PST50: Benchmark for Photorealistic Style Transfer

PST50 is the first benchmark dataset designed for rigorous evaluation of Photorealistic Style Transfer (PST). It includes high-resolution, professionally curated content-style pairs with ground truth stylizations, suitable for both paired and unpaired evaluation protocols.

This dataset is introduced in the paper:
SA-LUT: Spatial Adaptive 4D Look-Up Table for Photorealistic Style Transfer Project Page • Paper • Code

---

📁 Dataset Structure

PST50/
├── video/            # Content videos in log color space
├── content_709/      # Content images in Rec.709 color space
├── content_log/      # Content image in log color space
├── paired_gt/        # Ground truth stylized outputs for paired evaluation
├── paired style/     # Style references for paired evaluation
└── unpaired_style/   # Style references for unpaired evaluation

🧪 Evaluation Protocol

• Paired: Compare stylized outputs against paired_gt using metrics like LPIPS, PSNR, SSIM, and H-Corr.
• Unpaired: Use unpaired_style with stylized outputs for qualitative or perceptual studies.
• Video: Apply models to video sequences to test temporal consistency and real-time stylization.

---

📜 License: cc-by-4.0