Research map / Image generation / Fast diffusion sampling
Guided Star-Shaped Masked Diffusion
Viacheslav Meshchaninov, Egor Shibaev, Artem Makoian, Ivan Klimov, Nikita Balagansky, Daniil Gavrilov, Aibek Alanov, Dmitry Vetrov
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 |
|---|---|---|---|
| apply pure star-shaped sampler over the whole trajectory (ablation) | MDLM sampling · OWT 128 tokens | MAUVE: near-zero MAUVE at t_on=1.0 | Hurt |
| switch from MDLM to star-shaped sampler late in generation (t_on about 0.3) (ablation) | MDLM · OWT 128 tokens | MAUVE / perplexity: peak at t_on about 0.3 | Improved |
| add learned error predictor to unguided Star+ (G-Star+) (ablation) | Star+ · OWT 512 tokens, 32-512 steps | MAUVE: largest gap at 64-256 steps | Improved |
| use head-only trained 12-block predictor vs fully fine-tuned (ablation) | G-Star error predictor · OWT 128 steps | Pareto quality-diversity | No clear effect |
| shrink predictor to a single fine-tuned block (1B,F) (ablation) | G-Star error predictor · OWT 128 steps | Pareto quality-diversity | Hurt |
| lower error-predictor temperature (ablation) | G-Star · OWT 512 tokens | perplexity vs diversity: low T: best perplexity, lower diversity | Mixed |
| use 128-step G-Star-loop instead of 512-step ReMDM | G-Star-loop · OWT 512 tokens | latency (s): 9.16 -> 3.43 | Improved |
| add G-Star refinement to Dream-Instruct 7B | Dream-Instruct 7B · IFEval | score: 56.4 -> 59.3 | Improved |
| add G-Star refinement to Dream-Instruct 7B | Dream-Instruct 7B · MMLU | score: 69.9 -> 71.2 | Improved |
| add G-Star refinement to Dream-Instruct 7B | Dream-Instruct 7B · GSM8K | score: 81.5 -> 81.6 | No clear effect |
| use G-Star instead of ReMDM / MDLM | Conala conditional MDLM · Conala 128 steps | Qwen2.5-Coder ppl: 19.1 -> 16.4 | Improved |
| use 12-block full predictor (ablation) | G-Star · OWT | inference latency: about 50% increase | Hurt |
Closest papers by meaning
- Accelerated Sampling from Masked Diffusion Models via Entropy Bounded Unmasking (NeurIPS 2025 · similarity 0.90)
- Learn from Your Mistakes: Self-Correcting Masked Diffusion Models (NeurIPS 2026 · similarity 0.90)
- Remasking Discrete Diffusion Models with Inference-Time Scaling (NeurIPS 2025 · similarity 0.89)
- Beyond Masked and Unmasked: Discrete Diffusion Models via Partial Masking (NeurIPS 2025 · similarity 0.89)
- Simplified and Generalized Masked Diffusion for Discrete Data (NeurIPS 2024 · similarity 0.89)
- Masked Diffusion Models are Secretly Time-Agnostic Masked Models and Exploit Inaccurate Categorical Sampling (ICLR 2025 · 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.
