Add model card for Phoenix

README.md

@@ -0,0 +1,66 @@
+---
+license: apache-2.0
+pipeline_tag: image-segmentation
+---
+
+# Phoenix: Learning from Adversity — Semantic-Aware Mask Refinement through Adversarial Perturbation
+
+Phoenix is a model-agnostic refinement layer that turns coarse or noisy segmentation masks into precise object masks. It can be used as a deployment add-on for existing segmenters, a data-quality step for annotations, or an on-demand refinement tool.
+
+- **Paper**: [arXiv](https://arxiv.org/abs/2607.29059) (ECCV 2026)
+- **Project page**: [https://phoenix-eccv26.github.io](https://phoenix-eccv26.github.io)
+- **Code**: [https://github.com/naver-ai/Phoenix](https://github.com/naver-ai/Phoenix)
+- **Demo**: [https://huggingface.co/spaces/naver-iv/phoenix-demo](https://huggingface.co/spaces/naver-iv/phoenix-demo)
+- **Checkpoints**: [https://huggingface.co/naver-iv/phoenix-weights](https://huggingface.co/naver-iv/phoenix-weights)
+- **Evaluation data**: [https://huggingface.co/datasets/naver-iv/phoenix-eval-data](https://huggingface.co/datasets/naver-iv/phoenix-eval-data)
+
+## Quick start
+
+```python
+import numpy as np, phoenix
+from PIL import Image
+
+model   = phoenix.build_phoenix("ckpt/phoenix_efficientvit_xl1.pt")
+refiner = phoenix.PhoenixRefiner(model)
+
+image = np.array(Image.open("image.jpg").convert("RGB"))
+noisy = np.array(Image.open("noisy_mask.png").convert("L"))
+refined = refiner.refine(image, noisy, refine_iters=5)   # -> bool HxW mask
+```
+
+Command line:
+
+```bash
+python infer.py --checkpoint ckpt/phoenix_efficientvit_xl1.pt \
+    --image assets/examples/instance/00_image.jpg \
+    --mask  assets/examples/instance/00_noisy.png \
+    --output refined.png
+```
+
+## Checkpoints
+
+Download all checkpoints from [naver-iv/phoenix-weights](https://huggingface.co/naver-iv/phoenix-weights) into `ckpt/`.
+
+| File | Encoder | Used for |
+|---|---|---|
+| `phoenix_vit_h_instseg.pt` | ViT-H | Instance segmentation (Tables 1 & 2) and Cityscapes (Table S3) |
+| `phoenix_efficientvit_xl1.pt` | EfficientViT-XL1 | Lightweight instance segmentation and the demo |
+| `phoenix_vit_h_dis.pt` | ViT-H | DIS fine-grained segmentation (Table 3) |
+| `phoenix_efficientvit_xl1_dis.pt` | EfficientViT-XL1 | Lightweight DIS model used by the demo |
+| `phoenix_vit_h_voc.pt` | ViT-H | VOC semantic segmentation (Table S2) |
+
+## Citation
+
+```bibtex
+@inproceedings{kim2026phoenix,
+  title     = {Learning from Adversity: Semantic-Aware Mask Refinement
+               through Adversarial Perturbation},
+  author    = {Kim, Beomyoung and Hwang, Sung Ju},
+  booktitle = {European Conference on Computer Vision (ECCV)},
+  year      = {2026},
+}
+```
+
+## License
+
+Apache 2.0. See [LICENSE](https://github.com/naver-ai/Phoenix/blob/main/LICENSE) for details.