solution to ideogram 4 prompting

[HAMSAILY]

https://www.reddit.com/r/StableDiffusion/comments/1tzq04h/my_solution_to_ideogram_4_prompting/

github

https://github.com/iChristGit/comfyui-llamacpp-ideogram

workflow

https://www.reddit.com/r/StableDiffusion/comments/1tysann/workflow_ideogram4_with_lora_support_fixes/

[eruanno123]

Here I'm dropping some ideas for integrating Ideogram 4.0 support:

• Add Region support

  • Extend the existing Region system to support Ideogram's compositional deconstruction format.
  • Regions are already defined by layer opacity, so bounding boxes can be generated directly from non-transparent pixels.

• Add color palette extraction

  • Extract dominant colors from region pixels using algorithms such as Median Cut or K-Means.
  • Export these palettes into the generated Ideogram JSON.

About JSON generation, I see three possible approaches:

• Extend text prompt widget syntax

  • Add new prompt tags similar to the existing <lora:...> and <layer:...> syntax.

  • Examples:

    • <camera:tilted>
    • <lighting:softbox>

• New GUI for prompting

• LLM-based generation

  • Convert prompts into Ideogram JSON using an LLM (another dependency)

Compositional deconstruction supports two element types: obj and text

For text elements:

• Use vector layers as text sources?

  • Investigate support for vector layers in the KritaDocument layer API.
  • Convert text found within vector layer into Ideogram text element prompt.
  • Style provided by the prompt text box