
Version 2.0 update:
after I connected the grid pattern sometimes visible in bright or dark parts of the images with the fp4 training with OneTrainer, I switched to fp8 training, which got rid of the pattern. But I have to redo all my LoRAs to fix them all.
Due to the higher precision, the success rate is a bit higher than in v1.0 - i tested in batches of four and depending on the prompt you get more better images than with the previous version. Also try the lora-tag <lora:MaryJaneShoes-step19170:1>
Another thing which Flux is not really good at are Mary Jane shoes - also known as bar shoes. These come in a wide variety, with several attributes in different combinations. So instead of my previous one-concept LoRAs, I tried a multi-concept approach and with a bigger rank and alpha it worked. The final version 1.0 is a merge of 100 epochs of the full dataset with a second run of only the square toe and double strap images of 50 epochs, because these variants weren't strong enough.
The main trigger word is "Mary Jane shoes", and as always it seems that a lora-tag <lora:MaryJaneShoes:1> increases the reliability.
The LoRA is trained on these attributes:
round toe | pointy toe | square toe
low heel | medium block heel | medium kitten heel | high block heel | high stiletto heel | high square heel
and additionally the options wide strap, double strap and open-sided
Have fun to make some new combinations!
描述:
Version 2.0 is trained with fp8, getting rid of the grid pattern sometimes visible in the images.
训练词语: Mary Jane shoes,pointy toe,round toe,square toe,high stiletto heel,high block heel,medium block heel,medium kitten heel,low heel,open-sided,wide strap
名称: MaryJaneShoes-step19170.safetensors
大小 (KB): 149716
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success