Palette Director

Palette Director is a small utility LoRA for Flux.2/Klein workflows driven by a palette card and a structured prompt suffix. It is not a normal style LoRA and it is not meant to recolor finished images by itself.

The intended setup is:

• a palette-card image generated by the Palette Director ComfyUI node

• a prompt suffix that names palette roles such as background, main, human, material, and accent

• ReferenceLatent / conditioning routed into the sampler

The LoRA was trained to read this palette-card + prompt-suffix format and push generation toward the selected palette.

Recommended use:

• Use with the Palette Director custom nodes for ComfyUI — includes Palette Composer and Palette Director

• Example workflow is in the repo: examples/flux2_palette_director_workflow.json

• Start with LoRA strength around 1.

• Increase strength if the palette is too weak.

• Avoid strong color words in the prompt if you want the palette to lead.

Important limits:

• This is guidance, not exact hex color locking.

• The output may not use the selected colors one-to-one.

• If the prompt says things like red jacket or blue room, those words can fight the palette.

• Skin, sky, grass, metal, wood, and other strong object priors may keep their natural colors.

• The palette card can sometimes leak into the image as blocks, borders, or UI-like shapes.

• This model works best when the workflow is built around the palette-card conditioning method.
  
  Note:

  This model is also a learning project for me. This first version was made to test whether palette-card based color guidance can work in a practical workflow.

  My long-term goal is stronger, more predictable color control, but this v1 is not there yet. Treat it as palette guidance, not exact hex color locking.

  I deliberately included some weaker results in the gallery so you can inspect the limits before downloading. There is no guarantee every base model will follow the palette the same way. Tested with flux-2-klein-9b-kv-fp8.

  Over time, I plan to improve it by increasing the dataset size, improving data quality, and adding more varied examples. Because of that, some results may be inconsistent in this version.
  
  Examples:
  
  LoRA off. Node test only:
  
  
  
  LoRA on. Non-color prompt tests:
  
  
  
   
  
  

LoRA on. Color-containing prompt tests: