If its not too much, can you train with the data aligned using insightface code? the area of face/total area of a 128x128 image is much larger compared to that of ffhq data.
My hypothesis is that, with that data, the model should produce better result.
If its not too much, can you train with the data aligned using insightface code? the area of face/total area of a 128x128 image is much larger compared to that of ffhq data.
My hypothesis is that, with that data, the model should produce better result.