search

derrian-distro / LoRA_Easy_Training_Scripts

A UI made in Pyside6 to make training LoRA/LoCon and other LoRA type models in sd-scripts easy
GNU General Public License v3.0
998 stars 101 forks source link

Output results not desired #210

Closed smasonlawrence closed 2 months ago

smasonlawrence commented 2 months ago

Hi All, since there isn't a discussion board here I guess I'll have to put this under issues. Anybody able to offer some insight as to how I can improve my quality? I'm just overall not that impressed with the results. toml and output result attached... any suggestions on better settings to improve results?

vlengies_20240513150310_004500_00

vlengies.toml.txt

Jelosus2 commented 2 months ago

Ok, so I have some questions.

  1. What are you trying to train?
  2. How many images of the concept do you have?

Also is that image generated from the sampling option of the trainer? I see you had it enabled in the toml; if yes then try on the tool you use to generage instead of trainer's sampler.

smasonlawrence commented 2 months ago

I'm trying to train the actress Vanessa Lengies, I have 344 images, all high quality and cropped to her head or body. My captions are similar to this: vlengies, 1girl, solo, smile, brown hair, realistic, brown eyes, looking at viewer, long hair, lips, grin, teeth, upper body, curly hair

I ran it through SD with the attached out put, its okay..

stable-output

but hardly looks like the spitting image of what she actually looks like. photo-reference-of-actress

Jelosus2 commented 2 months ago

I don't train photorealism so I can't help you further, sorry.

smasonlawrence commented 2 months ago

Ok figured it out I think, the sampler in LoRA Easy Training Scripts output looks horrible for some reason, running it through SD with the actual training check point looks 100% better. So the checkpoint of the lora i was using had too many steps, but I couldnt tell that from the sample image that was generated by LoRA ETS. But in SD, I ran it at different strength's and found that it was over baked. So a strength of 0.6 was ideal. Just posting this if anybody else ever has a similar issue, the take away is this:

Final output image at 0.6 strenth is below, looks a lot better. stable-output-0_6

derrian-distro commented 2 months ago

just a note, the sample code is code kohya, the creator of the underlying system this uses, made. Personally, I don't use samples at all, instead preferring to just test in auto1111 instead.