{"id":"4a08582f-76b5-4993-86d3-ac9584d57864","arxiv_id":"2505.00687","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"GuideSR pairs a full-resolution guidance branch with a one-step latent diffusion branch and improves PSNR, SSIM, LPIPS, DISTS, and FID on DIV2K-Val, RealSR, and DRealSR.","lead":"This one-step diffusion super-resolution model feeds full-resolution structural features into a latent diffusion branch, reporting better fidelity metrics than seven diffusion SR baselines on three benchmarks. It could make high-quality, single-step upscaling practical for real-world images, but no code, checkpoints, or error bars are provided for independent verification.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The ablation does not isolate the full-resolution guidance mechanism; the missing control rows and the negligible 0.02 dB marginal gain over the long-skip baseline leave the central fidelity claim unverified.","rationale":"The reader's weakest assumption is that the full-resolution pixel-unshuffle plus concatenation mechanism's advantage over VAE-style conditioning is asserted but not measured in isolation. My stress-test identifies a specific, load-bearing support for that concern: the Table 2 ablation is missing exactly the control rows needed to attribute the reported PSNR gains to the stated mechanism. The marginal gain of Guidance without IGN over the long-skip baseline is only 0.02 dB, which strongly suggests that the mechanism's contribution is not established. This does not contradict the paper internally, and it does not warrant REJECT because the reported numbers could be correct; it does warrant keeping the CONDITIONAL verdict and requiring the missing ablation and a reproducibility release before the central claim can be accepted. The concrete check proposed would settle whether the full-resolution feature path is actually load-bearing by comparing it against an equally informative, architecture-matched control. No ad hominem is involved, and the concern is about experimental isolation rather than consensus disagreement.","tokens_in":13176,"tokens_out":8149,"duration_ms":87228,"concrete_test":"Retrain on the same RealSR training set and budget with a control that replaces the Guidance Branch features F'_r in Eq. 3 with a pixel-unshuffled copy of the input image, injected at the same UNet concatenation points and with a matching parameter count, while keeping the long-skip and IGN fixed. Also run the missing ablation row 'Baseline + Long-skip + Guidance (without IGN)' on RealSR. If the control reaches within 0.1 dB of the full GuideSR, or if the no-IGN guidance row is close to 26.80 dB, then the full-resolution pixel-unshuffle mechanism is not the source of the reported fidelity gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that full-resolution guidance features, pixel-unshuffled and concatenated into the UNet (Eq. 3, Fig. 2), preserve high-frequency structure that VAE downsampling loses and thereby drive the reported fidelity gains. This mechanism is never measured in isolation. In the Table 2 ablation on RealSR, the rows are: Baseline 26.65, Baseline + Long-skip 26.80, Baseline + Guidance 26.82, and Baseline + Guidance + IGN + Long-skip 27.08. There is no row 'Baseline + Long-skip + Guidance (without IGN)', no row 'Baseline + Long-skip + IGN', and no control that injects an equally informative low-level signal through the same concatenation points, such as a pixel-unshuffled copy of the input image or a VAE-encoded latent of the same full-resolution input. The marginal gain of the Guidance Branch over the long-skip baseline without IGN is only 0.02 dB (26.82 vs 26.80), and the full model's 0.28 dB gain over the long-skip baseline is attributed jointly to Guidance + IGN + Long-skip. Consequently, the 1.39 dB gain on DRealSR cannot be assigned to the proposed full-resolution guidance mechanism; it may come from the IGN, added parameters, the long-skip, or other training details. The description is also underspecified: the downsampling factor s in Eq. 3 and the exact UNet injection points and channel dimensions are not given, so the architecture is not reproducible from the text. The absence of a comparison to ClearSR (cited as [39]), which is the closest published method using low-resolution latent embeddings for diffusion SR, further prevents external calibration of the claimed gains.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"GuideSR is a single-step diffusion-based super-resolution model with a dual-branch design. The Guidance Branch processes the degraded input at full resolution using Full Resolution Blocks and an Image Guidance Network, producing a refined image R2 and multi-scale features that are pixel-unshuffled and concatenated into the UNet encoder of a LoRA-finetuned Stable Diffusion Turbo model. The Diffusion Branch operates in latent space and produces the final output R1, aided by a long-skip connection and zero-convolution skip connections from the VAE encoder to the decoder. Training uses a combined MSE, LPIPS, and GAN loss on both branches with weights 0.9 and 0.1 for R1 and R2, respectively. Experiments on DIV2K-Val, RealSR, and DRealSR show consistent gains over seven baselines in PSNR, SSIM, LPIPS, DISTS, and FID, including a 1.39 dB PSNR improvement on DRealSR over ResShift, while no-reference metrics (NIQE, MUSIQ, MANIQA, CLIPIQA) are worse for GuideSR, attributed to the perception-distortion tradeoff. The paper's central claim is that full-resolution guidance features, as opposed to VAE-downsampled conditioning, preserve high-frequency structure and drive the reported fidelity gains.","tokens_in":13546,"tokens_out":7184,"duration_ms":62056,"significance":"If the reported gains are reproducible, GuideSR offers a practical advance: state-of-the-art full-reference fidelity at single-step inference cost, and the idea of injecting pixel-unshuffled full-resolution features into a latent diffusion UNet is a plausible mechanism for improving fidelity. The paper is thorough in benchmarking seven baselines across three datasets with six reference-based metrics, and it honestly reports no-reference metrics even though they are worse. However, the central mechanism is not isolated in the ablations, the architecture is not fully specified, and the closest related method (ClearSR) is not compared quantitatively. The significance is therefore contingent on additional experiments that directly test the guidance mechanism.","major_comments":[{"comment":"The ablation in Table 2 does not isolate the contribution of the full-resolution guidance mechanism. The row 'Baseline + Guidance' (26.82 dB) differs from 'Baseline + Long-skip' (26.80 dB) by two components (Guidance vs. Long-skip), so the 0.02 dB difference is not a controlled comparison; the full model's 0.28 dB gain over 'Baseline + Long-skip' adds both Guidance and IGN simultaneously. There is no row 'Baseline + Long-skip + Guidance (without IGN)', no row 'Baseline + Long-skip + IGN', and no control that injects an equally informative low-level signal (e.g., a pixel-unshuffled copy of the input image or a VAE-encoded latent) at the same concatenation points. Consequently, the reported 1.39 dB PSNR gain on DRealSR cannot be attributed to the proposed full-resolution guidance; it could be due to the added parameters, the IGN, or the long-skip. Please add the missing control rows and report metrics beyond PSNR (SSIM, LPIPS, DISTS, FID) in the ablation.","section":"Section 4.3, Table 2"},{"comment":"The downsampling factor s in Eq. (3) is not specified, and the text does not state at which UNet encoder scales the features F'_r are concatenated or how the channel dimensions are matched between the pixel-unshuffled features and the UNet encoder outputs. The statement that pixel-unshuffle 'preserves all values' does not distinguish it from other downsampling operations; without a control comparing pixel-unshuffled guidance against VAE-encoded guidance at the same injection points, the paper does not establish that the proposed mechanism retains high-frequency structure better than the standard conditioning it criticizes. Please specify s, the injection scales, and the channel alignment, and add the corresponding control experiment.","section":"Section 3.1, Eq. (3)"},{"comment":"The paper cites ClearSR [39] as related work but does not include it in the quantitative comparison. Given that ClearSR is described as also providing latent low-resolution embeddings to help diffusion-based super-resolution 'see clearer', it is the closest published method to GuideSR; omitting it from Table 1 leaves the state-of-the-art claim unsubstantiated. Please either add a quantitative comparison on the shared benchmarks or explicitly justify its exclusion (e.g., if the test protocols differ).","section":"Section 4.1, Table 1"}],"minor_comments":[{"comment":"The abstract contains a grammatical error: 'and (2) a Diffusion Branch, which a pre-trained latent diffusion model to enhance perceptual quality' should read 'which uses a pre-trained latent diffusion model to enhance perceptual quality'.","section":"Abstract"},{"comment":"The text says 'paried LQ and 512×512 HQ images'; 'paried' should be 'paired'.","section":"Section 4, Datasets"},{"comment":"The sentence 'We utilizes dual discriminators' should be 'We utilize dual discriminators'.","section":"Section 3.3"},{"comment":"In the DIV2K-Val paragraph, 'surpassing the best previous method (ResShift) by 0.11dB and 0.0152' should specify that these are PSNR and SSIM gains, respectively.","section":"Section 4.1"},{"comment":"The fixed timestep t_f is not given; please state its value and how it was selected.","section":"Section 3.2, Eq. (4)"},{"comment":"The paper does not state whether the baseline numbers in Table 1 are taken from the original papers or re-evaluated under a common protocol; please clarify this to assess fairness of the comparison.","section":"Section 4.1, Table 1"},{"comment":"The abstract's 'state-of-the-art performance' is qualified in the body as applying to reference-based metrics; however, on no-reference metrics (NIQE, MUSIQ, MANIQA, CLIPIQA) GuideSR is consistently worse than several baselines. Please adjust the abstract or add an explicit qualifying phrase to avoid overclaiming.","section":"Section 4.1"},{"comment":"No code or checkpoints are released. Given the reproducibility concerns raised in the major comments, the authors should release code and model weights upon acceptance.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses an important problem and the proposed architecture is plausible, but the evidence for the core mechanism is currently insufficient. The missing ClearSR comparison is a notable omission because ClearSR is the closest prior work. I recommend major revision and encourage the authors to add the suggested ablations, specify the architecture details, and provide code to facilitate verification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a genuine architectural contribution. GuideSR combines a full-resolution guidance branch (FRB + IGN) with a LoRA-finetuned one-step SD Turbo diffusion branch, feeding pixel-unshuffled features into the UNet. The results are strong on three benchmarks, with consistent improvements over several baselines. That's real.\n\nThe central mechanism, though, is not actually isolated. The ablation in Table 2 is missing the critical rows. Baseline + Guidance + Long-skip (without IGN) is not reported, nor is Baseline + Long-skip + IGN. The comparison that would attribute the 0.28 dB gain of the full model to the full-resolution guidance mechanism is simply absent. The 0.02 dB difference between Baseline+Guidance and Baseline+Long-skip is not a measure of marginal gain—it's two separate single-component additions. So the paper's claim that full-resolution pixel-unshuffled guidance preserves high-frequency structure better than VAE conditioning is plausible but unverified. The stress-test note is right about the missing control: no injection of an equally informative low-level signal (like a VAE-encoded full-res latent) through the same concatenation points.\n\nAlso, the paper cites ClearSR ([39]) but never compares to it, which is a serious omission given ClearSR also operates on low-resolution embeddings in a diffusion SR setup. The architecture is not fully reproducible from the text: the pixel-unshuffle factor s and the exact UNet injection points are not specified. No code or checkpoints are released.\n\nWhat the paper does well: the motivation is clear, the dual-branch design is new relative to OSEDiff/DiffBIR/SeeSR, and the experiments span multiple datasets and metrics. The loss design (dual discriminators, shared weights) is sensible. The perception-distortion explanation for the poor no-reference scores is legitimate.\n\nRead this as a strong technical report rather than a complete validation. Deserves peer review; a referee should push for the missing ablations, the ClearSR comparison, and code release.","headline":"A promising one-step diffusion SR architecture with a plausible full-resolution guidance mechanism, but the ablation that would prove the mechanism distinct from long-skips is missing.","tokens_in":14103,"tokens_out":3230,"would_cite":true,"duration_ms":30333,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T45","68U10"],"pacs":[],"model":"deepseek-v4-flash","headline":"GuideSR claims that one-step diffusion super-resolution beats multi-step baselines by conditioning on full-resolution features instead of VAE latents.","keywords":["image super-resolution","one-step diffusion","latent diffusion model","structural fidelity","full-resolution guidance","real-world super-resolution","low-rank adaptation","channel attention"],"falsifier":"Train GuideSR twice with identical losses, data, and budget, but in one version replace the pixel-unshuffled Guidance Branch features with VAE-encoded features of the same size (or with a bilinear-downsampled full-resolution map). If the PSNR gap on DRealSR shrinks to near zero, the full-resolution mechanism, not the dual-branch training, is the source of the gain. A second check: run the Guidance Branch alone with the same losses; if it matches the dual-branch PSNR, the diffusion branch is not contributing.","tokens_in":12997,"feed_emoji":"🖼️","tokens_out":5653,"duration_ms":51559,"temperature":0.7,"pith_summary":"The paper claims that diffusion-based super-resolution loses fidelity because it conditions on VAE-downsampled latent codes that discard high-frequency structure, and that this can be fixed by extracting guidance features at the original image resolution and injecting them into a one-step latent diffusion model. Its proposed model, GuideSR, combines a full-resolution Guidance Branch with a LoRA-finetuned one-step Stable Diffusion Turbo branch, and reports state-of-the-art PSNR, SSIM, LPIPS, DISTS, and FID on DIV2K-Val, RealSR, and DRealSR while using a single inference step. A sympathetic reader would take the paper's central bet to be: full-resolution feature conditioning, rather than more training or bigger models, is what closes the fidelity gap in single-step diffusion SR.","feed_headline":"Full-res guidance lifts one-step diffusion SR past multi-step models","feed_subtitle":"GuideSR feeds original-resolution structure into a one-step diffusion UNet, gaining 1.39 dB on real-world data.","key_machinery":"The load-bearing mechanism is a dual-branch architecture. The Guidance Branch runs at full resolution: a shallow convolution extracts features, a chain of Full Resolution Blocks (residual-in-residual blocks with channel attention) refines them, and an Image Guidance Network applies guided attention to output a refined image and enriched features. Those features are downsampled by pixel-unshuffle (rearranging spatial values into channels, so no information is discarded) and concatenated into the UNet encoder at multiple scales, while a long-skip connection lets the UNet predict only a residual latent. The Diffusion Branch is a pretrained Stable Diffusion Turbo finetuned with LoRA (rank 8 on UNet, rank 4 on VAE), with zero-conv skip connections between VAE encoder and decoder, and trained with a weighted MSE+LPIPS+GAN loss supervised on both branch outputs. The key move is that guidance information enters at full resolution and is preserved through pixel-unshuffle rather than through lossy VAE encoding.","core_discovery":"GuideSR's central claim is that replacing VAE-latent conditioning with full-resolution feature guidance solves the fidelity bottleneck of one-step diffusion super-resolution. The Guidance Branch processes the degraded image at its original resolution with Full Resolution Blocks and an Image Guidance Network, then pixel-unshuffles the features to latent resolution and concatenates them into the UNet encoder of a LoRA-finetuned Stable Diffusion Turbo model. A long-skip connection makes the UNet predict only a residual latent. On DIV2K-Val, RealSR, and DRealSR the method reports higher PSNR, SSIM, LPIPS, DISTS, and FID than both multi-step (StableSR, DiffBIR, SeeSR, PASD, ResShift) and one-step (SinSR, OSEDiff) baselines, with the largest gain on DRealSR (1.39 dB over ResShift). The paper argues this shows structural fidelity and generative quality are not in conflict when the generative branch receives high-frequency guidance.","pith_inferences":["The paper does not isolate how much of the gain comes from pixel-unshuffle versus the adversarial training, the extra long-skip, or the LoRA finetune; a fair test would match all other variables and vary only the conditioning representation, and that test is not reported.","Pixel-unshuffle preserves all spatial values, so the Guidance Branch likely provides information that VAE encoding genuinely discards; if so, the same trick could improve other latent diffusion restoration tasks such as deblurring, deraining, or low-light enhancement.","The 1.39 dB gain on DRealSR over ResShift, a 15-step method, is large enough that reproducing it with an independent implementation would be a strong check; if it shrinks under matched training budgets, the architecture's contribution may be smaller than claimed.","One testable extension: feed the Guidance Branch features to a non-diffusion regressor with the same losses and see whether the diffusion branch adds perceptual value or mostly inherits the guidance branch's fidelity."],"forward_implications":["One-step diffusion SR can exceed multi-step methods on reference-based fidelity metrics, so 15-200 step inference is not necessary for state-of-the-art fidelity.","Full-resolution feature concatenation is a viable alternative to ControlNet-style conditioning for restoration, preserving structure without extra controllers.","The method improves real-world SR fidelity most on challenging datasets, suggesting full-resolution guidance generalizes beyond synthetic degradations.","The same dual-branch recipe may transfer to other latent-space generative restoration tasks that currently use VAE-latent conditioning.","With one-step inference, the restored output quality is available at near-GAN-level cost, making diffusion SR practical for deployment if memory footprint is addressed."],"supporting_citations":[{"why":"The one-step LoRA-finetuned diffusion baseline whose VAE-latent conditioning GuideSR is designed to improve.","marker":"[46]"},{"why":"Multi-step ResShift baseline that GuideSR surpasses by 1.39 dB on DRealSR, and whose degradation pipeline supplies training data.","marker":"[50]"},{"why":"Stable Diffusion Turbo, the pretrained one-step diffusion model used as the Diffusion Branch's generative prior.","marker":"[36]"},{"why":"Low-rank adaptation (LoRA) used to finetune the UNet and VAE with ranks 8 and 4.","marker":"[12]"},{"why":"Source of the zero-conv skip connections added between VAE encoder and decoder.","marker":"[31]"},{"why":"The perception-distortion tradeoff argument used to explain why lower no-reference metric scores accompany the best full-reference scores.","marker":"[2]"},{"why":"SinSR, a one-step diffusion SR competitor that GuideSR is compared against on all benchmarks.","marker":"[43]"},{"why":"StableSR, a multi-step baseline whose evaluation sets test GuideSR and whose 200-step results mark the multi-step baseline.","marker":"[40]"}],"fun_headline_variants":["One-step SR with full-res guidance beats multi-step diffusion models","GuideSR adds full-res structure, gains 1.39 dB in one-step SR","Full-res features unlock high-fidelity one-step diffusion SR","1.39 dB boost: GuideSR's full-res guidance beats multi-step SR","Single-step diffusion SR with full-res guidance outshines multi-step"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central premise is that full-resolution pixel-unshuffled features carry high-frequency structure that VAE downsampling loses, and that this information, not the extra training or the long-skip connection, is what drives the fidelity gains.","fun_headline_variants_meta":{"raw":{"variants":["One-step SR with full-res guidance beats multi-step diffusion models","GuideSR adds full-res structure, gains 1.39 dB in one-step SR","Full-res features unlock high-fidelity one-step diffusion SR","1.39 dB boost: GuideSR's full-res guidance beats multi-step SR","Single-step diffusion SR with full-res guidance outshines multi-step"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00024,"raw_usage":{"total_tokens":1540,"prompt_tokens":990,"completion_tokens":550,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":455}},"tokens_in":606,"tokens_out":550,"duration_ms":4881,"temperature":1.0,"reasoning_tokens":455,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:35:51.824287+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train GuideSR twice with identical losses, data, and budget, but in one version replace the pixel-unshuffled Guidance Branch features with VAE-encoded features of the same size (or with a bilinear-downsampled full-resolution map). If the PSNR gap on DRealSR shrinks to near zero, the full-resolution mechanism, not the dual-branch training, is the source of the gain. A second check: run the Guidance Branch alone with the same losses; if it matches the dual-branch PSNR, the diffusion branch is not contributing.","supporting_citations":[{"cited_title":"Fast high- resolution image synthesis with latent adversarial diffusion distillation","cited_arxiv_id":null,"evidence_quote":"Stable Diffusion Turbo, the pretrained one-step diffusion model used as the Diffusion Branch's generative prior."},{"cited_title":"The perception-distortion tradeoff","cited_arxiv_id":null,"evidence_quote":"The perception-distortion tradeoff argument used to explain why lower no-reference metric scores accompany the best full-reference scores."},{"cited_title":"Exploiting diffusion prior for real-world image super-resolution.International Journal of Computer Vision, pages 1–21, 2024","cited_arxiv_id":null,"evidence_quote":"StableSR, a multi-step baseline whose evaluation sets test GuideSR and whose 200-step results mark the multi-step baseline."}],"review_version":1}