This LoRA can generate unique iris patterns. 3D artists should be able to use this to texture eyes.
Eye colors you can prompt for: Blue, Gray, Hazel, Brown, Green. You might have to edit the colors afterwards if you want something more intense, as this model is based on natural human photos.
For the flux version these settings seem to work well for me:
1024px x 1024px size upscaled with hiresfix/latent upscale at 1.5x
CFG 2
20 steps
euler / simple
prompt: airis, green iris eye texture, a perfectly round black pupil centered in the texture of an iris with a circular white background, high quality render of a flat eye material, sharp focus
For the SD 1.5 version, use the default Stable Diffusion 1.5 model (finetuned models have weird effects on the iris texture)
Trained on a dataset of 512x512 and 768x768 images, although it seems to prefer generating at 512x512 and using hires fix. (Latent upscale at 0,6 denoise works well, although you can also use ESRGAN at 0,25. Different hires methods will change the patterns depending on what you're after.)
Generating at 768px will sometimes give results which aren't centered.
Trained on base v1.5 and tested with Euler A sampling method. Your results may vary if you use something else, but feel free to experiment.
Description
FAQ
Comments (7)
Can I ask how you got it so circular?
I suspect it has to do with my dataset, as I cropped all images to the exact same circular shape. Still, it doesn't always conform to this shape, about 30% of generations might be square/full frame
@aiceman Thanks, I'm just trying to get a perfect sphere shape in my data, but can't get it. I'll give a white background a try, :thumbsup:
this is great. as a 3D artist I know this will be really useful. Nicely done!
Nice iris textures!
Is it possible to use this Lora to improve existing pictures of eyes, by adding details into iris?
I tried it with Dezgo image to image but without success.
This model was specifically trained for creating flat iris textures from the front, to be used by texturing artists. I don't think it's going to be very useful for anything else, as it has been deliberately overtrained.
maybe when using facedetailer, or a mask detailer of the eyes





