Pixelated Eyes -

We've gone through the threads here related to low res outputs, applied the changes described below and still experiencing issues.

Has anyone else experienced issues lower quality outputs regardless of how high the quality is on the original source footage? Our outputs are noticeably lower in quality, particularly around the eyes (see attached). We found that VideoRetalking uses the GPEN-BFR-512 model for FaceEnhancement, so we switched to the GPEN-BFR-1024 model and updated all size references from 512 to 1024, hoping it would improve the results. Unfortunately, the outcome was the same. We also noticed that Super Resolution (SR) was disabled in the inference code, so we tried enabling it and implementing RealESRGAN_x4eplus.pth, but ran into some challenges there too. Has anyone else faced similar issues? If so, could you guide us on the best approach? Thanks!