Out of focus portraits

[bsalberto77]

Hi, I have a problem. When taking pictures of people, especially portraits, it happens to me that using the exact same prompt and the same instructions sometimes my face comes out totally focused and if I do it again at another time or another day they don't come out as clear. I'm going crazy because I've tried different prompts and it's always the same. Can someone help me and know why it could be? Thank you!!

[bbecausereasonss]

If you're using 2.0/2.1 this could be related to high-res fix (issue I was having)

[bsalberto77]

Thanks, yes I'm using 2.1 And what should I do then?

[bbecausereasonss]

I would recommend either:

1. Not using high-res fix
2. If using it, setting your starting image size at 768x768

I haven't found better solutions.

[bsalberto77]

But I never have Highres fixed checked. :-(

[kalkal11]

With the correct settings, SD upscale script in img2img is better, set denoising to around 0.3 and set the tile size to be square, a bit higher than default. LSDR is best upscaler but SwinIR is almost as good and faster. It'll reproduce the image in tiles.

[bbecausereasonss]

He's just talking about general inference with 2.1 portraits not up scaling.

[leakedglass]

@bsalberto77 hard to say without more information. Could be the Prompt or Negative prompt or something with the training, dataset/class images or settings.

I been doing a lot of training on people, particularly myself and I'm not having a focus issue. It might happen just like an other artifact issue but not a common problem.

I was previously having the issue that @bbecausereasonss mentioned with high-res fix while models trained on 2.1 but I changed something and now that was resolved. I still trying to backtrack and figure out what setting in training fixed it. I'm not sure of the rules on here, I don't know if I'm allowed to post examples. If I can I will.

[zero01101]

having the "upscale latent space image when doing hrfix" option enabled while using non-square aspect ratios (seemingly only in portrait?) gave me unusable output 100% of the time on 2.x models, guessing it was the latent space upscale option you're thinking of