I'd like to ask a question about training LoRA
I've run into an issue with a LoRA trained on Krea 2. I noticed that even without tagging the training images, the LoRA still takes effect when loaded—in fact, it works even without a trigger word.
For example, when I train a LoRA of a specific woman, whether the images were tagged or not, generating images with the LoRA loaded still applies the effect regardless of whether I enter the trigger word. On top of that, all female characters end up taking on the features of this LoRA. Why is this happening? This doesn't happen with Qwen Image or ZImage.
Yes, it will affect people especially if you have woman or female in the prompt. It sometime affects men as well if one is in the scene, giving the face of the subject. Not sure what can be done to stop that other than disabling the LoRA.
Yes, it will affect people especially if you have woman or female in the prompt. It sometime affects men as well if one is in the scene, giving the face of the subject. Not sure what can be done to stop that other than disabling the LoRA.
Sure enough, my tests confirmed that male characters are indeed affected as well, and it even leaks into the overall style. But when I tested LoRAs released by Fal, they don't have this issue at all. It's truly baffling.