Research map / Image generation / Fast diffusion sampling
DiffusionVL: Translating Any Autoregressive Models into Diffusion Vision Language Models
Lunbin Zeng, Jingfeng Yao, Bencheng Liao, Hongyuan Tao, Wenyu Liu, Xinggang Wang
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 |
|---|---|---|---|
| use training block size 1 instead of 8 (ablation) | DiffusionVL-3B · ChartQA | accuracy: 79.9 -> 82.8 | Improved |
| use training block size 16 instead of 8 (ablation) | DiffusionVL-3B · ChartQA | accuracy: 79.9 -> 77.4 | Hurt |
| increase training block size from 1 to 16 (ablation) | DiffusionVL-3B · DetailCaps throughput | TPS: 26.9 -> 41.6 | Improved |
| use block size 8 rather than 1 or 16 (ablation) | DiffusionVL-3B · DetailCaps | BERTScore: 29.8 -> 30.9 | Improved |
| add block-level noise instead of sequence-level noise (ablation) | DiffusionVL-7B · MMMU-Pro (vis.) | accuracy: 24.86 -> 24.97 | Improved |
| add block-level noise instead of sequence-level noise (ablation) | DiffusionVL-7B · MME (perp.) | score: 1527 -> 1519 | Hurt |
| use block diffusion instead of full diffusion (ablation) | DiffusionVL-7B · ChartQA | accuracy: 52.00 -> 84.20 | Improved |
| use block diffusion instead of full diffusion (ablation) | DiffusionVL-7B · MME (perp.) | score: 1607 -> 1519 | Hurt |
| lower dynamic remasking threshold from 1.0 to 0.2 (ablation) | DiffusionVL-3B · DetailCaps | TPS and BERTScore: TPS 31.5 -> 71.8; BERT 30.2 -> 18.6 | Mixed |
| lower dynamic remasking threshold from 1.0 to 0.6 (ablation) | DiffusionVL-3B · DetailCaps | TPS and BERTScore: TPS 31.5 -> 37.8; BERT 30.2 -> 29.7 | Mixed |
| start from AR-LM Qwen2.5-7B instead of dLLM LLaDA-8B for block diffusion VLM | DiffusionVL · ChartQA | accuracy: 29.8 -> 53.6 | Improved |
| use block diffusion instead of AR finetuning on the same AR-LM | DiffusionVL (Qwen2.5-7B) · MMMU val | accuracy: 45.4 -> 43.7 | No clear effect |
Closest papers by meaning
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.92)
- From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.92)
- Vision‑Language‑Vision Auto‑Encoder: Scalable Knowledge Distillation from Diffusion Models (NeurIPS 2025 · similarity 0.91)
- Dual Diffusion for Unified Image Generation and Understanding (CVPR 2025 · similarity 0.91)
- dKV-Cache: The Cache for Diffusion Language Models (NeurIPS 2025 · similarity 0.91)
- SimSD: Simple Speculative Decoding in 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.
