Research map / Image generation / Fast diffusion sampling
Post-Training Quantization on Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Apply 8-bit PTQ to the diffusion model | diffusion models · ImageNet 64x64, DDIM 100/250 steps and DDPM 4000 steps | FID: 21.70 -> 24.92; 21.63 -> 23.96; 20.82 -> 23.64 | Hurt |
| Apply 8-bit PTQ to the diffusion model | diffusion models · CIFAR 32x32, DDIM 100/250 steps | FID: 10.05 -> 14.18; 8.91 -> 11.66 | Hurt |
| Apply 8-bit PTQ to the diffusion model | diffusion models · CIFAR 32x32, DDPM 4000 steps | FID: 7.14 -> 7.10 | Helped |
| Calibrate with NDTC rather than the single-timestep baseline (ablation) | PTQ for diffusion models · ImageNet 64x64, DDIM 250 timesteps | FID: 49.37 -> 24.26 | Helped |
| Use varying uniformly sampled timesteps rather than one timestep (ablation) | PTQ for diffusion models · ImageNet 64x64, DDIM 250 timesteps | FID: 49.37 -> 26.19 | Helped |
| Calibrate using generated images rather than denoising-process noise samples (ablation) | PTQ for diffusion models · ImageNet 64x64 | FID: 33.15 -> 128.63 | Hurt |
| Use training-mimic calibration samples rather than noise samples (ablation) | PTQ for diffusion models · ImageNet 64x64 | FID: 33.15 -> 34.55 | Hurt |
| Use MSE rather than L1, cosine, or KL calibration distance (ablation) | PTQ for diffusion models · ImageNet 64x64 | FID: 100.52; 34.81; 47.27 -> 30.46 | Helped |
| Use Poisson rather than normal timestep sampling (ablation) | NDTC · ImageNet 64x64, DDIM 250 timesteps | FID: 23.96 -> 34.54 | Hurt |
| Use exponential rather than normal timestep sampling (ablation) | NDTC · ImageNet 64x64, DDIM 250 timesteps | FID: 23.96 -> 39.91 | Hurt |
| Set timestep-sampling mean to μ=1.5T/2 instead of μ=0.5T/2 (ablation) | NDTC · ImageNet 64x64, DDIM 250 timesteps | FID: 23.96 -> 39.09 | Hurt |
| Quantize the μ, Σ, or xₜ₋₁ output operation (ablation) | diffusion model · ImageNet 64x64, DDIM 250 timesteps | FID: 21.63 -> 21.38; 21.63 -> 21.96; 21.63 -> 21.94 | No clear effect |
Closest papers by meaning
- D^2-DPM: Dual Denoising for Quantized Diffusion Probabilistic Models (AAAI 2025 · similarity 0.93)
- TFMQ-DM: Temporal Feature Maintenance Quantization for Diffusion Models (CVPR 2024 · similarity 0.92)
- AccuQuant: Simulating Multiple Denoising Steps for Quantizing Diffusion Models (NeurIPS 2025 · similarity 0.92)
- MPQ-DM: Mixed Precision Quantization for Extremely Low Bit Diffusion Models (AAAI 2025 · similarity 0.91)
- Towards Accurate Post-training Quantization for Diffusion Models (CVPR 2024 · similarity 0.91)
- Training Diffusion-based Generative Models with Limited Data (ICML 2025 · 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.