Research map / Image generation / Fast diffusion sampling
Argmax Flows and Multinomial Diffusion: Learning Categorical Distributions
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 |
|---|---|---|---|
| replace variational dequantization with argmax lifting | AR flow · text8 | bpc: 1.43 -> 1.38 | Helped |
| replace variational dequantization with argmax lifting | coupling flow · text8 | bpc: 2.08 -> 1.82 | Helped |
| replace variational dequantization with argmax lifting | AR flow · enwik8 | bits per byte: 1.44 -> 1.42 | Helped |
| replace uniform dequantization with argmax lifting | AR flow · text8 | bpc: 1.90 -> 1.38 | Helped |
| use Multinomial Diffusion instead of Argmax coupling flow | text models · text8 | bpc: 1.82 -> 1.72 | Helped |
| use autoregressive Argmax Flow instead of Multinomial Text Diffusion | text models · text8 | bpc: 1.72 -> 1.39 | Helped |
| use autoregressive Argmax Flow instead of CategoricalNF | text models · text8 | bpc: 1.45 -> 1.43 | Helped |
| use Argmax Flow instead of discrete flow reimplementation | flows · text8 | bpc: 4.13 -> 1.38 | Helped |
| lower Cartesian base (binary instead of base 27) (ablation) | Argmax AR flow · text8 | bpc: 1.45 -> 1.44 | Helped |
| reduce diffusion steps from 1000 to 100 (ablation) | Multinomial Text Diffusion · text8 | sample time (s): 26.6 -> 2.4 | Helped |
| use Multinomial Diffusion instead of Argmax flow on segmentation | Cityscapes · bits per pixel | bpp: 0.37 test vs 0.33 train | Hurt |
| use softplus vs Gumbel thresholding for q (ablation) | Argmax Flow · Cityscapes | ELBO/IWBO | Mixed |
Closest papers by meaning
- Discrete Modeling via Boundary Conditional Diffusion Processes (NeurIPS 2024 · similarity 0.88)
- Glauber Generative Model: Discrete Diffusion Models via Binary Classification (ICLR 2025 · similarity 0.87)
- Rethinking Vector Field Learning for Generative Segmentation (NeurIPS 2026 · similarity 0.87)
- Diffusion PID: Interpreting Diffusion via Partial Information Decomposition (NeurIPS 2024 · similarity 0.87)
- Discrete Flow Matching (NeurIPS 2024 · similarity 0.87)
- Kaleido Diffusion: Improving Conditional Diffusion Models with Autoregressive Latent Modeling (NeurIPS 2024 · similarity 0.87)
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.