
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. With fp8, the training needed a lot less steps, and a rank 4 did work, but with far too many steps. So I increased to 88 images and rank 16 and that learned a lot better. The main keyword is looking over his/her glasses. Additionally you can use lowered on his/her nose, but that is not necessary in most cases. The LoRA tag <lora:LookingOverGlasses-step01760:1> increases reliablity.
With my LoRAs for the small wire-frame glasses you have a chance that Flux already makes this pose from its own, but here is also a LoRA to get it reliably: a woman "looking seductively over her glasses" which is also the keyphrase I used for the training. The strength varies wildly. You don't expect it, but it is easier with small glasses like in reality. If you use it with the small glasses LoRAs, you can go as low as 0.5 without anything else. With the big vanilla Flux glasses, you may need 1.0 and a additional <lora:LookingOverGlasses-30:1> tag. But so far, it looks like you can retain the shape from the first LoRA and add only the pose if you get the strength right.
描述:
训练词语: looking seductively over her glasses,looking over glasses
名称: LookingOverGlasses-30.safetensors
大小 (KB): 25641
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success