White Socks Girl_Test Version_flux.1-fp 8_Lora

Finally, I get a chance to try training FLUX.1's Lora. [Note: I will post the same images with their generation data in my homepage. It is a bit wired that TensorArt does not have the function which allows the images to contain the generation data as the local text2img does.] The trigger word here is "wsg", and it should function better if you also input propmts "white socks" and "no shoes“. Sometimes no shoes might not work, if that happen, try giving it more weight. Most of the poses in training set is sitting, so it might perform better in this pose, but you can try the pose you like anyway. The best weight based on my test is 0.75-0.8. For the sampler, I only used "Euler a", but it may perform better on other samplers which might fit the specific checkpoint better. And maybe it is because the images I used in training set are all 640*640 full body images, which only contains non-toe socks but might not be clear enough for FLUX.1 to learn, sometimes it is inevitable that the non-toe socks becomes toe socks. Anyway, it is the first time I give it a try with FLUX.1. I may change the dataset and parameters to figure out how to make it better later. Have fun with it.