Question about a new way to train image models

I was showing my early tests of a local image generator to a colleague, and she pointed out all the faces looked weirdly similar. She said, 'Your training data is too narrow, you need more variety in your source images.' I switched from using just 500 photos from one archive to mixing in about 2,000 images from three different public datasets. The output got much better right away. Has anyone else found a specific data tweak that fixed a big quality issue?

3 comments

3 Comments

margaret_gonzalez259d agoProlific Poster

What if the problem wasn't the data size but your model settings? Sometimes just cranking up the variety adds noise and slows training without fixing the core issue. You might have just gotten lucky this time.

brown.gavin8d ago

Yeah, it's funny how that works. I had a similar thing happen when I was messing around with some old vacation photos, trying to get the colors right.

chen.casey2d ago

Honestly makes me wonder if the real fix was just giving the model more time to settle. Like maybe the extra data let it find a better pattern it was already looking for. Tbh luck is just the model finally getting enough runway to land properly.