Research map / Architectures and efficiency / Quantization
QuEST: Low-bit Diffusion Model Quantization via Efficient Selective Finetuning
Haoxuan Wang, Yuzhang Shang, Zhihang Yuan, Junyi Wu, Junchi Yan, Yan Yan
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 time-embedding layer finetuning (TLA) to direct quantization (ablation) | QuEST · LSUN-Bedrooms LDM-4 W4A8 | FID: 6.95 -> 4.41 | Improved |
| add attention-related layer finetuning (CMA) after TLA (ablation) | QuEST · LSUN-Bedrooms LDM-4 W4A8 | FID: 4.41 -> 3.26 | Improved |
| add global loss supervision from full-precision output to TLA (ablation) | QuEST · LSUN-Bedrooms LDM-4 W4A8 | FID: 8.99 -> 6.41 | Improved |
| add global loss supervision to CMA (ablation) | QuEST · LSUN-Bedrooms LDM-4 W4A8 | FID: 11.41 -> 6.20 | Improved |
| supervise TLA with global loss only (no local loss) (ablation) | QuEST · LSUN-Bedrooms LDM-4 | FID: -7.13 FID, -9.39 sFID | Hurt |
| finetune only time-embedding and attention layers instead of full finetuning (ablation) | QuEST · LDM-4 W4A8 | FID: 5.38 -> 3.82 | Improved |
| integrate QALoRA finetuning instead of finetuning original layers (ablation) | QuEST · ImageNet 256x256 | FID: +5.62 FID | Hurt |
| optimize time embeddings with prediction-error loss instead of reusing precomputed features (ablation) | QuEST · LDM-4 | FID | Improved |
| quantize ImageNet LDM-4 to fewer bits (lower FID under quantization) | all methods · ImageNet 256x256 LDM-4 | FID: 11.28 (FP) -> 5.98 (QuEST W4A4) | Mixed |
| apply QuEST instead of PTQ for efficiency | QuEST · LSUN-Bedrooms LDM-4 W4A8 | time cost (GPU hours): 23.08 -> 15.25 | Improved |
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.
- EfficientDM — EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models (ICLR 2024) · better on 8 of 8 shared results
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
- EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models (ICLR 2024 · similarity 0.93)
- BitsFusion: 1.99 bits Weight Quantization of Diffusion Model (NeurIPS 2024 · similarity 0.93)
- Qua2SeDiMo: Quantifiable Quantization Sensitivity of Diffusion Models (AAAI 2025 · similarity 0.91)
- 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)
- 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.
