I made a SD1.5 LoRA model based on the legendary actress Petra Morgan
Prompt keywords:
p3tr4_morg4n, photorealistic
additional keywords that work good with this prompt: breasts, smile, 1girl, makeup, arms behind head, navel piercing,
Reccomnded weight is 0.8-1.0 for realistic models and >1.4 for anime models.
Description
FAQ
Comments (5)
A wonderful woman. She is also a great cook and
makes a great Lasagna alla bolognese (while naked)
While I am not familiar with the model you are referencing, I did have an issue with hair color. Using your prompt here on CivitAI, I constantly get brown hair. When I add in the word "blonde," it changes her face. Based on my limited knowledge of Lora, I think you did not run the Lora for enough steps. I see you have only made one model here.
When setting up the Lora, I focus on maximizing the total steps. I aim for around 6000 total steps, batching the images with 9 images per step. I aim for 50 head shots and 50 or more images at different distances, wide, medium, close-up shots.
Now, in my opinion, captioning, a step a lot of people half-ass, is the most important. I hand-label; first, I come up with a general set of features describing the subject, and I copy and paste that into every caption. Then, I describe the scene and the action. Finally, I re-paste the general features again into the caption. (Example: "hot blonde Petra, big tits, smiling at camera model, hot blonde Petra, big tits,").
Lastly, I rename the images with the same description (Petra_hot_blonde_big_tits_supermodel001.jpg). Now, some people do not use the name of the actor or model. I disagree and see that Stability suggests doing so with the celeb's full name because that word and likeness are already in the SD model, and basically, you just fine-tune it versus starting from scratch with unrelated words. The words in the caption are the Trigger words. I download all the models made. I do not publish from the modeling training page. I download all models and release a range of them. Generally, the last one created is not the best. Based on my standard 1.5 method, I will do 16 epochs, and generally find 8-12 to be good epochs based on 6000 total steps, with 3500-4500 steps being generally good for me.
Sorry if this was long and you knew it, but I thought you may not know some of this. again i am not a pro but these are the thing that have been working for me.
Seriously thanks fo rthe comment. I stopped making models, but this made me think of starting again. I always try to imrpove so crtiique is welcome. I'm not sure why your generation did not look like min,e I will need to look into it
@barlogspam176 Well the face in similar, it just she get brown hair. it could be just down to different based checkpoints. but to me that is a sign of under trained. I am not a pro. But i certainly think you should keep going.
I tested my lora again and it produces the girl likeness all the same. No idea where yours went wrong.