{"id":"58a66460-f5e0-428f-9a31-c9482524569c","arxiv_id":"2607.15711","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"DDR-SR routes each real-world low-resolution image to one of two diffusion experts based on a high-frequency-loss difficulty score, using a low-compression VAE for hard images and a high-compression VAE for easy images, reporting improved quality-efficiency trade-offs.","lead":"Image super-resolution typically runs the same expensive diffusion model on every photo. DDR-SR tries to first measure how hard each image is, then route easy images to a fast model and hard images to a detailed model, promising a tunable speed-versus-quality trade-off.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Inference-time routing requires ground-truth HR image (Eq. 2, Algorithm 2); no learned difficulty predictor is described, so the reported benchmark results may rely on oracle routing.","rationale":"The reader identifies the same load-bearing weakness: the difficulty estimator requires ground truth, making inference-time routing impossible without an unstated learned surrogate. This is not a stylistic or incremental issue; it undermines the paper's central claim of adaptive, practical super-resolution. The reported results in Table 1 and Table 2 are therefore suspect because the routing mechanism that differentiates DDR-SR from a simple two-expert ensemble is not operational in deployment. The paper also lacks an ablation comparing DDR-SR with each expert alone, so the demonstrated gains cannot be disentangled from the oracle routing. I agree with the reader's high-confidence reject verdict; the missing inference-time estimator is the decisive concern.","tokens_in":15270,"tokens_out":2085,"duration_ms":19157,"concrete_test":"Obtain the authors' inference code or ask them to specify the exact difficulty estimator used at test time. Run DDR-SR on the RealSR benchmark without supplying HR references: if the code calls E_high(I_gt) or otherwise requires ground truth to compute the routing decision, the reported results are oracle-based. If a learned proxy f(I_deg) exists, compare the routing decisions made by f against the oracle decisions from Eq. (2) and re-evaluate Table 1 metrics; if the proxy's routing degrades SSIM/LPIPS beyond a small margin, the central efficiency-quality claim fails under realistic conditions.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's central claim is that DDR-SR dynamically routes inputs between two experts at inference based on a difficulty score. However, the difficulty estimator defined in Eq. (2) computes d = 1 - E_high(I_deg)/E_high(I_gt), which requires the ground-truth HR image I_gt. Algorithm 2, line 2, explicitly passes I_gt to the estimator: d <- E_diff(I_deg, I_gt). In real-world deployment, I_gt is unavailable, so this routing cannot be executed as described. The paper does not introduce any learned proxy that maps only the degraded image I_deg to a difficulty score; Section 3.2 only defines the oracle metric, and no architecture, training loss, or implementation is given for a practical estimator. Consequently, either the experimental numbers in Table 1 were obtained with oracle access to ground-truth routing—which would inflate performance and undermine the practical claim—or the authors used an undocumented approximation that is neither reproducible nor validated. Without a deployable difficulty estimator, the core contribution of 'automatic dynamic routing' is not realized, and the tunable quality-speed trade-off advertised in the conclusion is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DDR-SR, a difficulty-aware dynamic routing framework for diffusion-based real-world image super-resolution. It defines a difficulty score based on high-frequency spectral energy attenuation between the degraded image and its ground-truth counterpart, splits training data into easy and hard subsets, and trains two LoRA-adapted expert networks: Expert-D4 with a 4x-compression VAE and Expert-D8 with an 8x-compression VAE. At inference, inputs are routed to one of the two experts according to a threshold on the difficulty score. The method is evaluated on DIV2K-val, RealSR, and DRealSR against several multi-step and one-step diffusion baselines, and the paper claims superior quality and a tunable quality/efficiency trade-off.","tokens_in":1196,"tokens_out":1263,"duration_ms":33513,"significance":"If the method worked as described, the idea of adaptively choosing VAE compression according to per-image difficulty would be useful for the Real-ISR community: it is a plausible way to preserve fine details for hard inputs while saving computation on easy ones. The paper also provides real benchmark numbers and a reasonable experimental setup. However, as written, the central mechanism is not implementable because the routing decision requires the ground-truth image, the efficiency claim is based on an incomplete comparison, and no ablation isolates the effect of routing. None of these issues is merely cosmetic: each concerns the core contribution of the paper.","major_comments":[{"comment":"The routing difficulty score defined in Eq. (2), d = 1 - E_high(I_deg)/E_high(I_gt), requires the ground-truth image I_gt. Algorithm 2 explicitly writes 'd <- E_diff(I_deg, I_gt)'. In real-world deployment I_gt does not exist, yet the paper nowhere specifies, trains, or evaluates a learned estimator that maps only I_deg to d. Without such a proxy, the automatic dynamic routing at inference cannot be executed, and the Table 1 results must have been obtained either with oracle access to ground truth (which would inflate reported gains) or with an undocumented approximation. Since dynamic routing is the paper's stated core contribution, this is a load-bearing flaw. The authors need to provide a fully specified deployable difficulty estimator and re-run the benchmark evaluation without using I_gt at inference.","section":"Section 3.2 and Algorithm 2"},{"comment":"The efficiency comparison reports only Expert-D4 (1.81T FLOPs, 1.63B parameters) and omits the full DDR-SR system. The actual inference pipeline also includes the difficulty estimator/router and, under the default median split, routes roughly half the inputs to Expert-D8 and half to Expert-D4. The paper never reports the expected or measured FLOPs/latency of the complete DDR-SR model, nor the router's additional cost. Consequently, the Conclusion's claim of 'a tunable trade-off between quality and speed' is not supported by the efficiency data presented. The authors should report full DDR-SR complexity, including router overhead, for at least a few routing ratios.","section":"Section 4.2, Table 2"},{"comment":"The ablation study is limited to LoRA rank configurations for Expert-D4 and Expert-D8 (Tables 3 and 4). It does not compare DDR-SR against Expert-D4-only or Expert-D8-only, nor does it vary the routing threshold tau. Without these comparisons, the reported improvement cannot be attributed to dynamic routing; it could come from the higher-capacity Expert-D4 architecture or from the mixed training strategy. The paper needs ablations that isolate the routing decision, e.g., D4-only, D8-only, DDR-SR at tau=median, and DDR-SR at other routing ratios, with matched training budgets.","section":"Section 4.3"}],"minor_comments":[{"comment":"Text says 'DiffBIR, SeeSR, PASD, and DiffBIR exhibit better performance' — DiffBIR is listed twice; the intended set is unclear.","section":"Section 4.2"},{"comment":"In the qualitative discussion, 'DRR-SR' is a typo for 'DDR-SR'.","section":"Figure 4 / Section 4.2"},{"comment":"The text says Expert-D4 has '1.65B parameters', but Table 2 lists 1.63B. Please reconcile.","section":"Section 4.2 / Table 2"},{"comment":"The formula for z_SR is typeset ambiguously; it should be z_SR = (z - sqrt(1 - alpha_bar_t) * eps_8(z, c_t, t)) / sqrt(alpha_bar_t) or written as an explicit one-step denoising expression.","section":"Eq. (3)"}],"recommendation":"reject","confidential_remarks":"The oracle-routing issue is fundamental: the inference algorithm as written cannot be run without the ground-truth image, and no learned proxy is described. This goes beyond a missing experiment; it invalidates the core contribution as presented. The efficiency comparison and ablation gaps further weaken the paper. I would not encourage revision unless the authors can supply a genuine difficulty predictor and recompute the benchmark results without I_gt during inference."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core of this paper is a difficulty-aware router that sends easy inputs to an 8x-VAE expert and hard inputs to a 4x-VAE expert, with both experts trained via LoRA from a one-step diffusion SR model. The motivation is sound: Figure 2 shows that the performance gap between methods grows on hard samples, which justifies sample-adaptive allocation. The mixed-training scheme (80/20 splits) and the user-controllable threshold are also reasonable design choices, and Table 1 gives real numbers across three benchmarks. This is a legitimate extension of ClassSR-style routing to diffusion-based SR, not a breakthrough but a useful engineering contribution if it worked as claimed.\n\nIt does not work as claimed. Algorithm 2 and Eq. (2) define the difficulty score using the ground-truth image I_gt. At deployment I_gt doesn't exist. The paper never describes a learned proxy that maps only the degraded image to a difficulty score, nor any approximation. So the benchmark results in Table 1 were either produced with oracle access to ground truth—which would favor the router in a way that can't be replicated in practice—or used some undocumented trick that isn't reproducible. This is not a minor omission; it's the load-bearing component of the entire system.\n\nThe other soft spots are smaller but real. The ablation study only varies LoRA ranks; it never compares the full DDR-SR against Expert-D4 alone or Expert-D8 alone, so we don't learn whether routing is actually responsible for the gains. Table 2 reports only Expert-D4's FLOPs, omitting the difficulty estimator and router overhead, so the efficiency claim is incomplete. Hyperparameters (λ1, λ2, α, β, τ, LoRA ranks) are given without sensitivity analysis; the threshold τ especially matters because it controls the entire trade-off.\n\nWho is this for? Anyone working on efficient diffusion-based SR would find the routing idea interesting, and the easy/hard benchmark analysis is a useful observation. But no one can build on the system as described until the difficulty estimator becomes a deployable module. I'd send it to peer review anyway: the idea is viable, the flaw is fixable, and a serious referee would ask for a learned difficulty predictor, routing-vs-fixed-expert ablations, and a full efficiency accounting. That path could turn this into a decent paper. As it stands, I would not cite it for the core claim.","headline":"The routing idea is real and the easy/hard analysis is nice, but as written the inference-time router needs the ground-truth image, so the central claim doesn't hold until a learned difficulty predictor is added.","tokens_in":16105,"tokens_out":1529,"would_cite":false,"duration_ms":15153,"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":"Routing by restoration difficulty—hard images to a detail-preserving 4×-VAE diffusion expert, easy ones to a fast 8×-VAE expert—yields better perceptual quality at lower compute than one-size-fits-all diffusion super-resolution.","keywords":["real-world image super-resolution","difficulty-aware routing","dynamic routing","diffusion models","VAE compression ratio","high-frequency energy","one-step diffusion","low-rank adaptation"],"falsifier":"Compute d for all test images using Eq. (2) with ground truth, then train a neural network that takes only I_deg as input and tries to predict d. If the predicted routing (using that proxy's scores) yields quality or efficiency no better than always sending every image to Expert-D4 (or always to Expert-D8), the central claim that difficulty-aware routing helps would be refuted. A second, simpler check: inspect whether the authors' reported benchmark numbers were produced using the ground-truth-based oracle d, since if they were, the same numbers would not be reproducible at deployment time whe","tokens_in":15177,"feed_emoji":"🖼️","tokens_out":10108,"duration_ms":75476,"temperature":0.7,"pith_summary":"Real-world image super-resolution models currently process every input with the same network, even though some images are far harder to restore than others. DDR-SR proposes to route each input to one of two diffusion experts: a high-capacity model with a 4×-downsampling VAE that preserves fine details for hard images, and an efficient model with the standard 8×-downsampling VAE for easy images. Difficulty is measured as the loss of high-frequency Fourier energy between the degraded image and its ground-truth counterpart. The paper reports that this dynamic routing beats state-of-the-art diffusion super-resolution methods on perceptual metrics while cutting FLOPs, and that the routing threshold gives users a tunable quality/speed knob. A sympathetic reading takes the central claim to be that the VAE compression ratio, not just the denoising network, is the right place to modulate capacity in diffusion-based restoration.","feed_headline":"Difficulty-based routing beats one-size-fits-all super-resolution","feed_subtitle":"Two diffusion experts—one detail-preserving, one fast—hit higher perceptual quality with fewer FLOPs.","key_machinery":"The load-bearing mechanism is the difficulty-aware router. It computes, via the 2D Fourier transform and a high-frequency mask, the score d = 1 − E_high(I_deg)/E_high(I_gt), which falls in [0,1] and measures how much high-frequency spectral energy the degradation has removed. That score is compared to a user-settable threshold τ (default: the training set's median difficulty) to send the input to either Expert-D4, which uses a 4×-downsampling VAE and is LoRA-tuned on 80% hard plus 20% easy samples, or Expert-D8, which uses the standard 8×-downsampling VAE and is LoRA-tuned on the complementary mix. Both experts share the same diffusion UNet and one-step denoising, so the routing decision is","core_discovery":"The paper's core discovery is that a pair of expert diffusion networks differing only in their VAE's spatial compression ratio—4× versus 8×—can be trained with an 80/20 mixed-sampling strategy to produce complementary strengths, and a simple threshold on the high-frequency energy attenuation score can then assign each input to the right expert. On DIV2K-val, RealSR, and DRealSR, the routed system ('DDR-SR-1') reports the best SSIM, LPIPS, and CLIPIQA among compared one-step and multi-step diffusion methods, and the 4× expert alone runs at 1.81T FLOPs versus 2.27T for the strongest one-step baseline. The author's claim is that the rigid 8× VAE compression in standard latent diffusion irrevers","pith_inferences":["The difficulty estimator as defined requires the ground-truth HR image; to make DDR-SR deployable, the authors would need to train a proxy network that predicts d from the degraded image alone, and the reported gains would then depend on how well that proxy recovers the oracle ranking.","The global Fourier energy score may be a coarse difficulty signal: an image with a small, heavily degraded text region could receive a low difficulty score if most of its energy is elsewhere; a spatially localized or region-weighted difficulty predictor might improve routing accuracy.","The 4×/8× choice is a two-point discretization of a continuous capacity spectrum; interpolating the VAE compression ratio between 4 and 8, or routing among more than two experts, might smooth the quality cliff at the threshold and yield a finer trade-off.","Because the two experts share the UNet, the router could in principle be folded into a single model that selects the VAE at inference time, reducing duplication of the diffusion backbone."],"forward_implications":["Diffusion super-resolution no longer needs a single fixed VAE compression ratio; capacity can be matched to input difficulty, improving perceptual fidelity on hard samples.","Since easy inputs dominate many real-world collections, average computational cost per image can drop below what a single high-capacity model would need.","The mixed-training recipe (80/20 in opposing directions) keeps each expert robust to routing mistakes without sacrificing specialization.","The user-adjustable threshold turns the quality/efficiency trade-off into a deployable knob, so practitioners can select a point on the curve at runtime.","The routing logic is model-agnostic in principle and could attach to any latent-diffusion restoration system whose VAE compression ratio can be varied."],"fun_headline_variants":["Dynamic routing matches image difficulty to VAE capacity","Two diffusion experts beat one-size-fits-all super-resolution","Difficulty-aware routing cuts FLOPs while boosting perceptual quality","Adaptive VAE compression for efficient real-world super-resolution","Routing images to 4x or 8x VAE experts saves compute and sharpens detail"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The difficulty score d is computed from both the degraded image and its ground-truth high-resolution counterpart; in real-world deployment the ground truth does not exist, and the paper does not describe a learned network that estimates d from the degraded image alone, so the automatic routing cannot be executed as specified without oracle access.","fun_headline_variants_meta":{"raw":{"variants":["Dynamic routing matches image difficulty to VAE capacity","Two diffusion experts beat one-size-fits-all super-resolution","Difficulty-aware routing cuts FLOPs while boosting perceptual quality","Adaptive VAE compression for efficient real-world super-resolution","Routing images to 4x or 8x VAE experts saves compute and sharpens detail"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000159,"raw_usage":{"total_tokens":1084,"prompt_tokens":784,"completion_tokens":300,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":213}},"tokens_in":528,"tokens_out":300,"duration_ms":3089,"temperature":1.0,"reasoning_tokens":213,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T22:31:37.497797+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute d for all test images using Eq. (2) with ground truth, then train a neural network that takes only I_deg as input and tries to predict d. If the predicted routing (using that proxy's scores) yields quality or efficiency no better than always sending every image to Expert-D4 (or always to Expert-D8), the central claim that difficulty-aware routing helps would be refuted. A second, simpler check: inspect whether the authors' reported benchmark numbers were produced using the ground-truth-based oracle d, since if they were, the same numbers would not be reproducible at deployment time whe","supporting_citations":[],"review_version":1}