Research map / Image generation / Fast diffusion sampling
DMQ: Dissecting Outliers of Diffusion Models for Post-Training Quantization
Dongyeun Lee, Jiwan Hur, Hyounguk Shon, Jae Young Lee, Junmo Kim
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 |
|---|---|---|---|
| replace SmoothQuant scaling with learned equivalent scaling | W4A8 quantized LDM-4 · FFHQ 256x256 | FID: 454.16 -> 30.37 | Improved |
| add Learned Equivalent Scaling (ablation) | baseline PTQ · FFHQ W4A8 | FID: 36.08 -> 33.46 | Improved |
| add adaptive timestep weighting (ablation) | LES · FFHQ W4A8 | FID: 33.46 -> 31.83 | Improved |
| add Power-of-Two Scaling (ablation) | LES + weighting · FFHQ W4A8 | FID: 31.83 -> 30.37 | Improved |
| add Power-of-Two Scaling (ablation) | LES + weighting · FFHQ W4A8 | sFID: 24.39 -> 22.72 | Improved |
| use adaptive weighting instead of uniform (ablation) | LES · FFHQ W4A8 | FID: 33.46 -> 31.83 | Improved |
| use linear early-step weighting instead of uniform (ablation) | LES · FFHQ W4A8 | FID: 33.46 -> 36.58 | Hurt |
| use quadratic early-step weighting instead of uniform (ablation) | LES · FFHQ W4A8 | FID: 33.46 -> 35.30 | Hurt |
| select PTS factors by voting instead of MSE (ablation) | PTS on all layers · FFHQ W4A8 | FID: 32.55 -> 31.91 | Improved |
| apply PTS only to skip connections instead of all layers (ablation) | voting PTS · FFHQ W4A8 | FID: 31.91 -> 30.37 | Improved |
| add MSE-selected PTS to all layers (ablation) | w/o PTS · FFHQ W4A8 | FID: 31.83 -> 32.55 | Hurt |
| custom W4A8 kernel with bit-shifting vs FP32 GEMM | PTS · latency at M=3072 | speedup: 5.17x | Improved |
Closest papers by meaning
- MPQ-DM: Mixed Precision Quantization for Extremely Low Bit Diffusion Models (AAAI 2025 · similarity 0.94)
- EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models (ICLR 2024 · similarity 0.93)
- Qua2SeDiMo: Quantifiable Quantization Sensitivity of Diffusion Models (AAAI 2025 · similarity 0.92)
- Towards Accurate Post-training Quantization for Diffusion Models (CVPR 2024 · similarity 0.92)
- TCAQ-DM: Timestep-Channel Adaptive Quantization for Diffusion Models (AAAI 2025 · similarity 0.92)
- SVDQuant: Absorbing Outliers by Low-Rank Component for 4-Bit Diffusion Models (ICLR 2025 · similarity 0.92)
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.
