Detailed eyes, I used a lot of realism to put this together with close up shots of the eyes to capture the fine details.
Description
FAQ
Comments (7)
So, of course, this LoRA is usefull only when you want to make a portrait.
But there is few issue with it, mostly if you want to use a character model (LoRA or Textual Inversion).
First, it add eyes makup (a slight one) which is hard to avoid with negative prompt.
Second, the weight of the pictures used change the face of the model you try to use.
For the first issu, i finally was able to deal with it.
But the second one, not at all.
I tried a lot of solutions to avoid that face changing. First, I used separate weight for te and unet. I tried a lot of different setup. None of them gave me good results. I also tried to fine tune the block weight of each block (which took me some times, so to say) and even with that, the LoRA model face I wanted to used was changed at the point you wouldn't say that is the same character.
Have you tried with inpainting (such as via adetailer, which has an eye detection model)?
@ptdurnvcoedqegsdnw No. When I test something, I don't do that with inpaint until it is specify that it is a inpaint model exclusively. If your model is too strong to be used in txt2img, you have to explain that on the description. You also have to tell us in the description the favorite weight we must setup to it, how many CFG and Steps, which checkpoint (to used and with which you trained your LoRA), etc...
If you don't, your LoRA is basically useless and a big waste of time for who'd like to test it (I spent mostly 2 hours trying your LoRA! )
@Alex_Ringess Not mine, btw, I just found this lora as well.
@Alex_Ringess Thank you for the feedback. I am new to training models. I used around 12 images to train this model all close-ups of photos of eyes. Maybe I should have used more photos. Not sure why it changed your character.
@ptdurnvcoedqegsdnw Did you find this Lora useful as the feedback will help me to make maybe better ones in the future? Maybe a Fine-tuned eyes 2. I have not long started with the training models and do want to create more.
@catastrophical0138 a version 2 can't hurt! I did find it useful, I'm only in early stages of testing but using adetailer's eye mesh detector with a denoise of 0.4 or 0.5 and lora weight of 0.5 seems to deliver the results I see in the sample images. I haven't tried a huge variety of eyes yet. But I can't speak to changing style since I applied it strictly with inpainting via adetailer, partly to avoid applying the lora in more areas than just the eyes.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.