People Silhouettes (Generative AI Experiments)
Created //
TL;DR // I used Stable Diffusion (SDXL 1.0) to generate silhouette figure images as assets for use in other artwork.
Background
I was creating some new building facade images. I needed silhouettes of people in the windows and I thought this might be a good use for generative AI.
Instead of drawing many different silhouettes, maybe Stable Diffusion could do it for me? After the images below, I'll detail the prompt I used.
The final silhouette images
Prompt details
These are the prompt details, as generated by Noiselith. Each image above was generated with a different seed, but the same prompt.
{
"type": "IMAGE",
"prompt": "black silhouette of single person, full body, white background, monochrome, black and white",
"negative_prompt": "color, shading, alien, too many limbs, furniture, desk, chair, window, building, ground, shadows on ground, multiple figures, multiple people, animal, frame, vignette",
"width": 1024,
"height": 1024,
"sampler": "DPM++ 2M Karras",
"steps": 20,
"cfg_scale": 7,
"random_generator_source": "NV",
"models": [
{
"name_hint": "SDXL 1.0 (VAE 0.9)",
"type_hint": "SD_CHECKPOINT",
"weight": null
}
],
"initial_image": null,
"is_fast_mode": null
}
Cleanup: Before and after
The images generated directly with Stable Diffusion were not as clean as the images above. Sometimes there were gradient backgrounds or other artifacts. These were all removed manually in Affinity Photo, to turn the images into the clean black and white masks seen above.
The manual cleanup time took... awhile. Especially when doing it for 50 images or so. There has to be a better way.
What's next?
If nothing else, it demonstrated that generative AI can be a quick way to generate libraries of assets for use in larger pieces of work. Think: simple image libraries, collections os well-defined shapes and outlines, collage source material, etc.
However... there has to be a better way to automate the post-processing cleanup job. That'll be something for a later day...
// ka
