{"id":"07d4248e-fb84-437b-a348-17c8234c727f","arxiv_id":"2412.07152","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Hero-SR adds a dynamic time-step selector and CLIP-based multimodal losses to a distilled one-step diffusion model, reporting state-of-the-art perceptual metrics on real-world super-resolution benchmarks.","lead":"This paper presents Hero-SR, a one-step diffusion model for real-world image super-resolution that chooses its starting noise step per image and uses CLIP-based text and image supervision to match human perceptual preferences. It reports higher no-reference perceptual quality scores than existing one-step and multi-step diffusion SR methods, while using only a single diffusion step.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA perceptual-quality claim is only supported by single-run no-reference IQA scores; without human preference data or error bars, the central 'human perception' claim is not yet established.","rationale":"Agree with the reader's weakest assumption. The single most load-bearing condition is that the six learned no-reference metrics used in Tables 2 and 3 are trustworthy proxies for human perception and that the differences are reliable. The paper's own contribution is 'human perception priors', so the SOTA claim cannot be separated from this condition. The reported evidence is single-run, no CIs, no human study, and the training objective is itself a CLIP-based perceptual-alignment loss, creating a plausible correlation between what is optimized and what is measured. The DTSM under-specification (Eq. 4, candidate set S and temperature) is a reproducibility concern, but the metric-validation concern is what directly supports or refutes the headline claim. I therefore keep the CONDITIONAL verdict: the core idea is plausible and the ablations are informative, but the human-perception claim needs a human preference test and error bars before acceptance as stated.","tokens_in":15725,"tokens_out":6652,"duration_ms":75297,"concrete_test":"Run a pairwise human preference study on 100 stratified samples from DRealSR and RealSR, comparing Hero-SR against SeeSR and OSEDiff (and a fixed-time-step Hero-SR variant), with at least 20 raters; report preference shares and 95% CIs. Independently, bootstrap the Table 2/3 metric deltas over images (1000 resamples) and report CIs; if Hero-SR is not preferred by humans, or the metric deltas overlap zero, the SOTA perceptual claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Hero-SR's outputs align with human perception better than one-step and multi-step SR methods. The evidence is Tables 2-3: single-run means on six learned no-reference IQA metrics. This is load-bearing because Hero-SR is explicitly trained with TD-PAL, which maximizes CLIP similarity to prompts like 'good image' and 'sharp image'. Learned no-reference metrics are themselves trained to approximate human ratings, and several are likely correlated with CLIP-based features, so the evaluation may reward the training objective rather than independently confirming perceptual quality. The paper provides no human preference study and no confidence intervals, and some reported deltas are tiny (e.g., RealSR MUSIQ 70.03 vs 69.99 in Table 2; in Table 3, SeeSR beats Hero-SR on DRealSR TRES 84.73 vs 83.59). Without a human preference test and error bars, the reported metric advantages cannot be distinguished from metric noise or metric overfitting. The DTSM also lacks specification of candidate set S, Gumbel temperature, and how t* is applied, but the evaluation weakness is the more direct threat to the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Hero-SR, a one-step diffusion-based super-resolution framework with two new components: a Dynamic Time-Step Module (DTSM) that selects a per-image starting diffusion step, and Open-World Multi-modality Supervision (OWMS) that adds CLIP-based text-domain and image-domain losses. The authors report experiments on DIV2K-val, DRealSR, and RealSR, comparing with one-step and multi-step diffusion SR baselines, and claim state-of-the-art performance on no-reference perceptual metrics such as MUSIQ, HyperIQA, TOPIQ, TRES, ARNIQA, and Q-Align. The paper includes ablations for DTSM, ID-SAL, TD-PAL, and individual perceptual attribute prompts, and states that code will be released.","tokens_in":15976,"tokens_out":6344,"duration_ms":56567,"significance":"If the perceptual-quality claim is substantiated, the paper makes a useful contribution: DTSM is a novel alternative to fixed-step one-step diffusion SR, and OWMS gives a simple, potentially reproducible recipe for injecting human-perceptual priors via frozen CLIP encoders. The module-level and attribute-level ablations are valuable, and the promise of public code is a strength. However, the current evaluation does not yet establish the central claim of superior alignment with human perception, and the DTSM specification is incomplete. The result is therefore promising but not yet fully supported.","major_comments":[{"comment":"The text makes blanket quantitative claims that are contradicted by the paper's own tables. It states that Hero-SR 'outperforms other methods across all datasets for no-reference perceptual metrics' and 'achieves leading results in full-reference metrics, surpassing other methods in PSNR, SSIM, and LPIPS.' Table 3 shows SeeSR achieves higher HyperIQA and TRES on DRealSR (0.6642 vs 0.6434 and 84.7264 vs 83.5888, respectively) and higher TOPIQ on RealSR (0.6891 vs 0.6881). Table 2 and Table 3 also show SinSR and ResShift exceeding Hero-SR on PSNR on some datasets (e.g., DIV2K PSNR 24.4111 vs 24.3663 for SinSR, RealSR PSNR 26.5344 vs 25.8271 for ResShift). These overstatements need to be corrected, and the claims should be aligned with the actual tables.","section":"Section 3.2, Eq. (4)"},{"comment":"The DTSM is underspecified and therefore not reproducible. Eq. (4) defines t* = Gumbel-Softmax(v, S), but the candidate set S is never described—its size, elements, and spacing are all omitted. The Gumbel-Softmax temperature is also never reported, and it is unclear whether a discrete or continuous relaxation is used at inference. Furthermore, the paper does not explain how the selected t* is fed to the U-Net (e.g., as a discrete step index, a time embedding, or a vector) and whether zLR is first noised to that step when t* is not the maximal step. Consequently, the DTSM ablation in Table 5 is difficult to interpret because the fixed-step baseline used in Variant-1 is not specified.","section":"Section 3.2, Eq. (4)"},{"comment":"The central claim that Hero-SR better aligns with human perception is not yet established by the reported evidence. All perceptual comparisons rely on single-run means of learned no-reference IQA metrics, with no error bars, significance tests, or human-preference study. This is particularly concerning because TD-PAL explicitly optimizes CLIP similarity between the generated image and positive perceptual prompts, and the evaluation metrics (MUSIQ, HyperIQA, TOPIQ, TRES, ARNIQA, Q-Align) are themselves learned models that may use CLIP-like features; the evaluation could therefore partly reward the training objective rather than independently confirm perceptual superiority. Some reported differences are very small (e.g., RealSR MUSIQ 70.0254 vs 69.9864 in Table 2), so without confidence intervals or a human study, these differences cannot be distinguished from metric noise.","section":"Section 4.2.1 and Section 4.3"}],"minor_comments":[{"comment":"The loss is named LTD-SAL in Eq. (9) but is referred to as TD-PAL in the text and as LTD-PAL in Eq. (13); please unify the notation.","section":"Eq. (9) and Eq. (13)"},{"comment":"The compared methods list contains the typo 'ResShit' instead of 'ResShift'.","section":"Section 4.1"},{"comment":"The ARNIQA reference is listed as CVPR 2024, but the paper actually appeared at WACV 2024.","section":"Reference [1]"},{"comment":"There are several typos: 'Fighure 2' in Section 3.1, 'sematic' in Section 3.3.2, and 'GAN-base' in Appendix A.","section":"Throughout"},{"comment":"Table 4 is titled 'perceptual attributions' while the text uses 'perceptual attributes'; please use consistent terminology.","section":"Table 4 and Section 4.3"},{"comment":"The claim 'we are the first to incorporate multimodal models into the training of Real-SR tasks' is too strong given prior semantics-aware Real-SR methods such as SeeSR, PASD, and SUPIR that use textual or CLIP-based guidance; please qualify this contribution.","section":"Section 1, contributions"},{"comment":"The limitation paragraph only mentions VAE reconstruction capacity; it should also acknowledge the lack of a human-preference evaluation for the perceptual claim.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The core idea is solid and the ablations support the module contributions, but the evaluation and specification gaps are load-bearing for the SOTA and human-perception claims. The paper is likely acceptable after a careful revision that corrects the overstated claims, adds statistical support or a human study, and fully specifies DTSM. I saw no citation-pattern issues; the code-release promise is a positive signal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: Hero-SR is a solid incremental contribution to one-step real-world super-resolution. The interesting part is the Dynamic Time-Step Module (DTSM), which picks a per-image starting diffusion step instead of always starting from the same noise level, and the ablations show it earns its keep. The OWMS supervision (CLIP text-pair plus image-semantic losses) is a combination of known ingredients but applied cleanly. If the numbers hold, this is a practically useful result: one-step SR that beats multi-step methods on several no-reference metrics while roughly matching PSNR/SSIM.\n\nWhat the paper does well: the experimental sweep is broad (DIV2K, RealSR, DRealSR; one-step, multi-step, GAN baselines in the appendix). The ablations in Table 5 are informative—removing TD-PAL craters MUSIQ/TOPIQ/TRES, and removing DTSM or ID-SAL also hurts. Training details are mostly there (SD-Turbo, LoRA rank 16, loss weights). The framework is cheap to train (2 L40s, 8 hours), which matters for adoption.\n\nSoft spots, in order of importance. First, the central claim—that Hero-SR aligns with human perception—is not backed by human data. The evidence is single-run means on learned no-reference IQA metrics. TD-PAL directly optimizes CLIP similarity toward 'good image' prompts, and several of those metrics are trained on human ratings and may correlate with CLIP features, so the evaluation could be partly rewarding the training objective. Some deltas are tiny (RealSR MUSIQ 70.03 vs 69.99; on DRealSR TRES, SeeSR actually beats Hero-SR 84.73 vs 83.59). Without error bars or a human preference study, the SOTA claim in the paper is not established. This is the load-bearing weakness.\n\nSecond, the DTSM is under-specified. Eq. (4) never defines the candidate set S or the Gumbel-Softmax temperature, and it is not clear how the selected t* is actually applied in the one-step denoising. That matters for reproducibility and for understanding what the module really learns. Third, a naming bug: Eq. (9) calls the loss LTD-SAL while the text calls it TD-PAL. Minor, but it will trip readers. Fourth, the novelty claim about being 'first' to incorporate multimodal models into Real-SR training is overstated; SeeSR and PASD already use semantic guidance, and CLIP-based perceptual losses are not new. The specific combination is new enough without overclaiming. No code is released, only a project page saying it will come.\n\nOverall: the core idea is plausible and the paper is above the median for this subfield. It needs major revision—human evaluation or at least error bars plus a metric-agreement analysis, full DTSM details, and toned-down claims—but it deserves referee time. I would not desk-reject it.","headline":"Solid incremental one-step SR paper whose adaptive time-step idea is worth knowing, but the human-perception claim needs human data before it can be taken at face value.","tokens_in":16508,"tokens_out":2872,"would_cite":true,"duration_ms":30210,"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":"Hero-SR is a one-step diffusion super-resolution model that adaptively picks a per-image diffusion step and applies CLIP text-image supervision to achieve state-of-the-art perceptual quality on real-world benchmarks.","keywords":["real-world super-resolution","one-step diffusion","dynamic time-step selection","CLIP multimodal supervision","perceptual quality","no-reference image quality assessment","semantic consistency","image restoration"],"falsifier":"Run a paired human preference study on the same test sets (DIV2K, DRealSR, RealSR) comparing Hero-SR output against the closest baselines such as SeeSR and OSEDiff; if human raters do not prefer Hero-SR at rates significantly above chance, the perceptual claim is falsified. Alternatively, recompute the no-reference metrics over multiple training runs and check whether the reported gaps (e.g., the 7.0% TOPIQ improvement) exceed run-to-run variance.","tokens_in":15521,"feed_emoji":"🖼️","tokens_out":9147,"duration_ms":76470,"temperature":0.7,"pith_summary":"The paper aims to show that a single diffusion step can produce real-world super-resolved images that meet human perceptual standards, provided the model adaptively selects its starting noise level per input image and supervises the output with CLIP-based multimodal guidance. On the DIV2K, DRealSR, and RealSR benchmarks, Hero-SR reports the best or second-best scores among compared one-step and multi-step diffusion methods across no-reference perceptual metrics such as MUSIQ, HyperIQA, TOPIQ, TRES, ARNIQA, and Q-Align. The authors argue this is significant because existing one-step methods fix a single starting time step for every image, ignoring the wide variation in degradation and structural complexity across real-world inputs, and because semantic consistency and perceptual naturalness are usually under-supervised. If the claim holds, perceptually aligned super-resolution can be achieved in a single diffusion step, lowering inference cost while matching or exceeding slower multi-step approaches.","feed_headline":"Beats 50-step super-resolution with one step on perceptual quality","feed_subtitle":"Adaptive diffusion steps and CLIP text-image supervision lift no-reference quality scores on real-world benchmarks.","key_machinery":"Two modules carry the argument. The Dynamic Time-Step Module (DTSM) maps a low-resolution image to a noise-level index via convolution, residual blocks, an MLP, and Gumbel-Softmax, enabling a differentiable per-image choice of denoising start point. The Open-World Multi-modality Supervision (OWMS) combines Text-Domain Perceptual Alignment Loss (TD-PAL), which uses CLIP to score cosine similarity between the generated image and positive/negative text prompt pairs for six perceptual attributes, and Image-Domain Semantic Alignment Loss (ID-SAL), which aligns the CLIP image embedding of the output with that of the ground truth. Together they define the total loss alongside MSE and LPIPS terms, with SD-Turbo as the base diffusion model fine-tuned by LoRA.","core_discovery":"The central claim is that human-perceptual alignment in one-step real-world super-resolution is achievable by jointly (1) dynamically selecting the diffusion time step $t^*$ according to the input's degradation and complexity, via a convolutional feature extractor and Gumbel-Softmax sampling, and (2) supervising the output with CLIP in both text and image domains: text prompts representing perceptual attributes (quality, sharpness, edge clarity, resolution, noise, clarity) pull the image toward positive attributes and away from negative ones, while image-domain alignment with the ground truth enforces semantic consistency. On DIV2K, DRealSR, and RealSR test sets, the model reports the best or second-best scores on no-reference metrics such as MUSIQ, HyperIQA, TOPIQ, TRES, ARNIQA, and Q-Align among compared one-step and multi-step methods, with a single diffusion step.","pith_inferences":["Beyond the paper, the DTSM design implies that the fixed start-step convention in one-step diffusion distillation may itself be a bottleneck; adaptivity could generalize to text-to-image or image translation with variable difficulty.","Our extension: the unstated candidate set $S$ and Gumbel-Softmax temperature in Eq.~4 are reproducibility-critical; we infer that sweeping them would change perceptual scores, so the paper's claim is only fully testable once those hyperparameters are disclosed.","Our extension: if no-reference IQA metrics are the optimization target, a risk is that the model over-fits those particular metrics rather than human perception; a human study would separate these cases.","We also infer that the framework may be extended by replacing the fixed six perceptual attribute prompts with an open-vocabulary or learned prompt set, potentially improving adaptation to new degradation types."],"forward_implications":["Per-image time-step selection can be applied to other one-step diffusion restoration tasks, since degradation level varies across images.","The reported results imply that a single diffusion step can match or exceed multi-step diffusion methods on no-reference perceptual metrics, making high-quality super-resolution more computationally accessible.","Incorporating CLIP-based text supervision in the training objective provides a way to inject human perceptual attributes without collecting human preference labels.","The combination of TD-PAL and ID-SAL suggests that both text-defined attributes and image-level semantic consistency contribute to perceptual quality, and ablations indicate each module is needed for top scores."],"supporting_citations":[{"why":"Supplies the frozen CLIP image and text encoders that both OWMS losses are built on.","marker":"[31]"},{"why":"Provides the Gumbel-Softmax reparameterization that makes DTSM's discrete time-step selection trainable.","marker":"[18]"},{"why":"Is the pre-trained SD-Turbo diffusion model that Hero-SR fine-tunes with LoRA.","marker":"[33]"},{"why":"Is the one-step fixed-time-step baseline that Hero-SR improves on and compares against.","marker":"[45]"},{"why":"Is the semantic-guidance multi-step baseline used in quantitative and qualitative comparisons.","marker":"[46]"},{"why":"Provides the StableSR test set used for evaluation and serves as a multi-step diffusion baseline.","marker":"[37]"},{"why":"Provides the Real-ESRGAN degradation pipeline used to synthesize LR-HR training pairs from LSDIR.","marker":"[39]"},{"why":"Introduces the CLIP-based positive/negative prompt comparison scheme that TD-PAL adapts.","marker":"[36]"},{"why":"Is a no-reference image quality metric used to evaluate perceptual quality of the outputs.","marker":"[44]"}],"fun_headline_variants":["One-step super-resolution with human perception priors","Adaptive diffusion step and CLIP lift one-step SR quality","Single-step diffusion SR with dynamic time-step and CLIP","Hero-SR: one-step real-world SR with human perception","One-step SR that preserves details and aligns with human perception"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's headline conclusion rests on the assumption that the no-reference quality metrics it optimizes for and reports (MUSIQ, HyperIQA, TOPIQ, TRES, ARNIQA, Q-Align) are faithful proxies for human perception; if those metrics disagree with human preference, or if the reported single-run differences are within metric noise, the state-of-the-art claim loses its support.","fun_headline_variants_meta":{"raw":{"variants":["One-step super-resolution with human perception priors","Adaptive diffusion step and CLIP lift one-step SR quality","Single-step diffusion SR with dynamic time-step and CLIP","Hero-SR: one-step real-world SR with human perception","One-step SR that preserves details and aligns with human perception"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000439,"raw_usage":{"total_tokens":2203,"prompt_tokens":897,"completion_tokens":1306,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":1225}},"tokens_in":513,"tokens_out":1306,"duration_ms":12555,"temperature":1.0,"reasoning_tokens":1225,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:04:05.024415+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a paired human preference study on the same test sets (DIV2K, DRealSR, RealSR) comparing Hero-SR output against the closest baselines such as SeeSR and OSEDiff; if human raters do not prefer Hero-SR at rates significantly above chance, the perceptual claim is falsified. Alternatively, recompute the no-reference metrics over multiple training runs and check whether the reported gaps (e.g., the 7.0% TOPIQ improvement) exceed run-to-run variance.","supporting_citations":[{"cited_title":"Learning transferable visual models from natural language supervision","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen CLIP image and text encoders that both OWMS losses are built on."},{"cited_title":"Categorical repa- rameterization with gumbel-softmax","cited_arxiv_id":null,"evidence_quote":"Provides the Gumbel-Softmax reparameterization that makes DTSM's discrete time-step selection trainable."},{"cited_title":"Adversarial diffusion distillation","cited_arxiv_id":null,"evidence_quote":"Is the pre-trained SD-Turbo diffusion model that Hero-SR fine-tunes with LoRA."},{"cited_title":"One-step effective diffusion network for real-world image super-resolution","cited_arxiv_id":null,"evidence_quote":"Is the one-step fixed-time-step baseline that Hero-SR improves on and compares against."},{"cited_title":"Seesr: Towards semantics-aware 10 real-world image super-resolution","cited_arxiv_id":null,"evidence_quote":"Is the semantic-guidance multi-step baseline used in quantitative and qualitative comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the StableSR test set used for evaluation and serves as a multi-step diffusion baseline."},{"cited_title":"Real-esrgan: Training real-world blind super-resolution with pure synthetic data","cited_arxiv_id":null,"evidence_quote":"Provides the Real-ESRGAN degradation pipeline used to synthesize LR-HR training pairs from LSDIR."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the CLIP-based positive/negative prompt comparison scheme that TD-PAL adapts."},{"cited_title":"Q-align: Teaching lmms for visual scoring via discrete text-defined levels","cited_arxiv_id":null,"evidence_quote":"Is a no-reference image quality metric used to evaluate perceptual quality of the outputs."}],"review_version":1}