Klein LoRA Recipes

The canonical Klein flow has two phases:

1. Train the LoRA on image-klein-base-9b.
2. Apply the LoRA on distilled image-klein-9b for practical generation.

The recipes are machine-readable JSON files in recipes/.

Reference-image eval protocols live separately in eval-recipes/. Use klein-times-square-kiss when you need a fixed composition test for comparing style pull across multiple Klein LoRAs.

Style LoRA

Use klein-style-lora for broad visual style.

Recommended use:

• image sets where the whole visual language matters
• film/frame/card/comic looks
• lighting, color, composition, texture, and medium transfer

Captioning:

• each caption starts with the trigger token
• each caption describes only visible content
• do not name the style in the caption
• do not mention that the image belongs to a dataset

The style recipe delegates to mere.run image train-lora --recipe klein-fast-style, which currently selects the fal-klein-fast LoRA target preset, the Klein base model, low-memory CUDA settings, and 250-step checkpoints. The plugin recipe adds explicit sample generation against image-klein-9b so remote runs produce preview artifacts without relying on implicit local defaults.

Character LoRA

Use klein-character-lora for identity portability.

Recommended use:

• a recurring person or character
• a mascot
• a fictional subject that should survive scene changes

Captioning:

• each caption starts with the trigger token
• include a class noun such as man, woman, robot, or character
• repeat stable identity anchors in every caption
• describe variables like pose, clothing, expression, crop, background, and medium
• never write same person, same character, previous image, or another view

Character LoRAs should be evaluated by changing the scene, pose, clothing, medium, and camera distance. If the identity only works in the training outfit or training background, the adapter is overbound.

The character recipe also uses klein-fast-style as its base and only overrides identity-focused capacity and caption dropout values. Keep additional experimental training flags in local custom recipe files until they are part of the core mere.run preset surface.