Research map / Image generation / Autoregressive and latent generation
Continuously Augmented Discrete Diffusion model for Categorical Generative Modeling
Huangjie Zheng, Shansan Gong, Ruixiang ZHANG, Tianrong Chen, Jiatao Gu, Mingyuan Zhou, Navdeep Jaitly, Yizhe Zhang
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 |
|---|---|---|---|
| augment masked diffusion with paired continuous latent (CADD) | MDM · CIFAR-10 | FID: 4.66 -> 2.88 | Improved |
| augment masked diffusion with continuous latent | MDM · ImageNet 32x32 | FID: 7.91 -> 3.74 | Improved |
| use CADD loss instead of MDM loss for code diffusion LM | DiffuCoder · HumanEval | pass@1: 67.1 -> 72.0 | Improved |
| use CADD loss instead of MDM loss for code diffusion LM | DiffuCoder · BigCodeBench-Hard | pass@1: 12.8 -> 17.6 | Improved |
| use CADD loss instead of MDM loss | DiffuCoder · MBPP | pass@1: 74.2 -> 75.7 | Improved |
| fine-tune DiffuCoder checkpoint with CADD loss | DiffuCoder · MBPP | pass@1: 74.2 -> 73.9 | Hurt |
| fine-tune DiffuCoder checkpoint with CADD loss | DiffuCoder · HumanEval | pass@1: 67.1 -> 73.8 | Improved |
| increase sampling steps from 1024 to 4096 (ablation) | CADD · OpenWebText generation | generative perplexity: 44.6 -> 35.3 | Improved |
| use CADD instead of MDLM on LM1B discrete perplexity | MDLM · LM1B | test perplexity: 31.8 -> 31.4 | Improved |
| use CADD instead of RDLM on Text8 | RDLM · Text8 | bits per character: 1.32 -> 1.35 | Hurt |
| use CADD vs MDLM zero-shot perplexity | MDLM · 7 zero-shot datasets | perplexity: CADD better on 4/7 tasks | Mixed |
| fine-tune MDLM checkpoint with CADD objective for 50B tokens | MDLM · OpenWebText generation | MAUVE and generative perplexity | Improved |
Closest papers by meaning
- Beyond Masked and Unmasked: Discrete Diffusion Models via Partial Masking (NeurIPS 2025 · similarity 0.91)
- Masked Diffusion Models are Secretly Time-Agnostic Masked Models and Exploit Inaccurate Categorical Sampling (ICLR 2025 · similarity 0.91)
- Continuous Latent Diffusion Language Model (NeurIPS 2026 · similarity 0.91)
- Simplified and Generalized Masked Diffusion for Discrete Data (NeurIPS 2024 · similarity 0.90)
- Simple and Effective Masked Diffusion Language Models (NeurIPS 2024 · similarity 0.90)
- Compositional Discrete Latent Code for High Fidelity, Productive Diffusion Models (NeurIPS 2025 · 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.
