Research map / Image generation / Fast diffusion sampling
MinerU-Diffusion: Rethinking Document OCR as Inverse Rendering via Diffusion Decoding
Hejun Dong, Junbo Niu, Bin Wang, Weijun Zeng, Wentao Zhang, Conghui He
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 |
|---|---|---|---|
| add Stage 2 uncertainty-driven fine-tuning after Stage 1 (ablation) | MinerU-Diffusion · OmniDocBench v1.5 (w/o GT layout) | Overall: 86.13 -> 88.94 | Improved |
| add Stage 1 foundational training before Stage 2 (ablation) | MinerU-Diffusion · OmniDocBench v1.5 (w/o GT layout) | Overall: 35.71 -> 88.94 | Improved |
| add Stage 2 on top of Stage 1 (ablation) | MinerU-Diffusion · OmniDocBench v1.5 (w/ GT layout) | Overall: 92.89 -> 93.37 | Improved |
| use dynamic confidence decoding instead of static step=6 (ablation) | MinerU-Diffusion · OmniDocBench v1.5 (w/ GT layout) | Overall: 88.31 -> 93.34 | Improved |
| use dynamic confidence decoding instead of static step=6 (ablation) | MinerU-Diffusion · decoding throughput | TPS: 91.56 -> 98.32 | Improved |
| increase static decoding steps from 6 to 32 (ablation) | MinerU-Diffusion · OmniDocBench v1.5 (w/ GT layout) | Overall: 88.31 -> 93.02 | Improved |
| raise confidence threshold from 0.5 to 0.95 (ablation) | MinerU-Diffusion · OmniDocBench v1.5 (w/ GT layout) | Overall: 87.61 -> 93.37 | Improved |
| raise confidence threshold from 0.5 to 0.95 (ablation) | MinerU-Diffusion · OmniDocBench v1.5 (w/ GT layout) | Table TEDS: 81.10 -> 90.99 | Improved |
| raise confidence threshold to 0.99 (ablation) | MinerU-Diffusion · OmniDocBench v1.5 (w/ GT layout) | Overall | Hurt |
| lower confidence threshold to 0.6 for speed (ablation) | MinerU-Diffusion · OmniDocBench v1.5 | TPS: 52 -> 164.8 | Improved |
| use block attention instead of full attention (ablation) | MinerU-Diffusion (LLaDA-MoE-7B) · OCR table generation | memory/latency/repetition | Improved |
| replace autoregressive decoding with diffusion decoding | MinerU-Diffusion vs AR · Semantic Shuffle (112 FOX images) | robustness to shuffled words | Improved |
Closest papers by meaning
- PiD: Fast and High-Resolution Latent Decoding with Pixel Diffusion (NeurIPS 2026 · similarity 0.86)
- From Denoising to Refining: A Corrective Framework for Vision-Language Diffusion Model (NeurIPS 2026 · similarity 0.86)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.86)
- Encoder-Decoder Diffusion Language Models for Efficient Training and Inference (NeurIPS 2025 · similarity 0.86)
- Würstchen: An Efficient Architecture for Large-Scale Text-to-Image Diffusion Models (ICLR 2024 · similarity 0.85)
- PP-OCRv5: A Specialized 5M-Parameter Model Rivaling Billion-Parameter Vision-Language Models on OCR Tasks (CVPR 2026 · similarity 0.85)
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.
