Custom training data has poor quality on hand

[ch1998]

I started from images using the enerf-outdoor dataset ，used colmap to generate camera parameters and init point clouds. The background quality in the result is very good, but once the characters move there is a noticeable loss of quality.
look at face and hand

[guanjunwu]

maybe you can enlarge the batch size and make the training process longer?
now I also face the same problem and I'm working hard to solve it QwQ