Add model card, pipeline tag and links to paper/code

README.md

@@ -1,3 +1,73 @@
----
-license: mit
----
+---
+license: mit
+pipeline_tag: unconditional-image-generation
+---
+
+# V-Co: A Closer Look at Visual Representation Alignment via Co-Denoising
+
+Official implementation of **V-Co**, a systematic study of visual co-denoising within a unified pixel-space diffusion framework.
+
+- **Paper:** [V-Co: A Closer Look at Visual Representation Alignment via Co-Denoising](https://huggingface.co/papers/2603.16792)
+- **GitHub Repository:** [HL-hanlin/V-Co](https://github.com/HL-hanlin/V-Co)
+
+## Method Description
+
+V-Co presents a principled study of visual representation alignment via co-denoising in pixel-space diffusion, systematically isolating the effects of architecture, CFG design, auxiliary losses, and feature calibration. The authors identify four key ingredients for effective visual co-denoising:
+
+1.  **Fully dual-stream architecture** to enable flexible cross-stream interaction.
+2.  **Structural masking** for unconditional classifier-free guidance (CFG) prediction.
+3.  **Perceptual-drifting hybrid loss** for stronger semantic supervision.
+4.  **RMS-based feature rescaling** for stable cross-stream calibration.
+
+Together, these findings yield a simple recipe for visual co-denoising. Experiments on ImageNet-256 show that, at comparable model sizes, V-Co outperforms the underlying pixel-space diffusion baseline (JiT) and strong prior pixel-diffusion methods while using fewer training epochs.
+
+## Evaluation
+
+To evaluate a pretrained V-Co model (e.g., VCo-B) on ImageNet 256x256, you can use the following command from the official repository:
+
+```bash
+torchrun --nproc_per_node=8 --nnodes=1 \
+    main_vco.py \
+    --P_mean -0.8 --P_std 0.8 \
+    --img_size 256 \
+    --batch_size 32 \
+    --gen_bsz 32 \
+    --resume '/path/to/vco_base/checkpoint.pth' \
+    --online_eval \
+    --model JiT-B/16-co \
+    --output_dir '/path/to/output_dir' \
+    --data_path '/path/to/imagenet/' \
+    --evaluate_gen \
+    --num_workers 12 \
+    --use_co_embed \
+    --use_dinov2 \
+    --use_dino_from_rae \
+    --use_gated_co_embed \
+    --use_mmdit \
+    --separate_qkv \
+    --use_conv2d_dino_proj \
+    --uncond_dino_null \
+    --dinov2_null_type 'attn_mask_asymmetric' \
+    --dinov2_drop_zero_loss \
+    --class_balanced_sampling \
+    --drifting_v3_loss \
+    --cfg_sweep '1.8'
+```
+
+## Citation
+
+```bibtex
+@misc{lin2026vcocloserlookvisual,
+      title={V-Co: A Closer Look at Visual Representation Alignment via Co-Denoising}, 
+      author={Han Lin and Xichen Pan and Zun Wang and Yue Zhang and Chu Wang and Jaemin Cho and Mohit Bansal},
+      year={2026},
+      eprint={2603.16792},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV},
+      url={https://arxiv.org/abs/2603.16792}, 
+}
+```
+
+## Acknowledgements
+
+The development of V-Co has been inspired by [JiT](https://github.com/LTH14/JiT) and [Latent Forcing](https://github.com/AlanBaade/LatentForcing).