Research map / Image generation / Fast diffusion sampling
MixDQ: Memory-Efficient Few-Step Text-to-Image Diffusion Models with Metric-Decoupled Mixed Precision Quantization
Tianchen Zhao, Xuefei Ning, Tongcheng Fang, Enshu Liu, Guyue Huang, Zinan Lin, Shengen Yan, Guohao Dai, Yu Wang
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 |
|---|---|---|---|
| add BOS-aware quantization of text embeddings (ablation) | MixDQ W8A8 on SDXL-turbo · COCO2014 | FID: 103.95 -> 31.65 (FID); CLIP 0.1478 -> 0.2652 | Improved |
| add mixed precision using SQNR-only sensitivity (ablation) | BOS-aware W8A8 SDXL-turbo · COCO2014 | FID: 31.65 -> 37.35 (FID); CLIP 0.2652 -> 0.2624 | Hurt |
| add metric-decoupled sensitivity (SSIM for content, SQNR for quality) to mixed precision (ablation) | MixDQ W8A8 on SDXL-turbo · COCO2014 | FID: 37.35 -> 17.03 (FID); CLIP 0.2624 -> 0.2703 | Improved |
| use MixDQ instead of Q-Diffusion at W4A8 | SDXL-turbo 1 step · COCO2014 | FID: 118.93 -> 17.68 | Improved |
| use MixDQ instead of Q-Diffusion at W8A8 | SDXL-turbo 1 step · COCO2014 | FID: 76.18 -> 17.03 | Improved |
| use FP8 non-uniform quantization instead of MixDQ | SDXL-turbo W8A8 · COCO subset 1024 prompts | FID: 83.39 -> 101.73 | Hurt |
| use EDA-DM layer-wise reconstruction over Q-Diffusion | SDXL-turbo W8A8 · COCO subset 1024 prompts | FID: 149.15 -> 137.98 | No clear effect |
| run the 1-step model instead of 2-step under quantization | SDXL-turbo · COCO prompts | quantization degradation | Hurt |
| use integer programming allocation vs random or naive sorting (ablation) | MixDQ mixed precision · SDXL-turbo Pareto frontier | CLIP Score vs bit-width | Improved |
| W4A16 vs FP16 latency | MixDQ · RTX 4080 | latency | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- PTQD — PTQD: Accurate Post-Training Quantization for Diffusion Models (NeurIPS 2023) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DGQ: Distribution-Aware Group Quantization for Text-to-Image Diffusion Models (ICLR 2025 · similarity 0.91)
- EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models (ICLR 2024 · similarity 0.89)
- MPQ-DM: Mixed Precision Quantization for Extremely Low Bit Diffusion Models (AAAI 2025 · similarity 0.89)
- Efficient Personalization of Quantized Diffusion Model without Backpropagation (CVPR 2025 · similarity 0.88)
- TFMQ-DM: Temporal Feature Maintenance Quantization for Diffusion Models (CVPR 2024 · similarity 0.88)
- ViDiT-Q: Efficient and Accurate Quantization of Diffusion Transformers for Image and Video Generation (ICLR 2025 · similarity 0.88)
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.
