I'm train it for test quality how good is checkpoint. If you need a LoRA that is possible to extract it from checkpoint.

To test how much a full checkpoint will improve quality compared to a LoRA trained with the same dataset. And they really improve color accuracy, reduce artifacts, but the style is still not fully captured. I use an Adafactor type of training, so more epochs = better understanding. And 500 epochs will be a lot for a model like this one. So I train up to 200 epochs to see how good or not good that is. Also, I tested how captions affect the trained model and found that precise captions lead to faster understanding of the dataset during training. I mean, right precise captions can increase the overall speed of training. For example, without captions, 150 epochs look great, whereas with detailed precise captions, 125 epochs also look great. So, to get a similar result, you need fewer epochs to train due to the right captions. I use a Gemini 2.0 Flash Experimental model for captions with my own ComfyUI workflow. Another point: some people use it online, so there's not a big difference between using a LoRA or a Checkpoint for generation.