Add model card for RAEv2

README.md

@@ -0,0 +1,38 @@
+---
+license: apache-2.0
+pipeline_tag: image-to-image
+---
+
+# Improved Baselines with Representation Autoencoders (RAEv2)
+
+RAEv2 simplifies and improves representation autoencoders (RAE), achieving over 10x faster convergence, better generation, and better reconstruction compared to traditional VAEs.
+
+[**Paper**](https://arxiv.org/abs/2605.18324) | [**Project Page**](https://raev2.github.io) | [**GitHub**](https://github.com/nanovisionx/RAEv2)
+
+## Overview
+Representation Autoencoders (RAE) replace traditional VAEs with pretrained vision encoders. RAEv2 introduces three key insights to improve this architecture:
+1. **Generalized Formulation**: Defining the representation as the sum of the last *k* encoder layers significantly improves reconstruction.
+2. **RAE and REPA Synergy**: Representation Alignment (REPA) and RAE are found to be complementary, allowing the same representation to serve as both an encoder and a target.
+3. **Free Guidance**: By re-parameterizing the DiT model output, REPA can provide classifier-free guidance without needing a second diffusion model.
+
+RAEv2 achieves a state-of-the-art gFID of 1.06 in just 80 epochs on ImageNet-256.
+
+## Sample Usage
+
+To reconstruct an image using a trained RAE model, you can use the sampling script provided in the [official repository](https://github.com/nanovisionx/RAEv2):
+
+```bash
+uv run python scripts/stage1/sample.py \
+    --config configs/stage1/sampling/dinov3l-k23-general.yaml \
+    --image assets/samples/sample_1.png
+```
+
+## Citation
+```bibtex
+@article{singh2026raev2,
+  title={Improved Baselines with Representation Autoencoders},
+  author={Singh, Jaskirat and Zheng, Boyang and Wu, Zongze and Zhang, Richard and Shechtman, Eli and Xie, Saining},
+  journal={arXiv preprint arXiv:2605.18324},
+  year={2026}
+}
+```