-space diffusion formulation (Eqs. 10-12) is coherent and the score-matching objective in Eq. (11) is standard. The idea of grounding degradations on standardized charts is interesting and could be a useful conditioning mechanism for multi-task restoration. The paper also provides extensive experiments and supplementary material, including perceptual metrics. However, the central claim—that Defusion is an all-in-one restorer for unknown, mixed, or real-world degradations—is not currently supported because the visual instruction at inference is constructed using the same degradation operator that generated the test LQ image. The ablation in Table 5 shows that the instruction is load-bearing (PSNR drops from 33.83 to 30.86 without it), so without a defined inference-time instruction-construction mechanism, the method is effectively an oracle-guided restorer on known degradation categories. The significance is therefore contingent on resolving this gap.","major_comments":[{"comment":"The paper does not provide any mechanism to construct the visual instruction v for an arbitrary input image at inference. The evaluation protocol uses the known test degradation category to build v; Supplement Sec. 6 states that \"the visual ground is augmented with the same category of degradation as a visual instruction.\" This gives the model an oracle hint about the degradation type and parameters. Since v is the only task-selection signal (Table 5 shows a drop from 33.83 to 30.86 PSNR without it), the reported results do not support the abstract's claims of handling \"unknown\" or \"real-world\" degradations. The authors must either provide a practical estimator for v from the LQ image alone, or explicitly reframe the method as requiring a user-supplied instruction and adjust the claims accordingly.","section":"Sec. 3.1 and Supplement Sec. 6"},{"comment":"For the mixed-distortion WED dataset, the visual instruction must be constructed from the known composition order of rain/snow/noise, and for the underwater datasets EUVP and TURBID no instruction-construction protocol is described at all. The underwater images contain a degradation that is not one of the seven training categories, yet Table 4 reports results without explaining how v is obtained. Without this information, the generalization claims in the abstract and conclusion are not supported by the experiments.","section":"Sec. 4.2, Tables 3 and 4"},{"comment":"The property that the model \"does nothing when v does not match the actual degradation\" is asserted but the training objective in Eq. (11) does not include any term for mismatched instructions. There is no description of how the model is trained to output y0 = 0 (i.e., T_v(xLQ) = xLQ) for a wrong v. This is important because any practical inference-time estimator of v will be imperfect; the manuscript needs to specify how mismatched instructions are handled during training and what the expected behavior is at inference.","section":"Sec. 3.2"}],"minor_comments":[{"comment":"The JPEG artifact removal (LIVE1) rows for AirNet, TransWeather, and WGWS-Net contain the same numbers as those methods' dehazing rows (e.g., AirNet 15.32/0.562, TransWeather 15.58/0.569, WGWS-Net 15.50/0.535), which appears to be a copy-paste error. Please verify and correct these entries.","section":"Table 1"},{"comment":"The expectation subscript in Eq. (11) is written as x_t, but the variable being diffused is y_t; the subscript should be y_t to be consistent with the conditioning variables.","section":"Eq. (11)"},{"comment":"The main text states that the model is trained for 100 epochs, while the supplement states 600K iterations; please clarify the relationship between these two statements.","section":"Sec. 4.1 and Supplement Sec. 8"},{"comment":"The figure caption refers to a loss L_inst, which is defined in Eq. (2) but stylized differently in the figure; please unify the notation.","section":"Fig. 2 and Sec. 3.1"},{"comment":"The supplement says the method takes \"the residual between the visual ground and its degraded version,\" while Sec. 3.1 describes subtracting the clean visual ground embeddings from those of the degraded instruction; please clarify whether the residual is computed in image space or in the embedding space, and how this residual is injected into the diffusion model.","section":"Supplement Sec. 6"}],"recommendation":"major_revision","confidential_remarks":"The central problem is the oracle nature of the visual instruction at inference. The supplement's explicit statement that the visual ground is augmented with the same degradation category as the test image makes the circularity clear, and the ablation confirms that the instruction is the main driver of performance. This is not a cosmetic flaw; it undermines the all-in-one and unknown-degradation claims. A major revision would need to add a practical inference-time instruction-construction module or a protocol for estimating the degradation from the LQ image, and then rerun the benchmarks with that protocol, ideally also comparing against a version without oracle hints. The degradation-space diffusion idea itself is worth preserving."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Defusion is worth a serious look: it proposes a concrete, novel conditioning signal—degradations applied to standardized TE42 charts—and combines it with a diffusion model that operates in the residual degradation space. The formulation in Eqs. (10)-(12) is coherent, the score-matching objective is standard, and the ablation in Table 5 shows the visual instruction is doing real work (removing it costs ~3 dB PSNR). The experimental coverage is broad: eight synthetic tasks plus real-world, mixed-distortion, and underwater sets, with both distortion and perceptual metrics reported. That is a solid contribution to the all-in-one restoration line, not a paradigm shift, but a genuine step.\n\nThe soft spots are real, and the biggest is the one the stress test flags: there is no inference-time mechanism to construct the visual instruction from an arbitrary LQ image. The supplement (Sec. 6) says the visual ground is augmented with 'the same category of degradation' as the test image, which on benchmarks is an oracle hint about the degradation family. For a genuinely unknown or mixed real-world degradation, the paper gives no estimator, prompt, or protocol that would produce the needed instruction. Since the model is trained to 'do nothing' when the instruction mismatches (Sec. 3.2) and the ablation confirms the instruction is load-bearing, the abstract's 'unknown degradations' and 'real-world generalizability' claims are not currently supported. This is not an internal inconsistency in the diffusion math, but it is a missing component that the central generalization claim depends on.\n\nOther issues are smaller but worth fixing. The paper cites InstructIR, AutoDIR, and DiffBIR in related work but never compares against them in the tables, which weakens the 'better than text-based priors' argument. And Table 1 has an obvious error: the Artifact removal (LIVE1) rows for AirNet, TransWeather, and WGWS-Net are identical to their Dehazing (Dense-Haze) entries. That is sloppy and should be corrected. These are all addressable with code release, a specified degradation-to-instruction pipeline (even a simple classifier or user-provided instruction), and a cleaned-up comparison table.\n\nBottom line: the core idea is promising and the diffusion formulation is sound, but the paper currently overclaims. It deserves a serious referee and, after major revision, could be a solid contribution. I would take it to a reading group to debate the conditioning scheme, and I would cite the visual-grounding idea if the authors fix the inference gap.","headline":"A genuinely new visual-conditioning idea for all-in-one restoration, with a sound degradation-space diffusion core, but the paper's headline claims about unknown degradations rest on an unspecified oracle protocol and need revision.","tokens_in":25460,"tokens_out":2310,"would_cite":true,"duration_ms":24067,"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":"Defusion is a single diffusion model, guided by visual instructions built from degraded test charts, that operates on the residual between a low-quality image and its restored version; the paper reports it beats prior unified models on…","keywords":["all-in-one image restoration","visual instruction","degradation space diffusion","diffusion model","image deblurring","image dehazing","image deraining","image denoising"],"falsifier":"Give Defusion an image whose degradation is not in the training families, such as a real photo combining motion blur, low light, and JPEG compression, run inference without supplying a visual instruction, and compare PSNR and LPIPS against task-specific restorers; if the gap is large or the output is not competitive, the all-in-one generalization claim collapses. As a second check, feed a deliberately mismatched instruction (a 'haze' chart on a blurry image) and verify the paper's claim that the model 'does nothing'—if it partially restores anyway, the controllability claim is falsified.","tokens_in":24419,"feed_emoji":"🖼️","tokens_out":8970,"duration_ms":81057,"temperature":0.7,"pith_summary":"The paper proposes Defusion, a single image-restoration model meant to handle many degradations at once: motion and defocus blur, haze, rain, raindrops, snow, noise, and JPEG artifacts. Its key move is to replace text prompts or learned task embeddings with a visual instruction: the same degradation applied to a standardized test chart built from textures, colors, and natural objects, so the model sees what the corruption looks like rather than reading a label. The diffusion process runs in 'degradation space', denoising the residual between the degraded image and its restored version instead of generating the image from scratch. On the reported benchmarks, the one model beats prior unified methods on every task and beats task-specific methods on several, including real-world and composite-degradation datasets. If the approach holds, a single restorer controlled by a visual description of its degradation is a practical step toward unknown and mixed degradations.","feed_headline":"One model beats task-specific restorers on 8 tasks","feed_subtitle":"A single model handles blur, rain, haze, snow, noise and artifacts by diffusing in degradation space.","key_machinery":"The load-bearing object is the degradation space $\\mathbb{D}_v$, defined as the set of differences $y_0 = x_{\\mathrm{LQ}} - T_v(x_{\\mathrm{LQ}})$ between a low-quality image and the restoration induced by a visual instruction $v$. Defusion adds Gaussian noise to this residual, trains a score model conditioned on $x_{\\mathrm{LQ}}$ and $v$ to reverse that process, and returns $x_{\\mathrm{LQ}} - \\hat{y}_0$ as the restored image. The visual instruction is built by applying the degradation to a TE42-based visual ground (randomly assembled regular textures, random textures, standard colors, and natural objects), tokenized by a vector-quantized autoencoder, and injected through copied cross-attention layers (VIA); low-quality structure enters through an Image Restoration Bridge (IRB) that injects hierarchical features into each U-Net block via AdaLN-Zero.","core_discovery":"Defusion's central claim is that guiding a diffusion model with a visual instruction—a standardized visual ground corrupted by the same degradation as the input image—and diffusing in the space of residuals between low-quality and high-quality images produces an all-in-one restorer that outperforms prior unified models across motion deblurring, defocus deblurring, desnowing, dehazing, raindrop removal, deraining, real denoising, and JPEG artifact removal, and that as a single model it even surpasses task-specific methods on several datasets. The paper further claims that the degradation-space formulation is more stable to train than image-space diffusion, that it imposes no special noise-schedule requirements and can therefore be adapted from any pretrained diffusion U-Net, and that it is controllable: when the visual instruction does not match the actual degradation, the model does nothing.","pith_inferences":["I infer the evaluation protocol is oracle-guided: the supplement says the visual ground is augmented with 'the same category of degradation' as the input, so for an arbitrary image whose degradation is unknown the paper gives no rule for constructing the instruction; adding a learned instruction estimator is the natural step that would make the all-in-one claim fully blind.","I infer the visual-ground pool could double as a calibration set: an unlabelled LQ image could be compared against instructions built from each candidate degradation, letting the model select its own instruction at inference and turning the 'does nothing when mismatched' property into an active selector.","I infer the estimated residual $\\hat{y}_0$ is itself a degradation map; if Defusion is right, that map could be reused for degradation segmentation, blind quality assessment, or synthesizing paired training data."],"forward_implications":["One trained model covers all eight tested tasks, and in the reported tables Defusion tops every unified baseline, beating task-specific models on several datasets as well.","The same model transfers to real-world and composite settings—NH-HAZE, LHP, RealSnow, RealBlur-J, a six-order rain–snow–noise mix, and underwater images—without task-specific fine-tuning.","Because the diffusion target is the LQ–HQ residual, the model preserves the low-quality structure and can be adapted from any pretrained diffusion U-Net without changing the noise schedule; the authors report stable training and fast DDIM sampling with four timesteps.","Diffusion in degradation space improves perceptual metrics (FID and LPIPS) relative to prior all-in-one diffusion restorers, as the generative prior from the pretrained U-Net is retained."],"supporting_citations":[{"why":"Defines the TE42 chart family whose textures, colors, and natural objects compose the visual grounds used to ground degradations.","marker":"[26]"},{"why":"Provides the VQ-GAN implementation and pretrained weights that the degradation tokenizer is adapted from.","marker":"[27]"},{"why":"Supplies the pretrained latent-diffusion U-Net that Defusion fine-tunes and injects conditions into.","marker":"[97]"},{"why":"Describes the IP-Adapter cross-attention injection that the Visual Instruction Adapter is built on.","marker":"[129]"},{"why":"Supplies the AdaLN-Zero mechanism that the Image Restoration Bridge uses to inject low-quality features into U-Net blocks.","marker":"[85]"},{"why":"Provides the protocol for mixing the individual restoration datasets into the all-in-one training set.","marker":"[137]"},{"why":"Is the leading unified baseline that Defusion is compared with and outperforms on the reported metrics.","marker":"[4]"}],"fun_headline_variants":["Visual-instruction diffusion restores any image degradation in one model","One diffusion model, guided by visual patterns, beats task-specific restorers","All-in-one restorer: visual instructions guide degradation diffusion","Visual-guided degradation diffusion restores all image types in one model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"At test time the method assumes you already know which degradation family the input image contains, because the visual instruction is built by applying that same family to a standard test chart; the paper gives no estimator that produces the instruction from an arbitrary unlabelled image.","fun_headline_variants_meta":{"raw":{"variants":["Visual-instruction diffusion restores any image degradation in one model","One diffusion model, guided by visual patterns, beats task-specific restorers","All-in-one restorer: visual instructions guide degradation diffusion","Visual-guided degradation diffusion restores all image types in one model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000816,"raw_usage":{"total_tokens":3545,"prompt_tokens":884,"completion_tokens":2661,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":2600}},"tokens_in":500,"tokens_out":2661,"duration_ms":17321,"temperature":1.0,"reasoning_tokens":2600,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:15:33.440234+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Give Defusion an image whose degradation is not in the training families, such as a real photo combining motion blur, low light, and JPEG compression, run inference without supplying a visual instruction, and compare PSNR and LPIPS against task-specific restorers; if the gap is large or the output is not competitive, the all-in-one generalization claim collapses. As a second check, feed a deliberately mismatched instruction (a 'haze' chart on a blurry image) and verify the paper's claim that the model 'does nothing'—if it partially restores anyway, the controllability claim is falsified.","supporting_citations":[{"cited_title":"Ingredient-oriented multi- degradation learning for image restoration","cited_arxiv_id":null,"evidence_quote":"Provides the protocol for mixing the individual restoration datasets into the all-in-one training set."}],"review_version":2}