Research map / Image generation / Fast diffusion sampling
Towards Accurate Post-training Quantization for Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes APQ-DM, a post-training quantization framework for diffusion models that uses timestep-group-specific quantizers found by differentiable search and selects calibration timesteps via a structural-risk-minimization-inspired criterion.
Problem
Post-training quantization of diffusion models suffers large errors because a shared quantizer is used across timesteps whose activation distributions differ, and randomly timestep-sampled calibration images generalize poorly.
Approach
Partition diffusion timesteps into groups with distribution-aware quantizers whose assignment is learned via differentiable search with entropy regularization, and choose calibration timesteps by a generalized SRM criterion combining importance-weight entropy and sampling counts (UCB-style).
Key results
- At W6A6 on LSUN-Bedroom (DDIM), APQ-DM reaches FID 6.57 on CIFAR-10 vs 9.19 for Q-Diffusion and 6.46 vs 7.48 FID against PTQ4DM on LSUN-Bedroom at W8A8.
- On class-conditional ImageNet LDM at W8A8, APQ-DM improves IS to 179.13 and FID to 11.58 versus PTQ4DM's 161.75 and 12.59.
- At W6A6 on LDM LSUN-Church, APQ-DM achieves FID 6.90 vs 11.05 for PTQ4DM, with similar computational cost.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use APQ-DM instead of the strongest listed baseline | DDIM W8A8 · CIFAR-10 | FID: 4.78 -> 4.24 | Helped |
| Use APQ-DM instead of the strongest listed baseline | DDIM W6A6 · LSUN-Church | FID: 32.85 -> 24.75 | Helped |
| Use APQ-DM instead of the strongest listed baseline | LDM W8A8 · ImageNet (conditional) | FID: 12.78 -> 11.58 | Helped |
| Use APQ-DM instead of the strongest listed baseline | LDM W6A6 · LSUN-Church (unconditional) | FID: 10.90 -> 6.90 | Helped |
| Increase timestep groups from 1 to 8 (ablation) | APQ-DM, CIFAR-10 W8A8 · CIFAR-10 | FID: 5.32 -> 4.24 | Helped |
| Increase timestep groups from 1 to 8 (ablation) | APQ-DM, CIFAR-10 W6A6 · CIFAR-10 | FID: 9.73 -> 6.57 | Helped |
| Increase timestep groups from 8 to 16 (ablation) | APQ-DM, CIFAR-10 W8A8 · CIFAR-10 | FID: 4.24 -> 4.29 | Hurt |
| Increase timestep groups from 8 to 16 (ablation) | APQ-DM, CIFAR-10 W6A6 · CIFAR-10 | FID: 6.57 -> 6.77 | Hurt |
| Use active timestep sampling instead of random sampling (ablation) | APQ-DM W8A8, calibration set size 1024 · CIFAR-10 | FID: 5.41 -> 4.24 | Helped |
| Use active timestep sampling instead of heuristic sampling (ablation) | APQ-DM W8A8, calibration set size 1024 · CIFAR-10 | FID: 5.21 -> 4.24 | Helped |
| Use active timestep sampling instead of random sampling (ablation) | APQ-DM W6A6, calibration set size 1024 · CIFAR-10 | FID: 8.92 -> 6.57 | Helped |
| Use active timestep sampling instead of heuristic sampling (ablation) | APQ-DM W6A6, calibration set size 1024 · CIFAR-10 | FID: 8.83 -> 6.57 | Helped |
Surprising findings
- Naive LSQ quantization at W6A6 collapses image quality (e.g., FID above 100 on LSUN-Bedroom DDIM), showing shared quantizers fail badly at low bitwidth.
Limitations
- Sensitivity to the entropy weight lambda, which the authors note can cause large FID changes.
- Gains are smaller at 8-bit where baselines are already close to full precision.
Experimental setup
Datasets: CIFAR-10 CelebA LSUN-Bedroom LSUN-Church CelebA-HQ ImageNet
Benchmarks: CIFAR-10 CelebA LSUN-Bedroom LSUN-Church CelebA-HQ ImageNet
Metrics: IS FID sFID
Baselines: LSQ PTQ4DM Q-Diffusion
Builds on: Q-Diffusion PTQ4DM LSQ DDIM Latent Diffusion Models
Closest papers by meaning
- Q-DiT: Accurate Post-Training Quantization for Diffusion Transformers (CVPR 2025 · similarity 0.92)
- Optimizing Quantized Diffusion Models via Distillation with Cross-Timestep Error Correction (AAAI 2025 · similarity 0.92)
- TCAQ-DM: Timestep-Channel Adaptive Quantization for Diffusion Models (AAAI 2025 · similarity 0.92)
- EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models (ICLR 2024 · similarity 0.91)
- AccuQuant: Simulating Multiple Denoising Steps for Quantizing Diffusion Models (NeurIPS 2025 · similarity 0.91)
- MPQ-DM: Mixed Precision Quantization for Extremely Low Bit Diffusion Models (AAAI 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.