{"id":"c947691e-8c49-464b-87c3-b6d0542b2e16","arxiv_id":"2505.12048","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"TSS accelerates diffusion super-resolution by concentrating denoising steps in early and late iterations and adapting the schedule per image region, improving perceptual scores with fewer steps.","lead":"This paper introduces a training-free way to speed up diffusion-based image super-resolution by sampling more heavily in the denoising stages that matter for fine detail. If it holds up, it could cut inference cost by roughly half while keeping or improving perceptual quality.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SDS's per-pixel timestep embedding (Eq. 9) is not a true per-pixel denoiser: the latent carries one global noise level, and the spatial embeddings are an untrained conditioning; A.9 concedes boundary artifacts, and Tab. 2 shows only marginal SDS gains.","rationale":"The paper's central claim requires that the spatial dynamic sampling step is both a valid extension of timestep conditioning and a measurable improvement. The first is not satisfied: Eq. 9 injects per-pixel embeddings into a latent that has a single global noise level, so these embeddings cannot represent the per-pixel noise levels they claim. This is not merely a boundary artifact but a fundamental mismatch in the mechanism; A.9 only acknowledges the boundary consequence. The second is weakly supported: Table 2 shows Step=14 SDS worsens NIQE and gives near-zero MUSIQ and Qalign gains, and no variance estimates are provided. If a random-permutation control reproduces the TSS gains, then the \"adaptive to image content\" claim is unsupported and the contribution reduces to TDS plus arbitrary spatial noise. This is the single most load-bearing concern because it targets the only component that makes the method spatial. The reader's weakest_assumption points to the same mechanism, so I agree. Since the issue is empirically settleable, the CONDITIONAL verdict stands: the paper should be accepted only with the control experiment, statistical reporting, and code. I therefore leave the reader's verdict unchanged.","tokens_in":20935,"tokens_out":11391,"duration_ms":119102,"concrete_test":"On RealPhoto60, rerun the Step=7 and Step=14 rows of Tab. 2 with the variance-derived spatial timestep map replaced by a random permutation of the same map (same multiset of per-pixel timesteps, positions shuffled), using at least 5 shuffles and reporting mean±std for NIQE/MUSIQ/Qalign. If the shuffled control matches TSS within the spread, the content-adaptive claim fails and SDS is just spatial noise; if it clearly degrades, the variance mapping matters, though the single-noise-level mismatch in Eq. 9 still needs a patch-wise or re-noised per-pixel control.","verdict_should_be":"UNCHANGED","load_bearing_attack":"At each reverse step the denoiser input z is a single latent at one global noise level, but SDS (Eqs. 5-9) adds a different timestep embedding Emb(t_{i,j}) at every pixel of that same z. The network therefore cannot be performing per-pixel denoising at different timesteps: the embeddings do not describe the actual noise level of any pixel. SDS is better described as a spatially varying, content-adaptive conditioning signal that the pre-trained network never saw, and A.9 concedes that large neighboring timestep intervals hurt quality. This is the load-bearing step because it is the only component that makes TSS \"spatial\"; if it is merely a learned-from-scratch modulation, the claimed mechanism and the \"without any extra training cost\" advantage are not established. The empirical support is also thin: in Tab. 2, adding SDS to TDS at Step=14 makes NIQE worse (3.59 vs 3.65) and changes MUSIQ by only +0.10, with no error bars. The central SOTA/acceleration claim therefore rests on a component whose mechanism is invalid as stated and whose measured contribution is close to noise.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Time-Spatial-aware Sampling (TSS), a training-free inference-time acceleration strategy for diffusion-based super-resolution. TSS combines Time Dynamic Sampling (TDS), a non-uniform polynomial timestep schedule that concentrates denoising steps in early and late iterations, with Spatial Dynamic Sampling (SDS), which modifies the timestep embedding spatially according to local variance. The method is motivated by an SNR analysis of SUPIR on RealPhoto60 and is evaluated by integrating TSS into StableSR, SUPIR, and PASD across six benchmarks with perceptual metrics and a full-reference appendix. The paper claims state-of-the-art perceptual quality with fewer steps than existing acceleration methods such as ResShift and SinSR.","tokens_in":21218,"tokens_out":6315,"duration_ms":64178,"significance":"If the claims hold, TSS would be a useful contribution because it is training-free, compatible with existing samplers, and shows consistent perceptual improvements in many of the reported settings. The paper has clear strengths: a comprehensible analysis of frequency- and spatial-domain denoising dynamics, extensive evaluation across three diffusion SR anchors, multiple datasets and metrics, additional FID results, comparison with DM-NonUniform, and a candid limitation discussion in A.9. The main weaknesses are an overstated universal MUSIQ claim, a spatial component (SDS) whose mechanism and measured contribution are not convincingly established, and per-model hyperparameter tuning that weakens the generality claim. These issues are addressable in revision, so the manuscript is not beyond repair.","major_comments":[{"comment":"The SDS mechanism is load-bearing for the claim that TSS is 'spatial' and training-free, but it is not convincingly established. At each reverse step the denoising network receives one latent z at a single global noise level, so the per-pixel embeddings Emb(t_i,j) do not describe the actual noise level of any pixel; the pretrained network was trained only with a global scalar timestep. The paper's own limitation statement in A.9 concedes that large neighboring timestep intervals can harm quality, which is consistent with SDS being an untrained content-adaptive perturbation rather than a true per-pixel denoiser. The empirical support in Table 2 is also weak: at Step=14, adding SDS to TDS makes NIQE worse (3.59 to 3.65) and changes MUSIQ by only +0.10, with no error bars. I request a control experiment that isolates whether the timestep semantics matter, such as replacing Emb(t_i,j) with a smooth variance-derived spatial bias of similar magnitude, and multiple-seed or paired results for the SDS ablation.","section":"§3.2, Eq. (9), A.9"},{"comment":"The abstract's claim of 'improving MUSIQ scores by 0.2–3.0' is not supported by the reported tables. In Table 1 (WebPhoto, SUPIR row), SUPIR-TSS MUSIQ is 58.72 versus 60.42 for SUPIR, a 1.70 drop; in DRealSR, PASD-TSS MUSIQ is 57.58 versus 58.21 for PASD. Several other rows show smaller gains within the noise level of the metric. The abstract, introduction, and conclusion should be qualified to state average or most-benchmark improvements, or the exceptions should be explicitly explained.","section":"Abstract, §4.2, Table 1"},{"comment":"The hyperparameters of TDS and SDS are tuned per anchor model; A.2 states that 'small differences in the values of the hyperparameters are taken to achieve the best performance,' and A.9 acknowledges sensitivity to model architecture. Since the reported gains are the basis of the generality claim, the paper should report the performance of TSS with a single fixed hyperparameter set across all anchors and datasets, or provide a systematic sensitivity analysis for each anchor rather than only for PASD in Table 10.","section":"A.2, Table 5, A.9"},{"comment":"No error bars or significance tests are reported, although some differences are small (e.g., Step=14 MUSIQ gain of +0.10 in Table 2, and several full-reference metrics that worsen in Table 11, such as PASD on RealSR PSNR dropping from 25.35 to 24.50). Because the central claim is that TSS 'consistently' improves metrics, the paper should provide variance across repeated runs or at least bootstrap confidence intervals for the key comparisons.","section":"§4.3, Tables 1, 2, 11"},{"comment":"The empirical motivation in Section 3.1 is based on SUPIR alone, yet TSS is applied to StableSR and PASD as well. The claim that the discovered temporal and spatial dynamics are generic to diffusion-based SR would be strengthened by repeating the frequency/spatial SNR analysis on at least one additional anchor model, especially since the hyperparameter ranges differ across models in Table 5.","section":"§3.1, Table 1"}],"minor_comments":[{"comment":"The table header contains a typo: 'amin, nmax' should be 'amin, amax'; the rows for StableSR and SUPIR also use inconsistent formatting.","section":"A.2, Table 5"},{"comment":"Equation (5) contains 'n(vvgi,j)' with a duplicated 'v'; it should be 'n(vgi,j)'.","section":"Eq. (5)"},{"comment":"The index ranges i in [0, H] and j in [0, W] should be [0, H-1] and [0, W-1] for a grid of size H x W.","section":"Eq. (6)"},{"comment":"The text in A.5 says 'As shown in Tab. 9' when referring to the DM-NonUniform comparison, but the table is labeled Table 8; all cross-references should be checked and corrected.","section":"A.5, Table 8"},{"comment":"Several figure captions in the appendix contain formatting errors, including 'NIQE=,7.59' and 'NIQE=12.00 MUSIQ=26.57' with missing commas or spaces.","section":"Fig. 10"},{"comment":"The text repeatedly refers to 'the full version' for implementation details and additional results; if the full version is the arXiv appendix, the references should say 'Appendix' consistently.","section":"§4.1, figure captions"}],"recommendation":"major_revision","confidential_remarks":"I am sympathetic to the paper and think the central TDS idea is promising and well supported by the reported experiments. The main uncertainties are whether SDS is a meaningful component beyond a content-adaptive bias, and whether the universal MUSIQ claim survives careful reading of Table 1. Both are fixable with additional experiments or by scaling back the claims. The manuscript is within the scope of the journal and should not be rejected on the basis of the empirical rather than theoretical nature of the method."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: the time-step resampling part (TDS) is a real and useful observation. Concentrating reverse-process steps at the beginning and end consistently improves no-reference perceptual scores for StableSR, SUPIR, and PASD at the same step count, and the early-stage ablation in Tab. 3 supports the claim that early steps matter for high-frequency recovery. The spatial part (SDS) is the weak link. The per-pixel timestep embedding in Eq. 9 does not actually make the network denoise each pixel at a different noise level, because the latent z has one global noise level and the pre-trained network was trained on a single scalar embedding. The ablation data agree: SDS adds almost nothing over TDS in Tab. 2, and at 14 steps it makes NIQE slightly worse (3.59 to 3.65). The authors' own A.9 concedes boundary artifacts when neighboring timestep embeddings are far apart. So the mechanism described in Sec. 3.2 is closer to a content-adaptive modulation than per-pixel denoising, and the \"spatial\" novelty rests on a marginal effect.\n\nWhat the paper does well: extensive evaluation across three anchor models and multiple real-world benchmarks, a direct comparison with DM-NonUniform, and an honest limitation appendix. The TDS result is robust across non-uniform functions and shows the expected behavior across models. The efficiency numbers (2.2x speedup for PASD at 7 steps) are concrete and reproducible.\n\nMain soft spots besides SDS: no error bars anywhere, so I can't tell whether the small deltas are meaningful. Hyperparameters are tuned per model (Tab. 5), which is fitting to the test set even if not training. The abstract's \"0.2-3.0 MUSIQ improvement\" has exceptions: WebPhoto SUPIR drops from 60.42 to 58.72, and full-reference PSNR often drops (e.g., PASD on DIV2K from 21.28 to 20.98). This is a perceptual trade-off, not a free lunch.\n\nWho this is for: people working on inference-time acceleration for diffusion SR. It deserves a serious referee, not a desk reject, but acceptance should hinge on releasing code, adding error bars, and rewriting the SDS claims to match what the module actually does.","headline":"TDS is a solid empirical acceleration trick for diffusion SR, but the spatial component doesn't do the mechanistic work claimed and the paper's strongest claims outrun the evidence.","tokens_in":21782,"tokens_out":2478,"would_cite":false,"duration_ms":27664,"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":"Diffusion super-resolution can run in half the steps by concentrating denoising iterations on the early and late timesteps where high-frequency detail is recovered.","keywords":["diffusion models","super-resolution","training-free acceleration","timestep sampling","spatial adaptive sampling","high-frequency recovery","perceptual quality","inference acceleration"],"falsifier":"Run TSS on a pre-trained diffusion super-resolution model while forcing large per-pixel timestep gaps between adjacent smooth and textured regions: if the spatially varying timestep embedding produces visible boundary artifacts or worse perceptual scores than uniform sampling, the central mechanism fails. Alternatively, measuring high-frequency SNR across a backbone outside the Stable Diffusion family and showing that the early/late two-stage pattern disappears would undercut the temporal motivation for TDS.","tokens_in":20686,"feed_emoji":"⚡","tokens_out":4880,"duration_ms":45838,"temperature":0.7,"pith_summary":"The paper argues that existing diffusion-based super-resolution methods waste iterations by sampling denoising steps uniformly, even though high-frequency image details are actually recovered in two concentrated windows: the earliest denoising steps and the final ones. On that basis it proposes Time-Spatial-aware Sampling (TSS), a training-free schedule that places more denoising steps in those critical windows and varies the schedule pixel by pixel according to local texture. The authors claim that attaching TSS to models such as SUPIR, PASD, and StableSR improves perceptual quality, with MUSIQ gains of 0.2 to 3.0, and matches or beats accelerated baselines such as ResShift and SinSR using about half the steps.","feed_headline":"Sampling trick halves diffusion super-resolution steps","feed_subtitle":"Time- and space-aware step allocation sharpens textures of existing SR models with no retraining.","key_machinery":"The Time Dynamic Sampling schedule is built from a two-piece polynomial resampling function $$f(t,a,n,T,T') = \\frac{t^n}{$a^{{n-1}}$}\\; (t<a), \\quad T - \\frac{(T-t)^n}{(T-a)^{n-1}}\\; (t\\geq a),$$ which concentrates sampling density near $t=0$ and $t=T$ when $n>1$ and collapses to uniform sampling as $n\\to 1$. Spatial Dynamic Sampling computes a smoothed local variance map $V_g$ from the low-resolution grayscale image, maps it linearly to per-pixel values of $a$ and $n$, and builds a per-pixel timestep tensor; the resulting spatial timestep embedding $t_{\\text{emb}}^{\\text{spatial}}$ is added to the network's main branch features at every spatial location, replacing the usual single global embedding. Together these two mechanisms let a pre-trained denoiser run one forward pass with a spatially varying timestep, allocating more steps to high-frequency recovery in textured regions while keeping smooth regions on a more uniform schedule.","core_discovery":"The central discovery is a two-stage temporal pattern in the denoising of high-frequency signals: the signal-to-noise ratio for high-frequency components rises visibly in the early iterations (timestep roughly 1000 to 700) and again in the late iterations (roughly 400 to 0), while the middle iterations can actually add noise. Spatially, smooth regions finish denoising early, whereas textured regions, which carry more high-frequency content, concentrate their recovery in the final steps and benefit from a more non-uniform schedule. The paper claims that by exploiting this pattern, Time Dynamic Sampling redistributes steps toward the two critical windows, and Spatial Dynamic Sampling gives each pixel its own timestep schedule based on local variance, delivered through a spatially varying timestep embedding, so that pre-trained diffusion super-resolution models produce sharper textures with fewer total iterations and no additional training.","pith_inferences":["Beyond the paper: the same two-window sampling idea could accelerate other diffusion restoration tasks such as inpainting, deblurring, and text-to-image generation, which the authors name as future work.","Beyond the paper: since Spatial Dynamic Sampling keys off the local variance of the low-resolution image, replacing variance with a semantic or saliency map is a testable extension that may handle smooth-but-structured regions better.","Beyond the paper: the reported artifacts from large neighboring timestep gaps imply an upper bound on spatial timestep contrast for pre-trained networks, so training models with spatial timestep embeddings could lift that bound."],"forward_implications":["TSS attaches to pre-trained diffusion super-resolution models with no training, so any future improvement in the base model immediately inherits the speedup.","At seven steps, SUPIR with TSS beats the fifteen-step ResShift and the one-step SinSR on NIQE, MUSIQ, and QAlign on most benchmarks, showing a step reduction of more than half.","TSS combines with standard accelerated samplers such as DDIM and UniPC, improving all perceptual metrics over each sampler alone.","The non-uniform schedule can be trigonometric, exponential, or polynomial, indicating that the exact functional form is not essential as long as sampling concentrates on the early and late stages."],"supporting_citations":[{"why":"Supplies the SUPIR model and RealPhoto60 dataset used for the frequency- and spatial-domain SNR analysis that motivates TSS.","marker":"[Yu et al., 2024]"},{"why":"Supplies the PASD anchor model on which TSS is evaluated and whose 7-step baseline is improved.","marker":"[Yang et al., 2025]"},{"why":"Supplies the StableSR anchor model and its 100-step baseline that TSS accelerates.","marker":"[Wang et al., 2024a]"},{"why":"Provides the ResShift accelerated baseline that TSS outperforms with fewer steps.","marker":"[Yue et al., 2024]"},{"why":"Provides the one-step distillation baseline SinSR that TSS matches or beats without training.","marker":"[Wang et al., 2024b]"},{"why":"Provides the DM-NonUniform schedule-optimization baseline against which TSS is directly compared.","marker":"[Xue et al., 2024]"},{"why":"Provides the DDIM accelerated sampler with which TSS is combined in the appendix experiments.","marker":"[Song et al., 2020]"},{"why":"Provides the UniPC sampler with which TSS is combined to show compatibility with general acceleration.","marker":"[Zhao et al., 2024]"}],"fun_headline_variants":["Time-space aware sampling cuts diffusion SR steps in half","No-retrain step scheduling doubles diffusion SR speed","Adaptive denoising schedule halves diffusion SR iterations","Smarter diffusion SR: fewer steps, sharper textures"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that a pre-trained denoising network, trained with one global noise level per image, accepts and correctly interprets a spatially varying per-pixel timestep embedding, even when neighboring pixels are at very different noise levels.","fun_headline_variants_meta":{"raw":{"variants":["Time-space aware sampling cuts diffusion SR steps in half","No-retrain step scheduling doubles diffusion SR speed","Adaptive denoising schedule halves diffusion SR iterations","Smarter diffusion SR: fewer steps, sharper textures"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000786,"raw_usage":{"total_tokens":3468,"prompt_tokens":945,"completion_tokens":2523,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":2461}},"tokens_in":561,"tokens_out":2523,"duration_ms":15755,"temperature":1.0,"reasoning_tokens":2461,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:41:15.633396+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run TSS on a pre-trained diffusion super-resolution model while forcing large per-pixel timestep gaps between adjacent smooth and textured regions: if the spatially varying timestep embedding produces visible boundary artifacts or worse perceptual scores than uniform sampling, the central mechanism fails. Alternatively, measuring high-frequency SNR across a backbone outside the Stable Diffusion family and showing that the early/late two-stage pattern disappears would undercut the temporal motivation for TDS.","supporting_citations":[{"cited_title":"Pixel-aware stable diffu- sion for realistic image super-resolution and personalized stylization","cited_arxiv_id":null,"evidence_quote":"Supplies the PASD anchor model on which TSS is evaluated and whose 7-step baseline is improved."},{"cited_title":"Resshift: Efficient diffusion model for image super-resolution by residual shifting","cited_arxiv_id":null,"evidence_quote":"Provides the ResShift accelerated baseline that TSS outperforms with fewer steps."},{"cited_title":"Accelerating diffusion sampling with optimized time steps","cited_arxiv_id":null,"evidence_quote":"Provides the DM-NonUniform schedule-optimization baseline against which TSS is directly compared."},{"cited_title":"Unipc: A unified predictor- corrector framework for fast sampling of diffusion mod- els","cited_arxiv_id":null,"evidence_quote":"Provides the UniPC sampler with which TSS is combined to show compatibility with general acceleration."}],"review_version":1}