LoRA model of Adult Actress Riley Reid.
trained on a variety of pictures of her, 81 pics in total, mostly ignoring general rules of Ai training datasets, 50% were close up, 50% were full body, Many of the pictures featured object and substances that partically covered her face. Still in my opinion it came out very good.
Trained to 4 epochs out of 5 with standard civitai training parameters. Captioned with JoyCaption Pre-Alpha
The only tag words is "Riley Reid" additional tags words, "wavy hair, light compelxition, green eyes" but arenot really necessary.
Description
FAQ
Comments (10)
Any comments/suggestions are welcome. do you believe It is undertrained and does not look like her?
Should I increase the epochs or just remove most of full body pictures?
I'm not entirely sure if it's under-trained, but I really like your LoRa. There's some inconsistency in the images some of them capture her perfectly, while in others, she resembles Melissa Moore a bit more. That said, it's still the best Riley Reid model I've come across so far.
From my recent experience with training Flux over the past few days, don't hesitate to push it a little further. Flux seems to handle a lot before showing signs of overtraining. You can safely go up to 3000 or even 4000 steps, so don't be afraid to set the image repeat high with the on-site trainer. I usually aim for 20 epochs, finding that 10 is decent, 15 is better, and 18/19 is often ideal 20 starts to feel a bit over-trained.
Also, be sure to balance your dataset with a mix of full-body shots and close-ups. If there's not enough variety, the model can sometimes overtrain on full-body shots, so the diversity in images really helps.
@MakeThemComeAliveAIArt thank you. Seriously this kind of advice is what I needed.
Yeah my modle was 4 epochs and 304. I didn't like the look of epoch 5 so i skipped it. next time i will train her for more epoch and do a better selection.
I agree that this model can be made to look much more like her.
@barlogspam176 20 Epoch is not ideal though it's whats best on the website, if you train locally you can do so much more.
Hi, thank you for making the lora. Since you requested comments, I agree it is kind of hit or miss to get a successful generation from this. When it works, it works really nicely. But sometimes there's no clear indication of who it is. I apologize I don't have specific suggestions for improvement as I don't know how to train. I am sure it is a lot of work and trial and error!
@EnragedAntelope Thanks. I'm making a v2 right now. it is arelady better. When I'm sure I'm gonna update this.
@barlogspam176 Awesome! Keep up the good work!
@EnragedAntelope Done. Also Melissa Moore is ready, needs just polishing .
New version seems easier to work with!
"object and substances that partically covered her face"... I guess we shouldn't speculate on what these are!