kiri-art/docker-diffusers-api

Blurred & noisy images when used with stable-diffusion-2 and stable-diffusion-2-1

Opened this issue · 8 comments

jan-thoma commented

``When i try to build the container from the Dockerfile. Passing either stabilityai/stable-diffusion-2-1 or `stabilityai/stable-diffusion-2-1. Im getting very noisy and blurry images with the parameters adapted from the example:

{
  "modelInputs": {
    "prompt": "Super dog",
    "num_inference_steps": 10,
    "guidance_scale": 7.5,
    "width": 1024,
    "height": 1024,
    "seed": 3239022079
  },
  "callInputs": {
    "MODEL_ID": "stabilityai/stable-diffusion-2-1",
    "PIPELINE": "StableDiffusionPipeline",
    "SCHEDULER": "LMSDiscreteScheduler",
    "safety_checker": true
  }
}

The output looks like this:

cbimage

Am i missing something important here or this a bug?

gadicc commented
gadicc commented

You may also get better results at the native model resolution of 768x768, but definitely the issue here is the small number of inference steps.

Also, FWIW, I personally preferred the outputs from "super dog" from the original 1.4 model, but it may just be a matter of taste 😅

jan-thoma commented

even with 50 steps, i get about the same output

{
  "modelInputs": {
    "prompt": "super dog",
    "num_inference_steps": 50,
    "guidance_scale": 7.5,
    "width": 768,
    "height": 768,
    "seed": 3239022079
  },
  "callInputs": {
    "MODEL_ID": "stabilityai/stable-diffusion-2-1",
    "PIPELINE": "StableDiffusionPipeline",
    "SCHEDULER": "LMSDiscreteScheduler",
    "safety_checker": true
  }
}

cbimage (1)

jan-thoma commented

Here are results from the other schedulers, with the same parameters as above

DDIMScheduler
DDIM

DPMSolverMultistepScheduler
DPMSolverMultistepScheduler

EulerAncestralDiscreteScheduler
EulerAncestralDiscreteScheduler

IPNDMScheduler
IPNDMScheduler

PNDMScheduler
PNDMScheduler

gadicc commented
gadicc commented

And thanks for sharing all those pics which made the issue much clearer! 🙏🙏

jan-thoma commented

Thanks for the help and the great work with this repo. Switching to dev did the trick. Wired up this container with GPT3 and Twitter to generate fictional metal albums. Works great so far (https://twitter.com/bare_metal_head)

gadicc commented

Awesome, thanks for reporting back and for your kind words 🙏

Very cool app! Thanks for sharing on the forums too, where I've linked back to it. Be in touch with any more issues, otherwise, Happy Diffusing! 😁