Research map / Image generation / Fast diffusion sampling
SegQuant: A Semantics-Aware and Generalizable Quantization Framework for Diffusion Models
Jiaji Zhang, Ruichao Sun, Hailiang Zhao, Jiaju Wu, Peng Chen, Hao Li, Yuying Liu, Kingsum Chow, Gang Xiong, Shuiguang Deng
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 |
|---|---|---|---|
| add SegLinear (segment-wise quantization from graph structure) (ablation) | SmoothQuant on SD3.5 DiT.0.norm1 · SD3.5 single layer | Frobenius error: 0.7041 -> 0.5381 | Improved |
| add SegLinear (ablation) | GPTQ on SD3.5 DiT.11.norm1 context · SD3.5 single layer | Frobenius error: 3.0176 -> 1.7637 | Improved |
| add SegLinear (ablation) | SVDQuant on SD3.5 DiT.0.norm1 · SD3.5 single layer | Frobenius error: 0.0861 -> 0.0714 | Improved |
| add SegLinear alone (ablation) | SD3.5 W8A8 quantization (SmoothQuant + AMax) · MJHQ-30K | FID: 23.35 -> 23.36 | No clear effect |
| add SegLinear alone (ablation) | SD3.5 W8A8 quantization · MJHQ-30K | Image Reward: 0.877 -> 0.899 | Improved |
| add DualScale alone (ablation) | SD3.5 W8A8 quantization · MJHQ-30K | FID: 23.35 -> 22.61 | Improved |
| combine SegLinear and DualScale (ablation) | SD3.5 W8A8 quantization · MJHQ-30K | FID / Image Reward: FID 23.35 -> 22.54; IR 0.877 -> 0.952 | Improved |
| use SegQuant instead of Smooth+ on FLUX W8A8 int | FLUX.1-dev · MJHQ-30K | FID: 26.21 -> 22.85 | Improved |
| use SegQuant instead of SVDQuant at W4A8 | SD3.5-DiT · MJHQ-30K | FID: 27.95 -> 27.30 | Improved |
| use SegQuant instead of Q-Diffusion on SDXL W8A8 fp | SDXL-UNet · MJHQ-30K | FID: 17.04 -> 17.03 | No clear effect |
| add segmentation and dual-scale steps | SD3.5 (RTX 4090) · Inference efficiency | Inference time (ms): INT8 278.86 -> SegQuant-8 287.85 | Hurt |
Closest papers by meaning
- Q-DiT: Accurate Post-Training Quantization for Diffusion Transformers (CVPR 2025 · similarity 0.91)
- Towards Accurate Post-training Quantization for Diffusion Models (CVPR 2024 · similarity 0.91)
- EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models (ICLR 2024 · similarity 0.90)
- TCAQ-DM: Timestep-Channel Adaptive Quantization for Diffusion Models (AAAI 2025 · similarity 0.90)
- Qua2SeDiMo: Quantifiable Quantization Sensitivity of Diffusion Models (AAAI 2025 · similarity 0.90)
- AccuQuant: Simulating Multiple Denoising Steps for Quantizing Diffusion Models (NeurIPS 2025 · 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.
