Research map / Image generation / Fast diffusion sampling
Truncated Diffusion Probabilistic Models and Diffusion-based Adversarial Auto-Encoders
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 |
|---|---|---|---|
| Truncate the diffusion chain and learn a separate implicit starting distribution | DDPM on CIFAR-10 · CIFAR-10 | FID: 3.21 -> 2.88 | Helped |
| Use 100 function evaluations instead of the full 1000-step DDPM sampling chain | DDPM on CIFAR-10 · CIFAR-10 | Recall: 0.57 -> 0.58 | Helped |
| Truncate diffusion and use a separate implicit generator | DDPM on LSUN-Church · LSUN-Church 256x256 | FID: 7.89 -> 3.98 | Helped |
| Truncate diffusion and use a separate implicit generator | DDPM on LSUN-Bedroom · LSUN-Bedroom 256x256 | FID: 4.90 -> 3.67 | Helped |
| Truncate the latent diffusion chain and learn its initial distribution | LDM on CUB-Bird · CUB-Bird | FID: 6.82 -> 6.72 | Helped |
| Truncate the latent diffusion chain and learn its initial distribution | LDM on MS-COCO · MS-COCO | FID: 6.3 -> 6.29 | Helped |
| Use a separately trained implicit generator with a pre-trained diffusion backbone (ablation) | TDPM+ on CIFAR-10, T_trunc=99 · CIFAR-10 | FID: 2.88 -> 2.85 | No clear effect |
| Use a pre-trained improved-DDPM backbone with the implicit generator (ablation) | TDPM+ on CIFAR-10, T_trunc=99 · CIFAR-10 | FID: 2.83 -> 2.25 | Helped |
| Replace the linear noise schedule with a cosine schedule (ablation) | TDPM-GAN on CIFAR-10 · CIFAR-10 | FID: 3.10 / 3.47; 3.30 / 3.16 | Mixed |
| Replace the linear noise schedule with a cosine schedule (ablation) | TDPM-CT on CIFAR-10 · CIFAR-10 | FID: 3.69 / 3.62; 3.97 / 3.24 | Mixed |
| Increase the truncation level from 4 to 99 steps (ablation) | TDPM on CIFAR-10 · CIFAR-10 | FID: 3.34 -> 3.10 | Helped |
| Use conditional transport instead of the GAN objective to fit the implicit distribution (ablation) | TDPM · Synthetic toy data | forward KL | No clear effect |
Closest papers by meaning
- Unified Generation, Reconstruction, and Representation: Generalized Diffusion with Adaptive Latent Encoding-Decoding (ICML 2024 · similarity 0.91)
- Stochastic Forward–Backward Deconvolution: Training Diffusion Models with Finite Noisy Datasets (ICML 2025 · similarity 0.90)
- Structure-Guided Adversarial Training of Diffusion Models (CVPR 2024 · similarity 0.90)
- Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.90)
- Bridging Data Gaps in Diffusion Models with Adversarial Noise-Based Transfer Learning (ICML 2024 · similarity 0.90)
- Ambient Diffusion Omni: Training Good Models with Bad Data (NeurIPS 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.