Refined Image Generation using Stable Diffusions Img2Img-Feature

This post is meant to test out the Img2Img feature and to reach our limits using it. Thanks to Albert Bozesan, whose video inspired me to give Img2Img a new try.

So, to start this off, we create a simple and rough sketch having our general idea in mind. This scene is also inspired by PromptHouse.xyz current challenge “Moment in Time”. You can find my post about my entry here.

first input image for img2img pipeline.

I know that this image lacks of artistic skills, but thats not what it’s meant for. We will use a first pass of img2img with the following prompt:

  • a boy and his father standing on a balkony
  • visiting the big solarpunk city for the first time
  • a big monument in the background and mountains in the far distance
  • volumetric, cinematic lighting, studio quality, sharp, elegant, vivid, P.A.Works, art by artgerm and greg rutkowski and alphonse mucha
First pass of Img2Img on our input image.

Now we will go ahead and do this process several times. until we receive a more detailed picture.

Third pass of img2img with our prompt.

Now that we have a solid base, we can begin shaping it to our imagination. Therefore we will only select a specific part of the input image and change the prompt to what we want to see in that specific region.

After collecting a few of these images we will go to our picture editing software and add the images back to their according place. We can use layer-masks to blend in the edges of the images or to hide unwanted features.

Repeating this procedure a dozen times, we can receive a result like this.

final output image

The Final output needed 14 iterations of picture bashing using about 200 generated images and a few hand painted brush strokes.

Here you can see a timelapse of the creation process. You can see how the individual iterations changed the input image until we receive the Final output image.

small gif of timelapse.

Leave a Comment

Your email address will not be published. Required fields are marked *