LoRA Training

Train a character LoRA

Fine-tune a LoRA on a specific character (original IP, illustrated, or photoreal) so it can be reproduced consistently across poses and scenes.

Estimated $2–$9 per training run · 15–45 min

What inputs you need

  • 15–40 character reference images (varied poses/angles)
  • Character-focused captions
  • Training config

What Badgr returns

  • Trained character LoRA
  • Training loss curve
  • Sample generations across poses

Recommended GPU routes

Estimated cost: $2–$9 per training run · Estimated runtime: 15–45 min

Example command

badgr run "python train_lora.py --config character_lora.yaml" \
  --gpu A100_40GB \
  --max-cost 9 \
  --max-runtime 60

Common setup failures Badgr avoids

Character identity drift across poses at inference time — training config enforces a minimum pose-variety threshold in the input set before starting.

Sample generations look like the reference pose instead of generalizing — the training set is checked for angle/expression variety before the run starts.

Ready to run this?

Launch from the dashboard, CLI, or Compute API. Max-cost protection included.