{"id":"2a01454a-12a1-47e0-a29b-66d6414ebaef","arxiv_id":"2412.19867","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A data-free method that tunes only the diagonal scales of Winograd transform matrices enables accurate 8-bit fully quantized Winograd convolution for diffusion models and ResNets.","lead":"This paper shows that fully quantized Winograd convolutions can be made accurate on large diffusion models by tuning only the scale factors of the Winograd transform matrices, using random noise instead of real training data. It reports near-lossless 8-bit image generation quality and significantly faster CPU runtime.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The data-free generalization claim overreaches: scales tuned on random noise may not minimize quantization error on real activations, and the paper supplies no theoretical guarantee that they do.","rationale":"The reader's weakest assumption is exactly the load-bearing point: random noise must be a sufficient proxy for real activations when tuning quantization scales. My analysis refines this by identifying the specific mechanism — the per-tap dynamic range profile of Y — that the learned scales are meant to equalize, and by noting that the paper's own calibration-data experiment (Appendix Table 8) shows the optimization is distribution-sensitive. The paper provides no theoretical guarantee, and the 'safely guaranteed' statement in the abstract is unsupported. However, the empirical evidence is substantial: the method works across diffusion models, ResNets, and transferred scales on CIFAR10 and MJHQ. Thus the central claim is likely correct for the tested regimes, but the generalization claim is overstated and the method's robustness to input distribution is unverified. A conditional verdict is appropriate, matching the reader's judgment; the concern does not change the verdict but underscores the missing evidence. The concrete test would settle whether the concern actually lands by directly probing the proxy assumption.","tokens_in":18490,"tokens_out":7662,"duration_ms":69824,"concrete_test":"For a fixed ResNet18 F(6,3) setup, learn Winograd scales three ways using the same Algorithm 1 but with different input distributions: (a) standard Gaussian noise as in the paper, (b) uniform noise on [-1,1], (c) Gaussian noise whose per-layer variance is matched to the activation statistics of 64 ImageNet images (used only to choose the noise distribution, not in the loss). Evaluate ImageNet top-1 accuracy for each. If the spread exceeds ~1%, the random-noise proxy is not distribution-robust and the data-free generalization claim fails. Additionally, compute the per-tap standard deviation of the Winograd-domain output Y (Eq. 5) on random noise inputs and on real validation inputs for a representative layer; report the Spearman rank correlation. If the correlation is below 0.9, the optimization target is mismatched with the quantity that determines real quantization error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that tuning only the diagonal scale matrices SB and SG on random Gaussian/uniform noise (Eq. 17, Algorithm 1) fully restores quality of fully quantized Winograd convolution. For this to hold, the noise distribution used in optimization must be an adequate proxy for real activation distributions when assessing quantization error. The paper asserts this is 'safely guaranteed' because no training data is used, but that is a non-sequitur: the method still optimizes against a specific synthetic input distribution, and could overfit to it. No formal bound or analysis is provided, despite the promise of a 'theoretically grounded analysis' in Section 1. The evidence is purely empirical transfer (CIFAR10, MJHQ), which is suggestive but not a guarantee. A sharper failure mode: the objective is per-layer SQNR on i.i.d. noise, but the quantization error that matters is the tap-wise dynamic range of Y (Eq. 5) under real activations, which have correlations, varying magnitudes, and time-step-dependent statistics in diffusion models. If the per-tap variance profile of Y for real inputs differs from that for random noise, the learned scales will not minimize the actual error. Notably, Appendix Table 8 shows that using 10k real prompts for calibration yields worse F(6,3) FID (36.18) than using random noise (26.58), suggesting the optimization is sensitive to the input distribution and that 'data-free' does not automatically mean distribution-robust. The claim would be load-bearing for the paper's headline contribution; if the noise-proxy assumption fails, the method's universality collapses to the tested configurations.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a data-free method for fully quantized Winograd convolution. The key ideas are (i) group-wise quantization applied to all stages of the Winograd pipeline (input transform, Hadamard product, output transform) and (ii) learning only the diagonal scale matrices S_B and S_G of the Winograd transform matrices, using random Gaussian or uniform noise inputs and a per-layer SQNR objective, with a single shared scale set across layers. The method is evaluated on InstaFlow-0.9B and Stable Diffusion v1.5 for text-to-image generation and on ResNet-18/34/50 for ImageNet classification. The authors report near-lossless W8A8 generation quality with fully quantized Winograd F(4,3) and F(6,3), top-1 accuracies close to full precision, and CPU kernel optimizations that yield a 31.3% convolution-layer speedup.","tokens_in":18782,"tokens_out":8689,"duration_ms":72232,"significance":"If the empirical results hold, the paper makes a useful practical contribution: a simple, calibration-free recipe for recovering quality in fully quantized Winograd convolution, which is relevant for on-device deployment of diffusion models. The strengths include evaluation across generation and classification tasks, transfer experiments to CIFAR-10 and MJHQ, release of code and kernels, and a method that is easy to describe and potentially easy to adopt. The main caveat is that the paper frames the data-free property as a theoretical generalization guarantee, which is not established by the presented analysis. The missing optimization details also need to be addressed before the method can be independently reproduced.","major_comments":[{"comment":"The paper does not specify how gradients are computed through the non-differentiable quantization function Q used in Eqs. (14)-(16). Because the objective in Eq. (17) depends on quantized outputs, backpropagation requires an approximation such as a straight-through estimator or a similar relaxation. Without this detail, the optimization procedure is underspecified, and the reported results cannot be reproduced or independently evaluated for correctness.","section":"Section 4.2, Eq. (17) and Algorithm 1"},{"comment":"The statement that generalization is 'safely guaranteed' because the method uses no training data is not justified. The method optimizes scales against random Gaussian/uniform noise inputs, which is itself a specific distribution; if the per-tap variance profile of the Winograd-domain output Y under real activations differs from that under noise, the learned scales may not minimize the actual quantization error. The 'theoretically grounded analysis' in Section 1 is a heuristic about row norms of B and G, not a generalization bound. The transfer results in Section 11 are empirical and suggestive, but they do not constitute the invoked guarantee. Please temper the claim or provide a principled analysis of when noise is a sufficient proxy.","section":"Abstract and Section 1"},{"comment":"The optimization target is per-layer SQNR on synthetic noise, whereas the reported metrics are FID, CLIP, and top-1 accuracy. No evidence is given that minimizing this proxy on noise translates to reduced quantization error under real activation distributions, especially for diffusion models where activations change across sampling steps. The assertion in Section 2 that group-wise quantization is 'inherently more robust to distribution changes' does not cover the learned-scale optimization. An ablation comparing the per-tap dynamic range of Y under noise versus real inputs (using a small calibration set for analysis only) would strengthen the argument.","section":"Section 4.2, Eq. (17)"},{"comment":"Key optimization hyperparameters are missing: learning rate, number of epochs, batch size, noise type and magnitude, the number of layers K selected per iteration, and how the shared scale set is updated when different layers are sampled across iterations. The paper also does not state the range or variance of the random noise used. These details are essential for reproducibility and for assessing the sensitivity of the method to optimization choices.","section":"Section 5.1 and Algorithm 1"},{"comment":"The comparison between learning with noise and learning with calibration data is confounded because the calibration-data runs use end-to-end training while the noise runs use layer-wise training. The F(6,3) FID gap (36.18 vs 26.58) therefore cannot be attributed to the input distribution alone. A matched comparison (layer-wise training with calibration inputs) is needed, or the confound should be acknowledged explicitly.","section":"Section 9, Table 8"}],"minor_comments":[{"comment":"The heading contains a typo: 'qantized' should be 'quantized'.","section":"Section 4 heading"},{"comment":"The alignment of row labels ('Standard scales', 'Learned scales') with the tile sizes is unclear; please reformat the tables so that each row's condition is explicit.","section":"Tables 1-4"},{"comment":"The paper leaves open whether Gaussian or uniform noise was used for the reported results; specify the distribution and its magnitude for reproducibility.","section":"Section 4.2"},{"comment":"The notation 'sqB sqB sqx' is terse and ambiguous; define the individual scale factors (e.g., left and right transform scales and input scale) more explicitly.","section":"Eqs. (14)-(16)"},{"comment":"The polynomial points are attributed to [2], but the standard points are conventionally taken from [15] and [41]; please adjust the citation to point to the original sources.","section":"Section 7.1"},{"comment":"The standard Winograd F(6,3) accuracy for ResNet-18 is 0.08%, which is near chance; a brief explanation of why the baseline collapses would help readers appreciate the challenge being addressed.","section":"Table 5"}],"recommendation":"major_revision","confidential_remarks":"The empirical contribution is solid and the paper is likely to be of interest to the deployment/quantization community. The main risk is the overclaimed data-free generalization guarantee; I would encourage the editors to ask the authors to either soften the guarantee or support it with analysis. The missing straight-through estimator detail is a concrete technical gap that must be fixed. I did not independently verify the baseline numbers taken from prior work (BQW, PAW+FSQ)."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core result is real and worth knowing: for fully quantized 8-bit Winograd convolution, tuning only the diagonal scale factors of the transform matrices—using random Gaussian or uniform noise as inputs—restores most of the quality lost by naive quantized Winograd. On InstaFlow-0.9B, FID drops from 217 to 24.5 for F(4,3) and from 327 to 26.6 for F(6,3), close to the FP16 23.0. On ResNet-18/34, it beats PAW+FSQ by 1.6-2.6 points on F(6,3). This is, as far as I know, the first full quantization of Winograd for large diffusion models, and the CPU kernel speedup (31.3% on conv layers) is a useful bonus.\n\nWhat is genuinely new: prior work learned full transform matrices with calibration data (PAW+FSQ) or quantized only part of the pipeline (BQW, LowINO). Learning only the diagonal scales, sharing one set across all layers, and doing it data-free is a clean twist that works well in the tested settings. I find the empirical evidence convincing for these cases.\n\nNow the soft spots, in proportion. The largest is the generalization claim. The paper says that because no training data is used, generalization is \"safely guaranteed.\" That is not a logical consequence: the scales are still fit to a synthetic input distribution, and no formal bound is supplied. The stress-test note points out an odd result in the appendix: calibrating with 10k real prompts gives a worse F(6,3) FID (36.18) than training on noise (26.58). I do not think this proves the noise proxy is wrong—the end-to-end calibration objective is likely harder to optimize—but it does show sensitivity to the optimization setup, undercutting any sweeping guarantee. The transfer results to CIFAR10 and MJHQ are good evidence, but evidence is not a guarantee.\n\nOther issues: the \"theoretically grounded analysis\" promised in the introduction is not delivered; the row-norm argument is heuristic. The paper never says how gradients flow through the non-differentiable quantization (straight-through estimator is not mentioned), nor does it give hyperparameters for Algorithm 1 (epochs, batches, learning rate). Without these, reproduction is harder than it should be. I also think \"near-lossless\" is too strong: InstaFlow F(6,3) loses 3.5 FID points, small but not negligible. And the ImageNet state-of-the-art claim is worded carefully but glosses over the fact that BQW beats this method on ResNet50 for both tile sizes.\n\nWho this is for: people working on low-precision inference for diffusion models, especially on CPUs. It deserves a serious referee, not a desk reject. I would send it to review with a request to add STE and hyperparameter details, soften the generalization language, and discuss the ResNet50 numbers. The central method is likely correct and useful; the paper just needs to match its claims to the evidence.","headline":"Tuning only the Winograd scale factors on random noise largely fixes fully quantized Winograd on diffusion models, but the 'safely guaranteed' generalization claim overreaches and key implementation details are missing.","tokens_in":19331,"tokens_out":5980,"would_cite":true,"duration_ms":62279,"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":"Tuning Winograd scale matrices on random noise makes fully quantized 8-bit Winograd convolution near-lossless for diffusion models.","keywords":["post-training quantization","Winograd convolution","group-wise quantization","diffusion models","learnable scales","data-free quantization","8-bit inference","CPU kernels"],"falsifier":"Take a model and learn the Winograd scales on random noise, then evaluate on inputs whose activations have extreme dynamic range, such as early diffusion timesteps, very high-resolution images, or a dataset with heavy-tailed outliers; if FID or top-1 accuracy drops substantially compared with scales learned on a small real calibration set, the data-free proxy is falsified. A more direct check is to compare the $8\\times8$ Winograd-domain output range ratios for noise versus real activations and see whether the learned scales compress both.","tokens_in":18266,"feed_emoji":"🖼️","tokens_out":11619,"duration_ms":93480,"temperature":0.7,"pith_summary":"This paper tries to establish that fully quantized 8-bit Winograd convolution—a fast-convolution method normally too lossy to use in quantized neural networks—can be made near-lossless by retuning only the diagonal scale factors of the Winograd transform, with no real training or calibration data. The scales are learned on random Gaussian or uniform noise, and the same learned scales are shared across all convolution layers of a model. If the claim holds, large text-to-image diffusion models could use 8-bit weights and activations together with Winograd $F(6,3)$ speedups while keeping FID and CLIP scores close to the full-precision model. The paper demonstrates this on InstaFlow-0.9B and Stable Diffusion v1.5, and shows ImageNet top-1 accuracy gains over prior Winograd post-training quantization methods, along with CPU kernel runtime improvements.","feed_headline":"Noise-learned scales restore 8-bit Winograd diffusion quality","feed_subtitle":"Fully quantized Winograd convolution regains near-FP16 FID and beats prior PTQ by up to 2.56% on ImageNet.","key_machinery":"The load-bearing object is the Vandermonde factorization of Winograd transforms, $B^T = S_B V^{-T}$, $G = S_G V$, $A^T = V^T S_A$, with diagonal scale matrices constrained by $S_A S_B S_G = I$. The method learns only $S_B$ and $S_G$ ($S_A$ follows from the constraint) by minimizing the SQNR loss between the quantized Winograd output and the full-precision convolution output on random noise inputs. This reweights the row norms of the transform matrices, compressing the cross-shaped dynamic-range spread in the Winograd-domain output $Y$ that group-wise quantization cannot absorb, and it does so without altering the algorithmic structure or requiring calibration data.","core_discovery":"The central discovery is that the severe accuracy drop in fully quantized Winograd convolution comes from the dynamic-range imbalance of the Winograd-domain output tile $Y$, not from quantization of the input transform or the Hadamard product. Group-wise quantization handles those stages well, but the output transform sees a cross-shaped pattern of very different magnitudes across the tile's taps, and a single or per-row scale cannot represent it. The paper shows this imbalance can be largely removed by learning only the diagonal scaling matrices $S_B$ and $S_G$ of the Vandermonde-based Winograd transforms, with $S_A$ determined by the identity $S_A S_B S_G = I$; the full transformation matrices themselves are never finetuned. Because the objective is minimized on random noise inputs, the learned scales transfer across datasets, and the result is 8-bit fully quantized Winograd $F(4,3)$/$F(6,3)$ that nearly matches full-precision image generation and classification quality.","pith_inferences":["The same mechanism could apply to other fast transforms whose intermediate domains have uneven magnitude spreads, such as FFT-based convolution or Strassen-like matrix multiplication, since the analysis identifies Vandermonde row-norm structure rather than Winograd specifically as the culprit.","Using a single set of scales shared by all layers implies a much smaller parameter search than per-layer calibration; a natural test is whether layer-specific scales further close the small remaining $F(6,3)$ gap without hurting transfer.","The data-free claim is only as strong as the noise proxy; a stress test on distributions with heavy-tailed outliers, such as very early denoising steps or high-resolution inputs, would show whether random noise is truly sufficient or merely adequate on the tested models."],"forward_implications":["8-bit fully quantized Winograd $F(4,3)$ and $F(6,3)$ on InstaFlow-0.9B and Stable Diffusion v1.5 keep FID and CLIP scores near the FP16 baseline, whereas standard group-wise Winograd collapses FID from about 22 to above 300.","On ImageNet, learned scales beat the PAW+FSQ baseline by 1.62% top-1 accuracy on ResNet-18 and 2.56% on ResNet-34 for Winograd $F(6,3)$, restoring most of the full-precision accuracy.","The same learned scales transfer across datasets without retraining: ImageNet-learned scales recover CIFAR-10 accuracy, and COCO-learned scales work on the MJHQ prompt set.","Highly optimized group-wise quantized Winograd kernels add a 31.3% runtime reduction on convolution layers and a 12.8% end-to-end speedup for InstaFlow-0.9B on CPU hardware.","Because only scales are tuned, the method avoids the cost of finetuning whole Winograd transformation matrices, which prior work found necessary for full quantization."],"supporting_citations":[{"why":"Defines the Winograd convolution algorithm F(m,r) whose full 8-bit quantization is the paper's target.","marker":"[15]"},{"why":"Derives the Winograd transforms from Vandermonde matrices with diagonal scaling matrices satisfying S_A S_B S_G = I, the structure the method learns.","marker":"[41]"},{"why":"BQW is the channel-balancing Winograd quantization baseline whose partial quantization the paper extends and outperforms.","marker":"[6]"},{"why":"PAW+FSQ is the calibration-based Winograd finetuning baseline the paper beats in ImageNet accuracy.","marker":"[5]"},{"why":"Q-Diffusion is the diffusion-specific PTQ baseline compared against group-wise quantization for text-to-image quality.","marker":"[17]"},{"why":"Prior work that learns full Winograd transformation matrices in a QAT setup; the paper contrasts its scale-only data-free learning with this costly approach.","marker":"[10]"},{"why":"InstaFlow-0.9B is one of the two diffusion models used to demonstrate near-lossless generation quality.","marker":"[21]"},{"why":"Stable Diffusion v1.5 is the second diffusion model used to demonstrate near-lossless FID and CLIP scores.","marker":"[30]"}],"fun_headline_variants":["Data-free scale tuning unlocks 8-bit Winograd for diffusion","Noise-only scale learning restores quantized Winograd quality","Group-wise scales bridge Winograd domain imbalance in 8-bit","Near-lossless 8-bit diffusion via noise-learned Winograd scales","Beat prior Winograd PTQ by 2.56% with data-free scale tuning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that random Gaussian or uniform noise captures the dynamic-range structure of real activations well enough that scales learned on noise transfer to actual images, so if real activation distributions differ in ways noise cannot mimic, the quality recovery would not generalize.","fun_headline_variants_meta":{"raw":{"variants":["Data-free scale tuning unlocks 8-bit Winograd for diffusion","Noise-only scale learning restores quantized Winograd quality","Group-wise scales bridge Winograd domain imbalance in 8-bit","Near-lossless 8-bit diffusion via noise-learned Winograd scales","Beat prior Winograd PTQ by 2.56% with data-free scale tuning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000642,"raw_usage":{"total_tokens":3025,"prompt_tokens":1088,"completion_tokens":1937,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":704,"completion_tokens_details":{"reasoning_tokens":1858}},"tokens_in":704,"tokens_out":1937,"duration_ms":15881,"temperature":1.0,"reasoning_tokens":1858,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:13:11.292265+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a model and learn the Winograd scales on random noise, then evaluate on inputs whose activations have extreme dynamic range, such as early diffusion timesteps, very high-resolution images, or a dataset with heavy-tailed outliers; if FID or top-1 accuracy drops substantially compared with scales learned on a small real calibration set, the data-free proxy is falsified. A more direct check is to compare the $8\\times8$ Winograd-domain output range ratios for noise versus real activations and see whether the learned scales compress both.","supporting_citations":[{"cited_title":"Fast algorithms for convolu- tional neural networks","cited_arxiv_id":null,"evidence_quote":"Defines the Winograd convolution algorithm F(m,r) whose full 8-bit quantization is the paper's target."},{"cited_title":"Frumkin, Boris Ginsburg, and Julien Demouth","cited_arxiv_id":null,"evidence_quote":"Derives the Winograd transforms from Vandermonde matrices with diagonal scaling matrices satisfying S_A S_B S_G = I, the structure the method learns."},{"cited_title":"Channel bal- ancing for accurate quantization of winograd convolutions","cited_arxiv_id":null,"evidence_quote":"BQW is the channel-balancing Winograd quantization baseline whose partial quantization the paper extends and outperforms."},{"cited_title":"Towards efficient and accurate winograd convolution via full quantization","cited_arxiv_id":null,"evidence_quote":"PAW+FSQ is the calibration-based Winograd finetuning baseline the paper beats in ImageNet accuracy."},{"cited_title":"Q-diffusion: Quantizing diffusion models","cited_arxiv_id":null,"evidence_quote":"Q-Diffusion is the diffusion-specific PTQ baseline compared against group-wise quantization for text-to-image quality."},{"cited_title":"Searching for winograd- aware quantized networks","cited_arxiv_id":null,"evidence_quote":"Prior work that learns full Winograd transformation matrices in a QAT setup; the paper contrasts its scale-only data-free learning with this costly approach."},{"cited_title":"Instaflow: One step is enough for high-quality diffusion-based text-to-image generation","cited_arxiv_id":null,"evidence_quote":"InstaFlow-0.9B is one of the two diffusion models used to demonstrate near-lossless generation quality."},{"cited_title":"High-resolution image synthesis with latent diffusion models","cited_arxiv_id":null,"evidence_quote":"Stable Diffusion v1.5 is the second diffusion model used to demonstrate near-lossless FID and CLIP scores."}],"review_version":1}