Research map / Image generation / Fast diffusion sampling
Transformers Learn the Optimal DDPM Denoiser for Multi-Token GMMs
Hongkang Li, Hancheng Min, Rene Vidal
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| increase the number of distinct patterns K per data point (ablation) | one-layer transformer denoiser on synthetic MTGM data · synthetic MTGM (d=64, M=8, P=256) | training iterations to converge | Hurt |
| make the pattern distribution more uniform (larger minimal mixture fraction) (ablation) | one-layer transformer denoiser on synthetic MTGM data · synthetic MTGM | iterations to converge (excess risk) | Improved |
| sample diffusion time steps only from the first 40% (higher time-averaged SNR) (ablation) | one-layer transformer denoiser on synthetic MTGM data · synthetic MTGM | convergence speed (excess risk) | Improved |
| increase tokens per data point P | trained transformer attention · theory, MTGM data | error in estimating mean pattern | Improved |
| train transformer with DDPM loss | one-layer single-head transformer · synthetic MTGM | DDPM test loss vs Bayes risk | Improved |
| train transformer with DDPM loss | one-layer single-head transformer · synthetic MTGM | attention weight on same-pattern keys | Improved |
| use fewer training samples for minority class (30% of digit 2) (ablation) | 6-layer 4-head DiT · MNIST digits 0-3 | FID of generated digit | Hurt |
Closest papers by meaning
- O(d/T) Convergence Theory for Diffusion Probabilistic Models under Minimal Assumptions (ICLR 2025 · similarity 0.90)
- Interpreting and Improving Diffusion Models from an Optimization Perspective (ICML 2024 · similarity 0.90)
- D^2-DPM: Dual Denoising for Quantized Diffusion Probabilistic Models (AAAI 2025 · similarity 0.89)
- Your Diffusion Model is Secretly a Noise Classifier and Benefits from Contrastive Training (NeurIPS 2024 · similarity 0.89)
- Training Diffusion-based Generative Models with Limited Data (ICML 2025 · similarity 0.89)
- Remix-DiT: Mixing Diffusion Transformers for Multi-Expert Denoising (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.
