Text-to-Image
Diffusers
Safetensors
StableDiffusionPipeline
stable-diffusion
Inference Endpoints

My installed Stable Difussion 2 generate bad images

#29
by ykh - opened

My installed Stable Difussion 2 always generate bad images, any friends know how to resolve it ? thank you .
As you can see below image, I just input "a cat" prompt, the result is very ugly.

image.png

Blow is my SD2 configuration settings:
image.png

image.png

  1. Try higher resolution - 768x768 and higher. SD 2.1 model is trained on 768x768 images and it's going to make bad images if you set smaller than that. Try 1024x1024.
  2. Use more words to describe the image. Example:
    A Japanese woman, wide shot, ultrarealistic uhd faces, Kodak Ultramax 800, pexels, 85mm, casual pose, 35mm film roll photo, hard light, detailed skin texture, masterpiece, sharp focus, pretty, perfect, super model, adorable, attractive, Hasselblad, candid street portrait : 6
  3. Do you use negative prompt? That is very important for making realistic images. Here is one example of negative prompt:
    blender, cropped, lowres, poorly drawn face, out of frame, poorly drawn hands, double, blurred, disfigured, deformed, repetitive, black and white : -2

Try CFG-scale 5-6-7

  1. Try higher resolution - 768x768 and higher. SD 2.1 model is trained on 768x768 images and it's going to make bad images if you set smaller than that. Try 1024x1024.
  2. Use more words to describe the image. Example:
    A Japanese woman, wide shot, ultrarealistic uhd faces, Kodak Ultramax 800, pexels, 85mm, casual pose, 35mm film roll photo, hard light, detailed skin texture, masterpiece, sharp focus, pretty, perfect, super model, adorable, attractive, Hasselblad, candid street portrait : 6
  3. Do you use negative prompt? That is very important for making realistic images. Here is one example of negative prompt:
    blender, cropped, lowres, poorly drawn face, out of frame, poorly drawn hands, double, blurred, disfigured, deformed, repetitive, black and white : -2

So it means this is not an issue, it is because the prompts that I input is not enough ?

Try CFG-scale 5-6-7

hmmm. I set the CFG to 5 or less than 5, then the image result is much better. why ?

  1. Try higher resolution - 768x768 and higher. SD 2.1 model is trained on 768x768 images and it's going to make bad images if you set smaller than that. Try 1024x1024.
  2. Use more words to describe the image. Example:
    A Japanese woman, wide shot, ultrarealistic uhd faces, Kodak Ultramax 800, pexels, 85mm, casual pose, 35mm film roll photo, hard light, detailed skin texture, masterpiece, sharp focus, pretty, perfect, super model, adorable, attractive, Hasselblad, candid street portrait : 6
  3. Do you use negative prompt? That is very important for making realistic images. Here is one example of negative prompt:
    blender, cropped, lowres, poorly drawn face, out of frame, poorly drawn hands, double, blurred, disfigured, deformed, repetitive, black and white : -2

It seems not work. I tried your prompts and negative, the result still seems bad, too much noise...

image.png

Same exact issue on my end. Everything I generate looks like a fractal DMT trip. I use an free online generator with the same settings including seed and get something you would expect.

same problem for me. is it a graphics card hardware issue?

I heard SD2.X does not good to use. most of us prefer to 1.5 still

were you guys able to find a solution?
I'm trying dreamlike-photoreal-2.0. I get bad results when using the model via the hugging face ecosystem, but I get amazing results from their UI:
https://dreamlike.art/create

I use the same prompts, same negative prompts and same params.

The solution (at least for me) was to use GPU...

Try to restart your PC (worked for me on SD 1.smth) I had a similarly ugly output and restarted, so the memory got deleted and SD works fine now

Sign up or log in to comment