Research map / Image generation / Fast diffusion sampling
Soft Truncation: A Universal Training Technique of Score-based Diffusion Model for High Precision Score Estimation
Dongjun Kim, Seungjae Shin, Kyungwoo Song, Wanmo Kang, Il-Chul Moon
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Randomize the truncation threshold across minibatches instead of using a fixed threshold (ablation) | Likelihood-weighted DDPM++ (VP) · CIFAR-10 | FID (ODE): 6.70 -> 3.96 | Improved |
| Use randomized truncation with P₀.₉ rather than fixed likelihood-weighted truncation (ablation) | DDPM++ (VP) · CIFAR-10 | FID (ODE): 6.70 -> 3.45 | Improved |
| Use randomized truncation P₁ rather than fixed likelihood-weighted truncation (ablation) | DDPM++ (VP) · CIFAR-10 | NLL: 3.03 -> 3.01 | Improved |
| Decrease σmin from 10⁻² to 10⁻⁵ (ablation) | NCSN++ · CIFAR-10 | NLL: 4.95 -> 2.97 | Improved |
| Decrease σmin from 10⁻² to 10⁻⁵ (ablation) | NCSN++ · CIFAR-10 | FID-10k: 6.95 -> 8.29 | Hurt |
| Use randomized truncation P₀ instead of a distribution that samples small truncation levels (ablation) | DDPM++ (VP) · CIFAR-10 | NLL: 3.01 -> 3.24 | Hurt |
| Use truncation distribution P₃ rather than P₀.₉ (ablation) | DDPM++ (VP) · CIFAR-10 | FID (ODE): 3.45 -> 6.54 | Hurt |
| Replace fixed likelihood weighting with Soft Truncation P₀.₉ (ablation) | DDPM++ (VP) · ImageNet32 | FID (ODE): 12.68 -> 8.42 | Improved |
| Apply Soft Truncation with variance weighting (ablation) | NCSN++ (VE) · CelebA | FID (PC): 3.95 -> 2.68 | Improved |
| Apply Soft Truncation with likelihood weighting (ablation) | DDPM++ (VP) · CelebA | FID (ODE): 3.95 -> 2.90 | Improved |
| Use Soft Truncation with P₁ instead of hard truncation at ε = 10⁻⁵ (ablation) | DDPM++ (VP) · CIFAR-10 | NLL: 3.03 -> 3.01 | Improved |
Closest papers by meaning
- A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models (ICML 2024 · similarity 0.88)
- Sparsely Supervised Diffusion (NeurIPS 2026 · similarity 0.87)
- One-Step Diffusion Distillation through Score Implicit Matching (NeurIPS 2024 · similarity 0.87)
- High-Fidelity Diffusion-Based Image Editing (AAAI 2024 · similarity 0.87)
- Attention-Driven Training-Free Efficiency Enhancement of Diffusion Models (CVPR 2024 · similarity 0.87)
- Towards Accurate Post-training Quantization for Diffusion Models (CVPR 2024 · similarity 0.86)
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.
