AI BriefingAFP-GIC Framework Published in IEEE Access for Controllable Generative Image Compression
Key point
The new AFP-GIC framework reduces decoder latency by 18.1% and parameter count by 20.5% compared to DC-VIC while supporting five bitrate operating points.
Details
AFP-GIC, a framework for controllable generative image compression, has been published in IEEE Access (2026) with its codebase and interactive playground now available. The system addresses bottlenecks in ultra-low bitrate compression, where standard learned codecs suffer from local distortion and generative models introduce hallucinations, by using an asymmetric Adaptive Fused Prior Transfer pipeline that enables prior-guided texture reconstruction without transmitting the fused prior.
Performance Metrics
Benchmarked on an NVIDIA RTX 4090, AFP-GIC demonstrates significant efficiency gains over the state-of-the-art DC-VIC model:
- Latency: Decoding time is reduced to 80.47 ms (vs. 98.27 ms for DC-VIC), an 18.1% improvement measured on 256×256 patches.
- Parameters: The model uses 120.6M inference parameters, which is 20.5% fewer than DC-VIC's 151.7M.
- Control: A single pretrained model supports toggling across 5 target bitrate operating points.
Resources
The authors have released the full codebase on GitHub and an interactive Hugging Face Space. Additionally, 2,760 reconstructed images and associated metric CSVs are packaged in GitHub Releases to facilitate academic cross-evaluation.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.