Research map / Architectures and efficiency / Quantization

Data-Free Group-Wise Fully Quantized Winograd Convolution via Learnable Scales

CVPR 2025 · Architectures and efficiency › Quantization · approach: Post-training quantization of ViTs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes data-free, group-wise fully quantized Winograd convolution for diffusion models and CNNs, learning only the Winograd transform scale factors on random Gaussian/uniform noise to reduce dynamic-range imbalance in the output transform.

Problem

Fully quantized Winograd convolution in diffusion models and CNNs suffers large quality loss because output-transform values have very imbalanced dynamic ranges, and existing fixes need costly training or domain-specific data that can overfit foundation models.

Approach

Group-wise quantize all Winograd pipeline stages (input transform, Hadamard product, output transform) and learn only the diagonal scale matrices S_B and S_G of the Vandermonde-based transforms, using an SQNR loss on random noise inputs with a single shared set of scales across layers, so no training data is needed.

Key results

• On InstaFlow-0.9B with AKL, W8A8 Winograd F(4,3) with learned scales reaches FID 24.51 vs 23.04 for W8A8 without Winograd, while standard-scale Winograd F(6,3) degrades to FID 326.96.
• On ImageNet with Winograd F(6,3), learned scales outperform PAW+FSQ by 1.62% (ResNet-18) and 2.56% (ResNet-34) top-1 accuracy.
• Optimized CPU kernels on Arm Graviton3 give a 31.3% convolution-layer runtime improvement from fully quantized Winograd and 12.8% end-to-end for InstaFlow-0.9B on a single thread.

Measured effects reported in the paper

Surprising findings

• Group-wise W4A8 on Stable Diffusion V1.5 gave a slightly better FID (20.62) than FP16 (21.72), which the authors attribute to FID unreliability.

Limitations

• Winograd F(6,3) with learned scales still shows a visible FID gap (26.58 vs 23.04 for W8A8 without Winograd on InstaFlow with AKL); the TAESD path is weaker for F(6,3) (27.69 vs 24.67).
• Results on ResNet-50 F(6,3) with learned scales (75.11%) are lower than PAW+FSQ (75.36%).
• Evaluation is limited to two diffusion models on COCO 5k and ResNets on ImageNet, with CPU-only runtime measurements.

Experimental setup

Datasets: MS-COCO 2017 (5k evaluation) ImageNet (evaluation) random Gaussian/uniform noise (scale learning)

Benchmarks: MS-COCO 2017 5k ImageNet

Metrics: FID CLIP score (ViT-g-14) top-1 accuracy runtime (wall-clock)

Baselines: Standard Winograd scales BQW PAW+FSQ group-wise W8A8 without Winograd FP16 model

Builds on: Winograd convolution Chinese remainder theorem / Vandermonde-based Winograd transforms group-wise quantization LLM group-wise quantization

Closest papers by meaning

• DGQ: Distribution-Aware Group Quantization for Text-to-Image Diffusion Models (ICLR 2025 · similarity 0.90)
• EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models (ICLR 2024 · similarity 0.88)
• TCAQ-DM: Timestep-Channel Adaptive Quantization for Diffusion Models (AAAI 2025 · similarity 0.87)
• Q-DiT: Accurate Post-Training Quantization for Diffusion Transformers (CVPR 2025 · similarity 0.86)
• ViDiT-Q: Efficient and Accurate Quantization of Diffusion Transformers for Image and Video Generation (ICLR 2025 · similarity 0.86)
• SVDQuant: Absorbing Outliers by Low-Rank Component for 4-Bit Diffusion Models (ICLR 2025 · similarity 0.86)