Research map / Image generation / Fast diffusion sampling
D^2-DPM: Dual Denoising for Quantized Diffusion Probabilistic 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 D^2-DPM, a post-training quantization method for diffusion models that models quantization noise as a time-step-aware Gaussian and corrects its mean and variance deviations during sampling via stochastic or deterministic dual denoising.
Problem
Post-training quantization of diffusion models adds quantization noise to the noise estimation network, causing mean and variance deviations that distort the sampling trajectory and degrade image quality.
Approach
Model the quantized output and quantization noise jointly as a time-step-aware Gaussian, then at inference subtract the conditional mean (deterministic variant) or a sampled noise estimate (stochastic variant) of the quantization noise, and adjust the diffusion coefficient to absorb the extra variance.
Key results
- On ImageNet class-conditional LDM-4 (scale 3.0, 20 steps, W8A8), S-D^2 reaches FID 8.11 vs 11.13 for full precision and 8.46 for PTQD, with 3.99x compression and 11.67x BOPs reduction.
- At W4A8 on LDM-4 (eta=0), D-D^2 achieves FID 9.71 vs 10.41 for PTQD.
- On LSUN-Bedroom unconditional (W4A8), D-D^2 gets FID 10.72 vs 12.57 for PTQD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply stochastic joint correction (SJC) (ablation) | PTQD on ImageNet 256×256, LDM-4, η=1.0, scale=1.5, 250 steps, W4A8 · ImageNet 256×256 | FID: 7.90 -> 6.91 | Helped |
| Apply stochastic joint correction (SJC) (ablation) | PTQD on ImageNet 256×256, LDM-4, η=1.0, scale=1.5, 250 steps, W4A8 · ImageNet 256×256 | sFID: 7.87 -> 7.49 | Helped |
| Add deterministic mean correction (DMC) (ablation) | PTQD on ImageNet 256×256, LDM-4, η=1.0, scale=1.5, 250 steps, W4A8 · ImageNet 256×256 | FID: 7.90 -> 7.14 | Helped |
| Add deterministic mean correction (DMC) (ablation) | PTQD on ImageNet 256×256, LDM-4, η=1.0, scale=1.5, 250 steps, W4A8 · ImageNet 256×256 | sFID: 7.87 -> 7.67 | Helped |
| Add deterministic variance correction (DVC) to DMC (ablation) | DMC on ImageNet 256×256, LDM-4, η=1.0, scale=1.5, 250 steps, W4A8 · ImageNet 256×256 | FID: 7.14 -> 6.81 | Helped |
| Add deterministic variance correction (DVC) to DMC (ablation) | DMC on ImageNet 256×256, LDM-4, η=1.0, scale=1.5, 250 steps, W4A8 · ImageNet 256×256 | sFID: 7.67 -> 7.42 | Helped |
| Use dual denoising instead of PTQD | LDM-4 on ImageNet 256×256, η=0.0, W8A8 · ImageNet 256×256 | FID: 8.46 -> 8.11 | Helped |
| Use dual denoising instead of PTQD | LDM-4 on ImageNet 256×256, η=1.0, W4A8 · ImageNet 256×256 | FID: 13.15 -> 12.60 | Helped |
| Use deterministic dual denoising instead of PTQD | LDM-4 on ImageNet 256×256, η=0.0, scale=1.5, 250 steps, W4A8 · ImageNet 256×256 | FID: 6.46 -> 5.56 | Helped |
| Use deterministic dual denoising instead of PTQD | LDM-4 on LSUN-Bedroom 256×256, η=1.0, 200 steps, W8A8 · LSUN-Bedroom 256×256 | FID: 9.16 -> 7.55 | Helped |
| Use stochastic dual denoising instead of PTQD | LDM-8 on LSUN-Church 256×256, η=0.0, 200 steps, W8A8 · LSUN-Church 256×256 | FID: 8.31 -> 7.82 | Helped |
Surprising findings
- The deterministic variant (D-D^2) performs comparably or better than the stochastic variant even when the stochasticity capacity is minimal, which the authors attribute to quantization-induced variance acting as extra Langevin-like noise.
- Several quantized configurations report lower FID than the full-precision model (e.g., S-D^2 at W8A8 on ImageNet with eta=0).
Limitations
- Relies on the empirical Gaussian assumption for quantization noise and diagonal/isotropic covariance simplifications.
- Experiments are limited to LDM models on image datasets at 256x256 resolution with W8A8 and W4A8 settings.
Experimental setup
Datasets: ImageNet LSUN-Bedrooms LSUN-Churches
Benchmarks: ImageNet 256x256 LSUN-Bedrooms 256x256 LSUN-Church 256x256
Metrics: FID sFID IS precision recall BOPs model size
Baselines: PTQ4DM Q-diffusion PTQD BRECQ Qrop
Builds on: BRECQ PTQD DDIM Latent Diffusion Models Adaround
Closest papers by meaning
- AccuQuant: Simulating Multiple Denoising Steps for Quantizing Diffusion Models (NeurIPS 2025 · similarity 0.91)
- Qua2SeDiMo: Quantifiable Quantization Sensitivity of Diffusion Models (AAAI 2025 · similarity 0.91)
- On Inference Stability for Diffusion Models (AAAI 2024 · similarity 0.90)
- Optimizing Quantized Diffusion Models via Distillation with Cross-Timestep Error Correction (AAAI 2025 · similarity 0.90)
- UDPM: Upsampling Diffusion Probabilistic Models (NeurIPS 2024 · similarity 0.90)
- EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models (ICLR 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.