Issues with Training Outcomes After Reducing Model Channel Count to 50% of the Original Paper Specification

[Damehou]

Hello,
I've been working on replicating a model from a paper but decided to modify the channel count to 50% of what's described in the original paper to potentially improve computational efficiency. However, I've encountered an issue where the training results now show noticeable grid-like artifacts, and the overall quality of the results significantly deviates from those presented in the paper.

Modifications Made:

• Reduced the channel count across all layers to 50% of the original specifications.

Observed Issues:

• The model produces grid-like artifacts in the output.
• There is a significant degradation in the quality of the results compared to the original model.

Questions:

1. Is it normal to see such a significant drop in performance and appearance of artifacts when reducing the model’s channel count?
2. Are there recommended strategies to mitigate these issues while still operating under reduced computational resources?
   Any guidance or recommendations on how to address these issues would be greatly appreciated.
   Thank you!