V2.1 : After a lot testing I improved the model a tad bit.This will increase the success rate around up to 40-50% I guess.For weighting 0.4-0.6 gives me nice results.Add "perfect hands"in positive and "disfigured hands" in the negative prompt.I won't be making a new version probably but we'll see :)
V1 : Okay I know nails and hands are the nemesis of ai. So I tried to mitigate this to some extend but results are okayish at best.I cherry picked the images so you will get a lot of errors as a result.Feel free to experiment and you can give feedback if you want.I might make a following to this to correct the mistakes.
Recommended settings: 0.25-0.5 weights with hiresfix 4x_foolhardy_Remacri upscaler.Ancestral sampler might give tiny bit better results.
Support me on buymeacoffee if you like my work.
Description
FAQ
Comments (8)
hey, I have seen your lora's that you have trained and they are astonishing. wanted to know what are the parameters you have been using to train them as for me. When I train to get better quality, in training it shows more than 40k steps and still not getting good results
Apperciated for your thoughts :) There isn't anything special that I make really.I usually use 15 to 25 images for per model and 107 to 135 repeats.But the key is that they are tailored specifically to benefit training. Stay away from anything obscuring the face even the hands ! I use Kohya SS for training like everyone else.For describing the image use as much as humane discriptions as possible.Don't use robotic discriptions like 1girl,shirt,pants,solo etc.For the training parameters I use "Aitrepreneur" settings and low vram json file. You can also look at my model discriptions in the external networks tab in automatic1111.There is a "i" word with a circle on the model that shows the parameters and discriptions.Like I said there isn't anything special that I make for training.It's the image set and discriptions what makes the difference.Hope that helps ^^
@PickleRick_1856 Oh sure, that will help me a lot, and if you could elaborate on the word human description, I couldn't get precisely what features to describe in captioning. I get the robotic description part.
and thanks for replying.
@Rozard What matters most is generally what you are seeing.For example: "trigger word " a woman wearing a collared white shirt and a necklace around her neck, a white room and a television with a couch in the background. Generate first with blip captions then redit like what I discribed up here. You will get the hang of it as you progress :)
@PickleRick_1856 oh okay i got it thank you for the help
@Rozard Anytime. Looking forward for your works :)
please make more colours, white, blue, red etc :)
I'd love to.But even the basic as this one doesn't work as intended.Unfortunately I almost give up on this one. It just not happening :(