AFP-GIC：可控生成式圖像壓縮的自適應融合先驗遷移

🚀 AFP-GIC: Controllable Generative Image Compression (Published in IEEE Access, 2026)

We are excited to release AFP-GIC, a controllable generative image codec, together with its pretrained model, inference framework, reconstructed images, benchmark metrics, and an interactive demo.

167:1 compression in the example shown: a 768×512 image becomes a 3.57 KiB bitstream at 0.0744 bpp, without resizing. This ratio compares the source PNG file with the compressed bitstream, including headers; results vary with image content and source format.

💡 The Core Challenge We Address

Generative low-bitrate compression must balance realistic detail with fidelity to the source image. Hallucinated textures and geometric distortions remain important challenges. AFP-GIC introduces an image-adaptive fused prior transfer pipeline that guides encoding and reconstruction according to image content, without transmitting the prior itself.

🏆 Key Technical Takeaways

📦 One Pretrained Model, Multi-Rate Control: Switch across five bitrate operating points using a single deployable pretrained model, eliminating the need to reload separate model weights between settings.
🏎️ Hardware-Efficient Execution: Delivers 18.1% lower decoder latency (80.47 ms vs. 98.27 ms) and 20.5% fewer inference parameters than DC-VIC, measured on an NVIDIA RTX 4090 using 256×256 patches.
🖼️ Content-Adaptive Reconstruction: Combines competitive PSNR/SSIM with improved no-reference naturalness measured by NIQE, targeting natural-looking reconstructions at very low bitrates.

🧪 Links and Evaluation Playground

Explore the implementation, test your own images, and inspect the results:

🤗 Live Interactive Playground: Launch Hugging Face Space
Upload your own images, compress and decompress them, compare reconstructions, and download bitstreams and metrics. The public demo currently performs inference on CPU, so processing can take some time.
💻 Official Source Code: GitHub Repository
PyTorch and CompressAI-based inference and evaluation tools, with a released pretrained model.
📥 Reconstructed Images and Metrics: GitHub Releases
Includes 2,760 reconstructed images, with per-image and average metrics across Kodak, CLIC2020, and DIV2K at five operating points, supporting direct baseline comparisons under matched evaluation protocols without rerunning the model.
📄 Official Publication: IEEE Xplore Digital Library

If you find AFP-GIC useful, please ⭐ star the GitHub repository and 🤗 like the Hugging Face Space. Your support helps more people discover and build on this work!