{"id":"286afc48-8341-44fc-9bf9-f4570ca0c35c","arxiv_id":"2412.09013","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"InvSR trains a noise predictor to initialize a frozen Stable Diffusion model at a high-signal intermediate step, delivering one-to-five-step super-resolution whose one-step output is competitive with dedicated one-step methods.","lead":"This paper trains a small 'noise predictor' network that, given a low-resolution image, chooses the noise added to the image's latent state so that a frozen Stable Diffusion model can denoise it into a high-resolution image in one to five steps. The method, InvSR, reports competitive visual quality at a fraction of the runtime of most diffusion-based super-resolution systems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (8) is dimensionally ambiguous: it mixes the pixel-space LR y0 with SD-Turbo's latent VAE space, and the paper never states the encoding; the one-step results rest on this undocumented step.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing point: Eq. (8) requires the LR input to already be in SD-Turbo's VAE latent space, yet the paper never states the encoding. My independent reading confirms this is the most serious issue in the central construction. The method otherwise has real supporting evidence: a released code and model, extensive ablations (loss terms, intermediate noise prediction, base model choice), and consistent benchmark results across three test sets. The mixed PSNR/SSIM versus perception metrics and the overbroad abstract wording are secondary; they qualify the strength of the empirical claim but do not threaten the mechanism as directly as the latent-domain ambiguity does. The latent encoding question is not a stylistic concern: if y0 is raw pixel input, then Eq. (8) combines a pixel-space image with latent-space noise under a pixel-space diffusion schedule, which is not a valid input for SD-Turbo's denoiser. If y0 is a latent, the missing encoder is a critical unstated component. The public code can likely settle this immediately, which is why I recommend keeping the reader's conditional verdict rather than escalating to rejection. The proposed test directly checks the implementation against the written equations, making the concern decisive either way.","tokens_in":18725,"tokens_out":6013,"duration_ms":64554,"concrete_test":"Run the released code and trace the tensor used as y0 at the line computing Eq. (8). If it is the output of the SD-VAE encoder (or any other encoder) rather than the raw LR image, rerun the ImageNet-Test one-step evaluation with the same encoder and verify that it reproduces the InvSR-1 row of Table 2 within expected metric noise. Then re-derive Eq. (8) in latent coordinates, e.g., x_t = sqrt(alpha_bar_t) E(y0) + sqrt(1-alpha_bar_t) f_w(E(y0), t), and confirm that substituting the raw pixel-space y0 breaks the pipeline. If the code uses an undocumented encoder, the paper must state it explicitly; if the code matches the literal pixel-space Eq. (8), feeding the result to SD-Turbo should fail or degrade, indicating the equations are not the implemented method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central construction Eq. (8) is never well-defined. SD-Turbo is a latent diffusion model; its denoiser epsilon_theta and the alpha_bar schedule are defined on VQGAN latents, not on RGB images. Yet the paper defines y0 as the LR image throughout Section 3 and never states that y0 in Eq. (8) is first VAE-encoded. If y0 is the pixel-space LR, then x_kappaM is a convex combination of an RGB image and a learned noise map of matching spatial size, while the subsequent call to epsilon_theta(x_kappaM, kappaM) requires a latent with different spatial dimensions and channels; the equations are dimensionally inconsistent and the learned f_w cannot produce a valid SD-Turbo input. If y0 is instead the VAE-encoded latent of the LR, the paper omits the encoder, and the stated two-downsampling-block architecture of f_w is not obviously compatible with SD's 8x-downsampled latent for the 128-to-512 task. The training objective Eq. (11) only references latent-space estimates and never clarifies the input domain of f_w, so the reported one-step results can rest on an undocumented preprocessing step. This is load-bearing because Eq. (8) is the sole mechanism connecting the LR observation to the frozen diffusion model; without a correct and stated definition of y0's domain, the diffusion-inversion framing and the one-step numbers are not reproducible from the paper alone.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes InvSR, a super-resolution method built on diffusion inversion with a frozen SD-Turbo backbone. A noise predictor f_w is trained to map the LR observation and a timestep to the noise that should be added to the LR to form the starting latent of a short reverse sampling chain (one to five steps). Intermediate steps use randomly sampled noise, and the training objective combines L2, LPIPS, and GAN losses on latent-space estimates. Experiments on ImageNet-Test, RealSR, and RealSet80 compare the method against GAN-based and diffusion-based SR baselines, and ablations in the supplement justify the partial-noise-prediction design.","tokens_in":19017,"tokens_out":7787,"duration_ms":77668,"significance":"If the construction is made precise, the contribution is practically valuable: a frozen diffusion backbone plus a 33.84M-parameter noise predictor achieves competitive perceptual quality at 117 ms for one-step ×4 SR while supporting a configurable number of sampling steps. The held-out evaluation on three datasets and the public release of code and models are strengths, and the ablations in Tables 4-6 of the supplement validate the main design choices. However, the central claim rests on Eq. (8), which as written is not well-defined because the domain of y0 is never specified; this must be resolved before the empirical results can be fully assessed.","major_comments":[{"comment":"The domain of y0 is never specified. Section 3 defines y0 as the LR image, and Eqs. (1)-(7) are written for DDPM tensors, but the denoiser εθ is SD-Turbo, a latent diffusion model whose α-bar schedule and noise prediction operate on VQGAN latents, not on RGB images. If y0 is the pixel-space LR, then x_{κM} in Eq. (8) is an RGB tensor while εθ(x_{κM}, κM) requires a 64×64 latent for the ×4 128-to-512 task; if y0 is instead the encoded latent, the encoder is omitted and the stated f_w architecture of two downsampling blocks does not obviously produce a 64×64 noise map from a 16×16 latent. Because Eq. (8) is the sole bridge from the LR observation to the frozen diffusion model, the one-step results are not reproducible from the paper alone. Please state the exact preprocessing of y0 and the input/output shapes of f_w, or rewrite all equations consistently in latent space.","section":"Sec. 3.2.2, Eq. (8), and Algorithm 1"},{"comment":"The text states that f_w predicts the mean and variance of a Gaussian via the reparameterization trick rather than directly estimating the noise map, but Eq. (8), Eq. (10), and Algorithm 1 all use f_w(y0, t) as the noise map itself. No sampling from the predicted distribution is specified, and the variance output is never referenced in the loss or inference. Please clarify whether f_w outputs a mean that is used directly as the noise map, or add the missing sampling/reparameterization step; otherwise the training objective Eq. (11) is under-specified.","section":"Sec. 3.2"},{"comment":"The paper says the whole loss is computed in the latent space and that the LPIPS loss is 'fine-tuned' in that space, but no definition is given for how LPIPS is applied to VQGAN latents or what fine-tuning means. Since LPIPS is a pixel-space perceptual metric, this sentence does not by itself define a computable loss. Provide the exact loss computation (which layers are used, what normalization is applied, and whether a separate network is trained), or move this clarification to the supplement.","section":"Sec. 3.2.3, Eq. (11), and training details"},{"comment":"The stochastic term σ_{τ_i} z_{τ_i} is never defined for the accelerated schedules used in the paper. For a diffusion step with τ_i = 250, a nonzero σ would inject substantial noise, so the reported 'one-step' result would include a stochastic perturbation at the final step, whereas the text frames one step as deterministic denoising. Specify the sampling algorithm (for example, DDIM with η = 0, or a specific σ schedule) and, if nonzero σ is used, state how the one-step numbers are obtained.","section":"Algorithm 1, line 4"}],"minor_comments":[{"comment":"The caption reads 'runtime (in milliseconds) highlighted by red in the sub-caption of the first example'; the phrasing is awkward and the parenthesis after 'Ours-2' is unclosed in the displayed text.","section":"Figure 1 caption"},{"comment":"The header 'Index of the sampled timesteps' should be 'Index of the sampling timesteps' for clarity.","section":"Table 1 header"},{"comment":"The expression f_w(y_{κ_i}, κ_i) should presumably be f_w(y_0, κ_i), since y_{κ_i} is never defined.","section":"Sec. 3.2.3"},{"comment":"The product notation for \\barα_t is garbled in the typeset version; it should be \\barα_t = \\prod_{s=1}^t α_s.","section":"Eq. (2)"},{"comment":"The sentence attributing textual inversion to 'Rinon et al.' should cite 'Gal et al.' [12].","section":"Related Work"},{"comment":"The phrase 'arbitrary number of sampling steps ranging from one to five' is slightly misleading; consider replacing 'arbitrary' with 'configurable' to avoid overclaiming.","section":"Abstract and Introduction"}],"recommendation":"major_revision","confidential_remarks":"The paper's empirical results may be valid, but the core equations need to be made precise before I can endorse publication. In particular, the ambiguity about whether y0 in Eq. (8) is an RGB image or a VAE latent is unusual given the authors' expertise; I suspect the implementation encodes the LR image and the text was written before that decision was finalized. I would encourage the editor to request that the authors state the latent-space encoding explicitly and, if possible, verify that the released code matches the corrected equations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a practical one-step-to-five-step diffusion SR method that does what it claims on held-out benchmarks, and the core idea—train a compact noise predictor to initialize a frozen SD-Turbo at a high-SNR latent—is genuinely new. The paper is worth a serious referee, but a couple of technical details around the central equation need fixing before publication.\n\nThe strength is the empirical package. They evaluate on a 3,000-image held-out set (ImageNet-Test) plus two real-world sets, with seven metrics and a fair set of baselines. The ablations in the supplement (loss weightings, choice of base diffusion model, and the unnecessary intermediate noise predictor) are the right ones, and they integrate cleanly with the story. Runtime of 117ms on an A100 for x4 is a concrete deployment advantage. Code and model are public.\n\nThe soft spots are real but not disqualifying. Most important: Eq. (8) is the load-bearing construction, and the paper never states whether y0 is the pixel-space LR or the VAE-encoded latent. The text says the loss is computed in latent space for SD-Turbo, but Eq. (8) itself is presented in generic DDPM notation without a single sentence tying y0 to the latent encoder. The described noise-predictor architecture (two downsampling blocks) doesn't obviously produce a 64x64 latent from a 128x128 input either. This is the kind of missing detail that makes reproduction harder than it should be, and it needs to be stated explicitly, not left to the code.\n\nTwo smaller issues. The 'optimal noise map' language overstates what a trained fitting procedure can claim; it's a learned initialization, not a derivation. And the abstract's 'superior or comparable' only holds for perceptual metrics—InvSR-1's PSNR is lower than GAN baselines and ResShift, which the paper itself shows. Also, the recent distillation-free one-step method (arXiv:2410.04224) is cited but not compared; that should be added. None of this overturns the central result; it just needs calibration.\n\nBottom line: send this to reviewers. The idea is novel, the experiments are substantial, and the ambiguity around Eq. (8) is fixable. I'd want to see a revised version that clarifies the latent encoding and softens the claims, but the work is solid enough to justify the round.","headline":"A genuinely practical one-to-five-step diffusion SR method with solid held-out benchmarks and one load-bearing clarity issue that should be fixed, not fatal.","tokens_in":19595,"tokens_out":5014,"would_cite":true,"duration_ms":48664,"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":"This paper claims that image super-resolution can be done from a frozen diffusion model by training a small noise predictor to construct a single intermediate starting state, enabling one-to-five-step sampling with single-step quality…","keywords":["diffusion inversion","image super-resolution","partial noise prediction","noise predictor","arbitrary-step sampling","Stable Diffusion","blind super-resolution","one-step diffusion"],"falsifier":"A direct falsifier is a controlled swap: in Eq. (8) at the trained starting steps, replace the predictor's output with zero-mean Gaussian noise of the same standard deviation, keep the frozen SD-Turbo sampler unchanged, and evaluate on ImageNet-Test; if one-step PSNR and LPIPS do not clearly worsen, the learned noise map is not what carries the super-resolution result, and a reproduction that fails to match Table 2's one-step numbers would likewise refute the headline claim.","tokens_in":18461,"feed_emoji":"🖼️","tokens_out":7355,"duration_ms":76813,"temperature":0.7,"pith_summary":"InvSR claims that image super-resolution can be reduced to choosing one good noise map. Instead of running a large diffusion model from pure noise or fine-tuning it, the method trains a compact 33.84M-parameter noise predictor that, given a low-resolution image and a starting timestep, produces the noise needed to build an intermediate diffusion state. From that state, a completely frozen Stable Diffusion Turbo model generates the high-resolution image in one to five sampling steps. The paper reports that even a single step beats or matches recent dedicated one-step diffusion SR methods across perceptual and distortion metrics, while allowing the user to trade fidelity for realism by changing the starting timestep. If this holds, the practical cost of diffusion-based SR drops sharply while keeping the large pretrained model untouched.","feed_headline":"One noise map turns frozen diffusion into one-step super-resolution","feed_subtitle":"A 34M-parameter predictor seeds SD-Turbo at any timestep, beating one-step SR rivals in 117 ms.","key_machinery":"The load-bearing object is the Partial Noise Prediction starting-state formula in Eq. (8): $x_{\\kappa_M} = \\sqrt{\\bar\\alpha_{\\kappa_M}}\\,y_0 + \\sqrt{1-\\bar\\alpha_{\\kappa_M}}\\,f_w(y_0, \\kappa_M)$. It collapses the full inversion problem, which would require predicting $T$ noise maps, into a single noise map for one chosen intermediate timestep, and it is the only place where the low-resolution image enters the diffusion trajectory. The noise predictor $f_w$, built on a VQGAN encoder with two downsampling blocks and self-attention, is trained to output Gaussian mean and variance parameters via the reparameterization trick, and its time embedding is what makes the starting timestep selectable at inference. Intermediate sampling steps are then ordinary DDIM/DDPM updates with random noise, relying on the high signal-to-noise ratio of timesteps at or below 250 to keep the frozen SD-Turbo denoiser on track.","core_discovery":"The central claim is that the reverse diffusion process for super-resolution does not need per-step noise optimization or backbone fine-tuning. The paper's construction starts from the DDPM marginal $x_\\tau = \\sqrt{\\bar\\alpha_\\tau}\\,x_0 + \\sqrt{1-\\bar\\alpha_\\tau}\\,\\xi$ and replaces the unknown high-resolution image $x_0$ with the given low-resolution image $y_0$, using a trained network $f_w$ to supply the noise: $x_\\tau = \\sqrt{\\bar\\alpha_\\tau}\\,y_0 + \\sqrt{1-\\bar\\alpha_\\tau}\\,f_w(y_0, \\tau)$. The predictor is time-conditioned and trained with L2, LPIPS, and GAN losses in the latent space of SD-Turbo, with starting timesteps 250, 200, 150, and 100; at inference the user picks a starting step and runs one to five DDIM-style steps. The paper's evidence is that InvSR-1 reaches 24.14 dB PSNR and 0.2517 LPIPS on ImageNet-Test, outperforming the one-step OSEDiff baseline on all seven reported metrics, and that the predicted noise map is visibly correlated with the low-resolution image, which the authors interpret as showing that the predictor has found an LR-dependent inversion trajectory.","pith_inferences":["An extension the paper leaves implicit is to make the starting step self-tuning: feed an automatic blur or noise estimate into the timestep selection so users do not have to choose the step count manually.","Since the predicted noise map is visibly correlated with the low-resolution image, the predictor can be viewed as injecting the missing high-frequency residual into the latent; the same construction could seed other conditional restoration tasks, such as deblurring or deraining, with a frozen diffusion backbone.","The paper does not state how the low-resolution image is encoded into SD-Turbo's latent space; a natural reading is that the same VQGAN encoder is reused, and an ablation replacing that encoder with a differently trained one would reveal how much of the super-resolution quality comes from the frozen prior versus the predictor."],"forward_implications":["One-step super-resolution becomes a practical setting: InvSR-1 runs in 117 ms on an A100 and beats the dedicated one-step OSEDiff on all seven metrics reported on ImageNet-Test.","The number of sampling steps can be chosen per image, so noise-dominated images can use one step while blur-dominated images can use three to five steps to recover finer structure.","Because the diffusion backbone stays frozen, improvements in the base generative model can be inherited by retraining only the small noise predictor.","Reference metrics decline as sampling steps increase, so the multi-step mode is a realism tool rather than a simple accuracy booster."],"supporting_citations":[{"why":"Supplies the DDPM forward marginal $q(x_t|x_0)$ that Eqs. (7) and (8) build the starting state from.","marker":"[16]"},{"why":"Defines the latent diffusion framework and the Stable Diffusion model whose latent space the method operates in.","marker":"[43]"},{"why":"Supplies SD-Turbo, the frozen denoising backbone that performs the actual reverse sampling.","marker":"[44]"},{"why":"Provides DDIM, the accelerated deterministic sampler used to run one to five inference steps.","marker":"[46]"},{"why":"Supplies VQGAN, whose encoder architecture is reused for the noise predictor and whose latent space carries the whole pipeline.","marker":"[10]"},{"why":"Provides the RealESRGAN degradation pipeline used to synthesize the training low-resolution and high-resolution pairs.","marker":"[55]"},{"why":"The dedicated one-step diffusion super-resolution baseline that InvSR claims to surpass on all seven metrics on ImageNet-Test.","marker":"[59]"},{"why":"The one-step SinSR baseline that InvSR compares against in the single-step evaluation.","marker":"[57]"},{"why":"Supplies the ResShift degradation settings for the ImageNet-Test benchmark and the four-step ResShift baseline.","marker":"[69]"}],"fun_headline_variants":["One learned noise map predicts LR-dependent path for super-resolution","Train a noise predictor once, then do super-resolution in 1–5 steps","Frozen diffusion backbone, arbitrary steps: just predict the noise","Super-resolution without per-step optimization: predict noise, sample","Single-step SR beats OSEDiff by predicting the right noise map"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The one-step result rests on the assumption that a low-resolution image perturbed by the learned noise map lands inside the narrow distribution of latent states from which the frozen denoiser can recover the high-resolution image.","fun_headline_variants_meta":{"raw":{"variants":["One learned noise map predicts LR-dependent path for super-resolution","Train a noise predictor once, then do super-resolution in 1–5 steps","Frozen diffusion backbone, arbitrary steps: just predict the noise","Super-resolution without per-step optimization: predict noise, sample","Single-step SR beats OSEDiff by predicting the right noise map"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000239,"raw_usage":{"total_tokens":1530,"prompt_tokens":976,"completion_tokens":554,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":592,"completion_tokens_details":{"reasoning_tokens":465}},"tokens_in":592,"tokens_out":554,"duration_ms":5906,"temperature":1.0,"reasoning_tokens":465,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:21:50.546987+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct falsifier is a controlled swap: in Eq. (8) at the trained starting steps, replace the predictor's output with zero-mean Gaussian noise of the same standard deviation, keep the frozen SD-Turbo sampler unchanged, and evaluate on ImageNet-Test; if one-step PSNR and LPIPS do not clearly worsen, the learned noise map is not what carries the super-resolution result, and a reproduction that fails to match Table 2's one-step numbers would likewise refute the headline claim.","supporting_citations":[{"cited_title":"Denoising dif- fusion probabilistic models","cited_arxiv_id":null,"evidence_quote":"Supplies the DDPM forward marginal $q(x_t|x_0)$ that Eqs. (7) and (8) build the starting state from."},{"cited_title":"High-resolution image syn- thesis with latent diffusion models","cited_arxiv_id":null,"evidence_quote":"Defines the latent diffusion framework and the Stable Diffusion model whose latent space the method operates in."},{"cited_title":"Adversarial diffusion distillation","cited_arxiv_id":null,"evidence_quote":"Supplies SD-Turbo, the frozen denoising backbone that performs the actual reverse sampling."},{"cited_title":"Denois- ing diffusion implicit models","cited_arxiv_id":null,"evidence_quote":"Provides DDIM, the accelerated deterministic sampler used to run one to five inference steps."},{"cited_title":"Taming transformers for high-resolution image synthesis","cited_arxiv_id":null,"evidence_quote":"Supplies VQGAN, whose encoder architecture is reused for the noise predictor and whose latent space carries the whole pipeline."},{"cited_title":"Real-esrgan: Training real-world blind super-resolution with pure synthetic data","cited_arxiv_id":null,"evidence_quote":"Provides the RealESRGAN degradation pipeline used to synthesize the training low-resolution and high-resolution pairs."},{"cited_title":"One-step effective diffusion network for real-world image super-resolution","cited_arxiv_id":null,"evidence_quote":"The dedicated one-step diffusion super-resolution baseline that InvSR claims to surpass on all seven metrics on ImageNet-Test."},{"cited_title":"Sinsr: diffusion-based image super- resolution in a single step","cited_arxiv_id":null,"evidence_quote":"The one-step SinSR baseline that InvSR compares against in the single-step evaluation."},{"cited_title":"Effi- cient diffusion model for image restoration by residual shift- ing","cited_arxiv_id":null,"evidence_quote":"Supplies the ResShift degradation settings for the ImageNet-Test benchmark and the four-step ResShift baseline."}],"review_version":1}