This commit implements several improvements to the face swapping pipeline,
focusing on enhancing output quality and providing optimizations for performance.
Key changes include:
1. **Upgraded Face Swapping Model:**
* Switched from `inswapper_128_fp16.onnx` to the full-precision `inswapper_128.onnx` model. This is expected to provide higher fidelity face swaps.
* Updated the model download logic in `modules/processors/frame/face_swapper.py` accordingly.
2. **Optimized Face Enhancement (GFPGAN):**
* Modified `modules/processors/frame/face_enhancer.py` to set the `upscale` parameter for `GFPGANer` from `1` to `2`. This can improve the detail and perceived quality of faces processed by the enhancer.
3. **Improved Color Matching for Swapped Faces:**
* Implemented statistical color transfer in `modules/processors/frame/face_swapper.py`. This matches the color profile of the swapped face region to the original target face's region, leading to more seamless and natural blending.
* This feature is controlled by a new (assumed) global flag `statistical_color_correction`.
4. **Optimized Mouth Masking Logic:**
* Reduced default `forehead_extension_factor` in `create_face_mask` from `5.0` to `2.5` for slightly faster mask computation.
* Reduced default `mouth_mask_blur_kernel_size` in `create_lower_mouth_mask` from `(15, 15)` to `(9, 9)` to speed up this blur operation.
* These parameters are now fetched using `getattr` to allow future configuration via global variables (e.g., `modules.globals.forehead_extension_factor`).
5. **Performance Analysis & Other Considerations:**
* Identified model inference (swapper, enhancer) as primary GPU workloads.
* Noted that mouth masking (CPU-bound) and the new color correction add overhead. Making these features optional (which they are, via global flags like `mouth_mask` and `statistical_color_correction`) is important for you to balance quality and performance.
* Reviewed face detection usage and found it to be reasonably efficient for the modular pipeline structure.
These changes aim to significantly improve the visual quality of the face swaps and provide some performance tuning options.
rehanbgmi
google-labs-jules[bot]
KRSHH
Kenneth Estanislao
zjy-dev
Jasurbek Odilov
Neurofix
Jonah Hewett
Chou Chamnan
inwchamp1337
Giovanna
Teo Jia Cheng
VilkkuKoo
Gordon Böer
KUNJ SHAH
killerlux
NeuroDonu