Research map / Image generation / Fast diffusion sampling
Infinite Mask Diffusion for Few-Step Distillation
Jaehoon Yoo, Wonjung Kim, Chanhyuk Lee, Seunghoon Hong
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 single deterministic mask with infinite stochastic mask (IMDM) | MDM with ReDi distillation · LM1B unconditional 2 steps | Gen. PPL: 530.36 -> 284.62 | Improved |
| replace single mask with IMDM | MDM with SDTT+ReDi · LM1B unconditional 2 steps | Gen. PPL: 353.90 -> 165.85 | Improved |
| replace single mask with IMDM | MDM with SDTT · LM1B unconditional 2 steps | Gen. PPL: 584.22 -> 524.31 | Improved |
| replace single mask with IMDM | MDM with SDTT+ReDi · LM1B unconditional 16 steps | Gen. PPL: 55.36 -> 52.51 | Improved |
| replace MDM with IMDM | distilled MDM · synthetic {00,11} one-step validity | validity (%): 49.8% -> 97.7% | Improved |
| replace MDM with IMDM | distilled MDM · synthetic {00,11} | factorization error: 0.693 -> 0.082 | Improved |
| apply IMDM with SDTT+ReDi against Di4C | Di4C · OpenWebText unconditional 2 steps | Gen. PPL: 747.66 -> 500.84 | Improved |
| apply IMDM with SDTT+ReDi against SDTT | SDTT · OpenWebText conditional 2 steps | MAUVE: 0.201 -> 0.810 | Improved |
| apply IMDM at 860M scale | SDTT · OpenWebText 860M unconditional 4 steps | Gen. PPL: 339.89 -> 90.62 | Improved |
| reduce noise dimension from 768 to 256 (ablation) | IMDM · LM1B 2 steps | Gen. PPL: 165.85 -> 294.41 | Hurt |
| use Gaussian instead of Uniform noise (ablation) | IMDM · LM1B 2 steps | Gen. PPL: 165.85 -> 167.48 | No clear effect |
| change noise scale from 1.0 to 0.5 (ablation) | IMDM · LM1B 2 steps | Gen. PPL: 165.85 -> 179.19 | Hurt |
Closest papers by meaning
- Masked Diffusion Models are Secretly Time-Agnostic Masked Models and Exploit Inaccurate Categorical Sampling (ICLR 2025 · similarity 0.92)
- Accelerated Sampling from Masked Diffusion Models via Entropy Bounded Unmasking (NeurIPS 2025 · similarity 0.92)
- Scaling up Masked Diffusion Models on Text (ICLR 2025 · similarity 0.92)
- Theoretical Benefit and Limitation of Diffusion Language Model (NeurIPS 2025 · similarity 0.91)
- Simple and Effective Masked Diffusion Language Models (NeurIPS 2024 · similarity 0.91)
- Looped Diffusion Language Models (NeurIPS 2026 · similarity 0.91)
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.
