Add model card for BARD-VL

README.md

@@ -0,0 +1,46 @@
+---
+pipeline_tag: image-text-to-text
+---
+
+# BARD-VL
+
+BARD (Bridging AutoRegressive and Diffusion) is a framework designed to convert pretrained autoregressive vision-language models (VLMs) into decoding-efficient diffusion VLMs (dVLMs). By combining progressive supervised block merging with stage-wise intra-dVLM distillation, BARD-VL maintains the strong multimodal capabilities of models like Qwen3-VL while achieving up to 3× decoding throughput speedup.
+
+- **Paper:** [BARD: Bridging AutoRegressive and Diffusion Vision-Language Models Via Highly Efficient Progressive Block Merging and Stage-Wise Distillation](https://huggingface.co/papers/2604.16514)
+- **Repository:** [https://github.com/fudan-generative-vision/Bard-VL](https://github.com/fudan-generative-vision/Bard-VL)
+- **Project Page:** [https://fudan-generative-vision.github.io/Bard-VL](https://fudan-generative-vision.github.io/Bard-VL)
+
+## Model Description
+
+Autoregressive VLMs offer strong multimodal capability, but their token-by-token decoding imposes an inference bottleneck. BARD addresses this by converting these models into same-architecture, large-block diffusion VLMs. Key components include:
+- **Progressive Supervised Block Merging (PBM):** Gradually enlarges the decoding block size.
+- **Stage-Wise Distillation (SWD):** Recovers performance lost at larger blocks using a fixed small-block diffusion anchor.
+
+Experimental results show that BARD-VL establishes new state-of-the-art performance among comparable-scale open dVLMs at 4B and 8B scales.
+
+## Usage
+
+To use this model, please follow the installation and setup instructions in the [official GitHub repository](https://github.com/fudan-generative-vision/Bard-VL). You can run inference for image or video understanding using the provided `inference.py` script:
+
+```bash
+python3 inference.py \
+  --model_id fudan-generative-ai/Bard-VL-B4-Mask-4B-Instruct \
+  --block_size 4 \
+  --denoising_steps 4 \
+  --confidence_threshold 0.6
+```
+
+## Citation
+
+```bibtex
+@article{chen2026bard,
+  title={BARD: Bridging AutoRegressive and Diffusion Vision-Language Models Via Highly Efficient Progressive Block Merging and Stage-Wise Distillation},
+  author={Chen, Baoyou and Xia, Hanchen and Tu, Peng and Shi, Haojun and Mu, Shan and Yuan, Weihao and Zhu, Siyu},
+  journal={arXiv preprint arXiv:2604.16514},
+  year={2026}
+}
+```
+
+## Acknowledgements
+
+This repository builds on top of [NVIDIA NeMo AutoModel](https://github.com/NVIDIA-NeMo/Automodel).