Retrain system with larger SignWriting

[AmitMY]

Currently, we represent SignWriting as 512x512 images, but we have a lot of whitespace, and the model does not follow correctly.

Instead, I would like to try with 2x images

Original	New

[bigbraindump]

training res: 512x512

Results:
Normal Pipeline

Img2Img Pipeline

notes:

• model follows seed more than control image, line art style maintained
• similar behavior to previous pr (addressing issue #7, results and next steps #8) when validating 256 trained model incorrectly with 512 images
• loss monitoring and lr adjustments didn't improve results
• subjectively, 256 res training showed better control/learning

question:
Any suggestions or comments?