hunyuan：

when you've got around 100 or more images in your training set for the same concept, using 10 repeats and a learning rate of 0.00002 (diffusion-pipe), you'll usually see great results around epoch 8. By epoch 10 or 12, your Lora model will be pretty much stable. Now, if you're working with about 50 images, you'll notice that things stop changing much after around 20 epochs. Basically, the fewer images you have in your training set, the more epochs you'll need to train. I think this experience could be helpful for anyone training Hunyuan models.

Flux：etc.