Krea Recipe Builder

Answer three questions and download a ready-to-use recipe for the KG Krea 2 Image Guide Card V10. Every number is filled in from the render-validated tuning tables — no knowledge of treatments, pulls, or layer arrays required.

1Name it

The name becomes the choice in the card's Use image for dropdown.

2What should the reference image give your picture?

Pick the one job this recipe is for. The right preparation, safety policy, and tuning follow automatically.

Names WHAT the encoder studies from the image. Naming what to skip works too ("…not the person or the background"). Strongest on object-bound aspects like clothing; broad moods like weather mostly ride the image itself.

3How strongly should it come through?

This sets the recipe's response band. The card's strength slider still rides on top — these choices decide what that slider feels like.

4Fine-tune (optional — the defaults are the tested values)

Open fine-tuning

1.0×
1.0×

These scale the recipe's 12-band layer table exactly like the card's manual Structure layers pull / Finish layers pull dials. At 1.0× the tested family table is used as-is (and omitted from the file).

Hard limit the card can't exceed, whatever the slider says. Leave blank for the tested value.

Your recipe file


      

        
        
      

    

    

      
Install it

      
    
        
  1. Save the file into the node pack's custom_recipes/ folder
            (or <ComfyUI user dir>/krea_reference/recipes/).
    2. 
        
  2. In ComfyUI, refresh node definitions (or restart).
    3. 
        
  3. Your recipe now appears in Use image for on the V10 guide card.
    4. 
      

      

        Working examples to compare against ship in
        custom_recipes/starter-pack.yaml: borrow the
        weather, borrow the clothing style (a focus recipe),
        and cinematic color grade.
      

    

    

      Two-minute check — fix a seed and render at card strength
      0.4, 0.65, and 0.9. You should get
      quiet → clear → strong. If the reference's subject sneaks in,
      untick "may appear" above; if nothing changes, pick a bolder loudness.
      The stack's prepared_references output shows exactly what
      the encoder studied.