Research map / Image generation / Fast diffusion sampling
Error Propagation Mechanisms and Compensation Strategies for Quantized Diffusion Models
Songwei Liu, Chao Zeng, Chenqian Yan, Xurui Peng, XING WANG, Fangmin Chen, Xing Mei
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 TCEC to SVDQuant W4A4 | SVDQuant (SDXL) · MJHQ | FID: 20.6 -> 18.1 | Improved |
| add TCEC to SVDQuant W8A8 | SVDQuant (SDXL) · MJHQ | PSNR: 26.4 -> 27.3 | Improved |
| add TCEC to SVDQuant W8A8 | SVDQuant (SDXL-Turbo) · MJHQ | FID: 24.3 -> 24.5 | No clear effect |
| add TCEC to SVDQuant W4A4 | SVDQuant (SDXL-Turbo) · MJHQ | FID: 24.6 -> 23.9 | Improved |
| add TCEC to SVDQuant W4A4 | SVDQuant (PixArt-Sigma) · MJHQ | PSNR: 17.6 -> 18.3 | Improved |
| add TCEC to ViDiT-Q W4A8 | ViDiT-Q (OpenSora) · VBench | Imaging Quality: 61.07 -> 64.97 | Improved |
| add TCEC to ViDiT-Q W8A8 | ViDiT-Q (OpenSora) · VBench | Dynamic Degree: 61.11 -> 61.09 | No clear effect |
| use two-step approximate error solving instead of iterative solving (ablation) | TCEC (SDXL) · MJHQ | FID: 16.5 -> 16.0 | Improved |
| use empirical rule for lambda1 instead of grid search (ablation) | TCEC (SDXL) · MJHQ | FID: 16.2 -> 16.0 | No clear effect |
| increase calibration set from 256 to 1024 (ablation) | TCEC (SDXL) · MJHQ | PSNR: 26.7 -> 27.3 | Improved |
| increase calibration set from 1024 to 2048 (ablation) | TCEC (SDXL) · MJHQ | PSNR: 27.3 -> 27.4 | No clear effect |
| add TCEC to SVDQuant at 2 inference steps | SDXL-Turbo · MJHQ | FID: 27.7 -> 27.7 | No clear effect |
Closest papers by meaning
- Optimizing Quantized Diffusion Models via Distillation with Cross-Timestep Error Correction (AAAI 2025 · similarity 0.93)
- On Error Propagation of Diffusion Models (ICLR 2024 · similarity 0.92)
- AccuQuant: Simulating Multiple Denoising Steps for Quantizing Diffusion Models (NeurIPS 2025 · similarity 0.92)
- D^2-DPM: Dual Denoising for Quantized Diffusion Probabilistic Models (AAAI 2025 · similarity 0.92)
- 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 0.91)
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.
