Klein Reference Evaluation Recipes

Reference-image LoRA evals test whether an adapter can pull a neutral source image into the learned style while preserving a fixed composition. They are recipes for comparing adapters, not model-training recipes.

Machine-readable eval recipes live in eval-recipes/.

Times Square Kiss

Use klein-times-square-kiss when comparing Klein style LoRAs against the same source photograph.

Fixed settings:

mere.run image generate \
  --model image-klein-9b \
  --ref-image ./street-kiss.png \
  --strength 0.55 \
  --width 1024 \
  --height 768 \
  --steps 16 \
  --seed 424243 \
  --lora ./style.safetensors \
  --lora-scale 1.5 \
  --prompt "$PROMPT" \
  --output ./runs/kiss-eval/style-times-square-kiss.png

Prompt shape:

STYLE_TRIGGER a sailor in uniform kissing a nurse in a crowded city street,
full body embrace, her right arm around his shoulder, her left arm hidden behind
his back, his arms around her waist, natural human anatomy, no extra limbs, no
extra hands, crisp faces, detailed fabric texture, clear depth separation, high
detail, clean sharp cinematic film still, STYLE_DIRECTION

Guidance:

• Use distilled image-klein-9b for generation.
• Keep --strength 0.55 and --lora-scale 1.5 as the first comparison pass.
• Use seed 424243 for the cleaner anatomy baseline.
• If arms duplicate, keep the same LoRA and try the side-view repair variant from eval-recipes/klein-times-square-kiss.json before retraining.
• Do not make the prompt so style-specific that the base model can solve it without the adapter.
• Store private source images, LoRAs, and generated eval outputs outside the public core repo. A project-local path such as data/lora-evals/times-square-kiss/ is appropriate for artifacts.