Research map / Image generation / Fast diffusion sampling
Sparse-LaViDa: Sparse Multimodal Discrete Diffusion Language Models
Shufan Li, Jiuxiang Gu, Kangning Liu, Zhe Lin, Zijun Wei, Aditya Grover, Jason Kuen
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 dense MDM parameterization with sparse truncation, register tokens and caching | LaViDa-O · GenEval | latency (s/image): 21.27 -> 10.86 | Improved |
| replace dense MDM parameterization with sparse one | LaViDa-O · GenEval | overall score: 0.77 -> 0.78 | No clear effect |
| use sparse parameterization | LaViDa-O · ImgEdit | latency (s): 63.98 -> 22.55 | Improved |
| use sparse parameterization | LaViDa-O · ImgEdit | overall score: 3.71 -> 3.79 | Improved |
| use sparse parameterization | LaViDa-O · MathVista | latency (s): 10.41s -> 3.72s | Improved |
| use sparse parameterization instead of Fast-dLLM caching | LaViDa-O · MathVista | accuracy: 56.1 -> 56.7 | Improved |
| use sparse parameterization | LaViDa-O · image understanding suite (Table 5) | accuracy: MME 488 -> 450; MMMU 45.1 -> 43.6; ChartQA 80.0 -> 82.0 | Mixed |
| use sparse parameterization | LaViDa-O · MJHQ-30k | FID: 6.68 -> 7.63 | Hurt |
| train baseline on same 20M data subset vs sparse | Sparse-LaViDa · MJHQ-30k | FID: 8.11 -> 7.63 | Improved |
| add 64 register tokens (vs 0) (ablation) | Sparse-LaViDa · DPG-Bench | score: 80.3 -> 82.4 | Improved |
| add 64 register tokens (vs 0) (ablation) | Sparse-LaViDa · MJHQ-30k | FID: 9.32 -> 7.63 | Improved |
| add 64 register tokens (vs 0) (ablation) | Sparse-LaViDa · GenEval | overall score: 0.76 -> 0.78 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- LaViDa-O — Lavida-O: Elastic Large Masked Diffusion Models for Unified Multimodal Understanding and Generation (ICLR 2026) · better on 8 of 12 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Beyond Masked and Unmasked: Discrete Diffusion Models via Partial Masking (NeurIPS 2025 · similarity 0.90)
- 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)
- Remasking Discrete Diffusion Models with Inference-Time Scaling (NeurIPS 2025 · similarity 0.89)
- Dual Diffusion for Unified Image Generation and Understanding (CVPR 2025 · similarity 0.89)
- The Design Space of Tri-Modal Masked Diffusion Models (NeurIPS 2026 · 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.
