Research map / Image generation / Fast diffusion sampling
HQ-DM: Single Hadamard Transformation-Based Quantization-Aware Training for Low-Bit Diffusion Models
Shizhuo Mao, Hongtao Zou, Qihu Xie, Song Chen, Yi Kang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Single instead of Double Hadamard transformation (ablation) | HQ-DM W4A4 · ImageNet 256x256 LDM-4 | IS: 4.37 -> 248.30 | Improved |
| use Single instead of Double Hadamard transformation (ablation) | HQ-DM W4A4 · ImageNet 256x256 LDM-4 | FID: 179.40 -> 6.58 | Improved |
| use Single instead of Double Hadamard transformation (ablation) | HQ-DM W8A8 · ImageNet 256x256 LDM-4 | latency: 1.0688 -> 0.9419 s/sample (1.13x) | Improved |
| add Hadamard transform to naive PTQ + LSQ (ablation) | Naive PTQ · ImageNet 256x256 LDM-4 W4A4 | IS: 5.09 -> 33.29 | Improved |
| add Hadamard transform to naive PTQ + LSQ (ablation) | Naive PTQ · ImageNet 256x256 LDM-4 W4A4 | FID: lowering the FID by 125.43 | Improved |
| add Hadamard transform to naive QAT + LSQ (ablation) | Naive QAT · ImageNet 256x256 LDM-4 W4A4 | IS: 159.24 -> 203.69 | Improved |
| add Hadamard transform to naive QAT + LSQ (ablation) | Naive QAT · ImageNet 256x256 LDM-4 W4A4 | sFID: 8.38 -> 10.98 | Hurt |
| add Hadamard transform to LoRA only (ablation) | LoRA distillation · ImageNet 256x256 LDM-4 W4A4 | FID: 8.55 -> 6.52 | Improved |
| add Hadamard transform to LoRA only (ablation) | LoRA distillation · ImageNet 256x256 LDM-4 W4A4 | IS: 185.51 -> 229.75 | Improved |
| add learnable LSQ scales to LoRA + Hadamard (ablation) | HQ-DM · ImageNet 256x256 LDM-4 W4A4 | IS: 229.75 -> 248.30 | Improved |
| lower activation bits from W4A4 to W4A3 | HQ-DM vs EfficientDM · ImageNet 256x256 LDM-4 | IS: 220.20 -> 248.30 at W4A4; 8.74 -> 49.62 at W4A3 | Improved |
| HQ-DM vs EfficientDM on LSUN-Churches | HQ-DM · LSUN-Churches LDM-8 | FID: 12.83 -> 9.64 (W4A4); 20.97 -> 14.67 (W4A3) | Improved |
Closest papers by meaning
- MPQ-DM: Mixed Precision Quantization for Extremely Low Bit Diffusion Models (AAAI 2025 · similarity 0.91)
- BiDM: Pushing the Limit of Quantization for Diffusion Models (NeurIPS 2024 · similarity 0.90)
- Qua2SeDiMo: Quantifiable Quantization Sensitivity of Diffusion Models (AAAI 2025 · similarity 0.90)
- EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models (ICLR 2024 · similarity 0.90)
- SVDQuant: Absorbing Outliers by Low-Rank Component for 4-Bit Diffusion Models (ICLR 2025 · similarity 0.90)
- Towards Accurate Post-training Quantization for Diffusion Models (CVPR 2024 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
