What this workflow does

👉 Generates a sticker/icon with transparency, relying only on the sdxl-turbo model and some computer vision utilities.

The potentially interesting part of this workflow is the creation of the initial canvas, which is painted with a color that is complementary to the generation's subject color profile.

How to use this workflow

👉 Write the subject to generate in the "POSITIVE PROMPT" node, located inside the "INPUT" group.

Specifying the color of the subject should improve results.

The "Configs" group contains:

• The model and the controlnet model loader nodes, which should load the sdxl-turbo and the xl-canny models.
• The text at the start of the prompt (which is prepended to the "POSITIVE PROMPT").  
• The seeds for each section of the workflow.
• Image adjustments for the last output.

All the icon variants are sent to preview nodes, don't forget to save them manually!

Tips about this workflow

👉 Consider replacing or adding nodes in sections 3 to 5 of the workflow; using segmentation** models and sdxl/upscalers to improve the quality of the result.

• The first generation is set with an ancestral sampler; you might want to change this for reproducibility.
• I left notes explaining different sections of the workflow, but the verbosity might be offputting.

** even without relying on segmentation models, the segmentation can be improved. Before the submission of this workflow, I used the BuildColorRangeHSVAdvanced; however, the warning message triggered by the node and the use of undocumented code would be somewhat jarring. The used BuildColorRangeHSV will build the interval relative to the samples without taking into consideration the nature of the background, that is: if the background is colored, it would be preferable to have more leeway with respect to values and less leeway with respect to hue; and if the background is black or white would be preferable to have less leeway with respect to values. With this in mind, I tried to adjust the leeway to an acceptable value via experimentation, using Kmeans to reduce color variation prior to segmentation in order to use a tighter range.