Do you have any recommendations for LoRA training of characters?

Tried this model with same configuration as example images, i get distorted faces or hands. I use DrawThings and it needs to import model as ckpt format. It seems it does not work well with this one, most of the checkpoints from the site works just fine.

V1.3
- faces are hit or miss (distinctive look)
- eyes and lashes need some improvement

but i think you're on a good way, just needs some improvement. imo the best realistic illust Model so far

Thanks for good job but 1.3 fp32 version is not good as before version. Like this, https://imgur.com/a/glotrFm of course bad image is newer version. :) Anyway, I appreciate the effort and hope for better development. :D

This is still the best realistic Illustrious model. It’s surprising how much better Illustrious performs compared to Pony, even though both are based on SDXL.

Works great with 8-step Hyper LoRA, CFG from 1 to 2, Eular A SGM Uniform, Hi-res of 1280x1856 at 0.4 denoise. CyberRealistic v6.5 as optional refiner. Face A-detailer may sometimes be needed.

Positive prompt:
(score_9, score_8_up, score_7_up, score_6_up, photorealistic, realistic, film grain, depth of field, masterpiece, best quality, amazing quality, very aesthetic, absurd res, newest, scenery, detailed, detailed background, detailed skin, skin pores, (glossy image, shiny image:1.3), colourful image, dramatic lighting, volumetric lighting, dramatic shadows).

or

(absurd res, newest, scenery, detailed, detailed background, detailed skin, skin pores, (glossy image, shiny image:1.3), colourful image, dramatic lighting, volumetric lighting, dramatic shadows).

Negative prompt:
score_1, score_2, score_3, score_4, score_5, glitched, distorted, blurry, low res, low quality, bad quality, worst quality, low detail, worst detail, cropped, censor, censored, text, signature, watermark, username, artist name, chibi, child, young, short, ugly, bad eyes, closed eyes, deformed, disfigured, bad anatomy, bad hands, monochrome, grayscale, sketch, oldest, simple background,

The sydney sweeney model haha

It makes great bodies and follows the prompts well, but it really struggles with faces, I can't get anything decent out of it

The 1.3 version has seen significant improvements, particularly in terms of realistic and the plausibility of the background. But compared to V1.1, characters generated by V1.3 are seem too be a little too Western-oriented. I used prompts "Tifa \(FF series\)" while the face of character still looks like a westerner.

Still, good job.

V1.3_FP16 looks better than V1.1_FP16 in nearly every single way except for the faces, especially the eyes. I'm not sure what i'm doing wrong but I can't get the eyes to look good on 1.3, even with ADetailer they still look bad. Anyone else having this problem?

V1.1 still looks amazing though, it's by far my favorite Illustrious model

I wonder if a lora for realism would work as well as a full checkpoint

Incredible