I trained this lora on 25 images of Emma Watson for 1600 steps, enjoy :D
v1.2 improved quality, no trigger required. This is the same model I posted on huggingface a few days ago.
v1.1 updated the model to require the trigger emmwat, after training extensively on triggers, using the person's name reduces model accuracy. please let me know if you prefer the model trigger or not. thanks ^^
v1.0 does not require a trigger word
Below is my flux LORA training guide:
https://civarchive.com/articles/6678
Description
FAQ
Comments (17)
Do you think using her name reduced the accuracy because it already existed in some (probably intentionally mutilated) form in the model or just because it associated with a regular person. Like if you made an unknown persons the trigger would it work like the made up trigger word.
yes i do, if you generate images of Emma Watson with out the lora it always generates a girl with similar features to her but does not look like her. this bleeds in with the lora and reduces the likeness of the model.
@punzel Interesting, had a feeling it was happening with some of my loras. Time to name a few celebs as john smith.
@punzel By the way have you tried the civitai trainer? Your guide is fantastic and gave me great results but civit is about half the price plus you can generate buzz from your work to even lower that.
@JaneB ooo ill have to check that out!! thanks!
@punzel Please report back, I trust your opinion more than I trust mine :)
@JaneB i will! :3
Thanks for improving. DO you believe if the person is not very famous adding her real name would affect the LoRA performance?
PS we are rediscovering things from the SDXL loRA
Use alphanumerical triggernames, use 2000 steps, etc.
hii! if the trigger generates a similar person each time, it may affect the model; if it's random, it shouldn't.
This lora is almost perfect, great job!
amazing job, What settings did you use in the last version?
hii, thank you!! the guide linked in the description has all the settings i used!
Very nice 👍🙂
So good
Hello another question, In the workflow it says unet, What exactly model are you using? Could you share the model?
https://huggingface.co/black-forest-labs/FLUX.1-dev/tree/main
flux1-dev.safetensors
@punzel where is the "Workflow" your talking about? Cant find it..