I have 3090, 30 training images, and training time is 2 hours 20 minutes/ SOLVED

[trollver9000]

This cant be normal if your 3060 was 1 hour? Is it too many images?

[ThetaCursed]

The difference in training time is most likely related to your image resolutions and bucket settings. Here is why:

• Image Sizes & Bucketing: If you didn't use the built-in Smart Aspect Ratio Bucketing tool, or if your Min Side (Base) and Max Side (Limit) are not set to 512 and 768 respectively, the training will take significantly longer. Processing oversized or unbucketed images slows down the GPU dramatically, even on a RTX 3090.
• Dataset Size: 30 images is definitely not too many. Personally, I regularly use up to 100 images without changing the default training steps and it runs perfectly fine.

Run the Smart Aspect Ratio Bucketing tool on your dataset first, make sure the min/max sides are 512/768, and try training again. Your 3090 should then process it much faster.

[trollver9000]

with 22 images and bucketing process now it takes 50 minutes to train, is that a good result ? So You say 30 images vs 100 images gets same training time ? Whats good image count to not overtrain?