We are working on CoderAI (/storage/coderai), a local AI inference server with a web studio UI. The codebase is Python/FastAPI on the backend and vanilla JS/HTML on the frontend (codai/admin/templates/chat.html).
We want to expand the face swap feature (codai/api/faceswap.py) to support multiple backends and models, so users can choose the best option for their hardware and use case. The same philosophy should apply to deblur (/v1/images/deblur) and unpixelate/upscale (/v1/images/unpixelate) — currently both use single hardcoded approaches (OpenCV Wiener and Real-ESRGAN respectively), but users should be able to pick between alternatives there too.
The goal is maximum user choice across all three features:
For face swap, consider offering at minimum:
- The current insightface + inswapper_128 path
- SimSwap or other ONNX-compatible swapper models
- Optional post-processing enhancers (CodeFormer, GFPGAN, or similar) that can be toggled on top of any swapper
- A facefusion-based path if it simplifies supporting multiple models via subprocess
For deblur, alternatives to pure OpenCV signal processing could include ML-based blind deblurring models (e.g. NAFNet, Restormer, or similar lightweight restoration networks).
For unpixelate/upscale, alternatives to Real-ESRGAN could include ESRGAN variants, SwinIR, HAT, or any other super-resolution model the user has downloaded.
The API should accept a model or backend parameter so the caller can select which implementation to use. Missing models should fail gracefully with a clear error rather than silently falling back. The web UI should expose the available options (discovered at runtime based on what's installed/downloaded) as a selector in the relevant panels.
Read the existing implementations before proposing changes to understand the current structure, file layout, and how the UI communicates with the backend.