Research map / Image generation / Fast diffusion sampling
The Diffusion Duality, Chapter II: $\Psi$-Samplers and Efficient Curriculum
Justin Deschenaux, Caglar Gulcehre, Subham Sekhar Sahoo
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 |
|---|---|---|---|
| replace Duo's full-vocabulary curriculum with the top-k sparse approximation (Duo++) | Duo · LM1B / OpenWebText | test PPL: 29.9 -> 30.0 (LM1B), 25.2 -> 25.2 (OWT) | No clear effect |
| replace Duo's curriculum with the efficient curriculum | Duo · training cost | peak memory and training time: -33% peak memory, -25% training time | Improved |
| increase k in the efficient curriculum from 2 to 5 (ablation) | Duo++ · LM1B | test PPL: 30.0 -> 30.2 | No clear effect |
| use the efficient curriculum instead of Duo's on multiple-choice tasks | Duo · 7 MCQ downstream tasks | accuracy | Mixed |
| use Psi-samplers (predictor-corrector) instead of ancestral sampling for uniform-state diffusion | Duo++ · OpenWebText generation | Gen. PPL at matched entropy | Improved |
| use Psi-samplers with Duo++ instead of ancestral sampling | Duo++ · CIFAR-10 | FID and IS | Improved |
| add nucleus sampling (p=0.9) (ablation) | MDLM and Duo on OWT · OpenWebText generation | Gen. PPL | Improved |
| add nucleus sampling at high NFE (ablation) | MDLM on CIFAR-10 · CIFAR-10 | FID | Mixed |
| use piecewise constant kappa_t instead of ReMDM loop schedule (ablation) | Psi-samplers · CIFAR-10 / OWT | FID/IS, Gen. PPL | Improved |
| use uniform-state with Psi-sampler vs masked diffusion with ReMDM | Duo++ vs MDLM · OpenWebText generation | Gen. PPL | Improved |
Closest papers by meaning
- Informed Correctors for Discrete Diffusion Models (NeurIPS 2025 · similarity 0.91)
- Encoder-Decoder Diffusion Language Models for Efficient Training and Inference (NeurIPS 2025 · similarity 0.90)
- The Diffusion Duality (ICML 2025 · similarity 0.90)
- Scalable Discrete Diffusion Samplers: Combinatorial Optimization and Statistical Physics (ICLR 2025 · similarity 0.90)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.90)
- Dual-Rate Diffusion: Accelerating diffusion models with an interleaved heavy-light network (NeurIPS 2026 · 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.
