Research map / Image generation / Fast diffusion sampling
Directly Denoising 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 Directly Denoising Diffusion Models (DDDM), which train a U-Net to map noisy inputs to clean images conditioned on its own previous-iteration estimate, plus a Pseudo-LPIPS loss, enabling strong one- and two-step generation without distillation or ODE solvers.
Problem
Diffusion models are slow to sample, while few-step methods typically rely on ODE solvers or distillation from pretrained teacher diffusion models, adding cost and constraints.
Approach
Train the denoiser conditioned on its own estimate of x0 from the previous training iteration (iterative fixed-point update of x0 estimate in the probability flow ODE framework), so one-step and multistep sampling share one model without ODE solvers or distillation; additionally use Pseudo-LPIPS, sqrt(LPIPS + c^2) - c, as the loss metric.
Key results
- DDDM-deep achieves FID 2.57 (one-step) and 2.33 (two-step) on unconditional CIFAR-10, and 1.79 with 1000-step sampling.
- DDDM with T=8000 reaches FID 1.74 at 1000 steps on CIFAR-10, reported as state of the art.
- On ImageNet 64x64, DDDM attains FID 4.21 (one-step) and 3.53 (two-step), trailing iCT-deep (3.25 and 2.77).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase sampling from 1 to 2 steps (ablation) | DDDM-deep (T=1000) · CIFAR-10 | FID: 2.57 -> 2.33 | Helped |
| Increase sampling from 2 to 1000 steps (ablation) | DDDM-deep (T=1000) · CIFAR-10 | FID: 2.33 -> 1.79 | Helped |
| Increase sampling from 1 to 2 steps (ablation) | DDDM-deep (T=1000) · ImageNet 64x64 | FID: 3.47 -> 3.08 | Helped |
| Increase sampling from 2 to 1000 steps (ablation) | DDDM-deep (T=1000) · ImageNet 64x64 | FID: 3.08 -> 2.11 | Helped |
| Use the deeper architecture (ablation) | DDDM · CIFAR-10, 1 NFE | FID: 2.90 -> 2.57 | Helped |
| Use the deeper architecture (ablation) | DDDM · ImageNet 64x64, 1 NFE | FID: 4.21 -> 3.47 | Helped |
| Increase T and training epochs together (ablation) | DDDM · CIFAR-10 | FID: 2.90 -> 2.86 -> 2.83 -> 2.82 | Helped |
| Use Pseudo-LPIPS instead of LPIPS (ablation) | DDDM · CIFAR-10 | FID | Helped |
| Use Pseudo-LPIPS instead of Pseudo-Huber (ablation) | DDDM · CIFAR-10 | FID | Helped |
| Use Pseudo-LPIPS instead of L1 or squared L2 (ablation) | DDDM · CIFAR-10 | FID | Helped |
| Vary the Pseudo-LPIPS hyperparameter c (ablation) | DDDM · CIFAR-10 | FID | Mixed |
| Increase sampling steps (ablation) | DDDM · CIFAR-10 and ImageNet 64x64 | FID | Helped |
Surprising findings
- FID is not strictly monotonically decreasing with the number of sampling steps, though the overall trend is downward.
- Pseudo-LPIPS consistently beats LPIPS across a wide range of c values; the paper also notes that using ImageNet-trained LPIPS and Inception may bias FID on ImageNet.
Limitations
- Stores a previous x0 estimate per training sample, adding memory overhead (29.5GB for ImageNet 64x64).
- Evaluation may be biased because ImageNet is used in both LPIPS and the Inception network for FID; no human evaluation.
- ImageNet 64x64 results lag iCT/iCT-deep, which the authors attribute to smaller batch size and fewer iterations and possibly suboptimal loss hyperparameters.
Experimental setup
Datasets: CIFAR-10 ImageNet 64x64
Benchmarks: CIFAR-10 ImageNet 64x64
Metrics: FID Inception Score Precision Recall
Baselines: DDIM DPM-solver DEIS UniPC DFNO TRACT Diff-Instruct Consistency Distillation Consistency Training iCT StyleGAN2-ADA EDM
Builds on: DDPM DDIM Consistency models Pseudo-Huber loss ADM U-Net
Closest papers by meaning
- Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.91)
- Diffusion Bridge Implicit Models (ICLR 2025 · similarity 0.91)
- Residual Denoising Diffusion Models (CVPR 2024 · similarity 0.91)
- Denoising Diffusion Step-aware Models (ICLR 2024 · similarity 0.90)
- Learning to Discretize Denoising Diffusion ODEs (ICLR 2025 · similarity 0.90)
- UDPM: Upsampling Diffusion Probabilistic Models (NeurIPS 2024 · similarity 0.89)
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.