After Training 100+ Krea 2 Character LoRAs: Data Quality Beats Quantity for Likeness

Cold-Worldliness5392 · reddit · 2026-10-08

The author shares hands-on lessons from training 100+ character LoRAs for Krea 2, with the core finding that more training does not mean better likeness.

Key takeaways:

Data quality > quantity: 30 genuinely useful images beat 100 mediocre ones. Common dataset problems: near-duplicates, heavy filters/face editing, low resolution, one dominant angle, group photos, images from a single photoshoot.
Face coverage isn't enough: A LoRA can nail the face but know nothing below the shoulders. Mix tight face shots, waist-up, full body, profiles, different lighting/expressions/clothing — the goal is "understand this person," not "recognize this face."
Too many similar images reduce flexibility: if 70% of the dataset shares the same hairstyle/angle/outfit, the model bakes those into the identity.
The final epoch isn't automatically best: save multiple checkpoints and test each; earlier checkpoints often have better facial likeness, more natural skin, and less overfitting. Checkpoint selection is part of training.
Test outside the dataset's comfort zone: identity must survive new outfits, unusual angles, and lighting.
Body type drifts even when the face is excellent: Krea 2 can preserve facial identity while drifting to a generic body — use physique-check prompts (height, shoulder width, muscularity).
Trigger words matter less than people think; dataset and training quality dominate.

Original post →

After Training 100+ Krea 2 Character LoRAs: Data Quality Beats Quantity for Likeness

More from Multimodal