← sources

End-to-End workflow for training high-quality LoRAs?

source: reddit (reddit:StableDiffusion)published 2026-10-28 14:21ingested 2026-10-29 02:40

retained text

I'm exploring LoRA training (primarily for realistic human characters) and would love to learn what workflows people are using successfully.

Curious about your end-to-end process:

\- Dataset: How do you generate/select images? How many, and what kind of variety?

\- Preparation: Captioning, cropping, resolution, and dataset cleanup?

\- Training: Tools, settings, epochs/steps, and base models?

\- Validation: How do you evaluate results and decide when to retrain?

I'm particularly interested in maintaining character identity and consistency across different poses, expressions, lighting, and environments.

Would appreciate any lessons learned, common mistakes, or workflows you'd recommend!

document fetched 2026-10-29 00:33 · canonical https://reddit.com/r/StableDiffusion/comments/1xhpf4t/endtoend_workflow_for_training_highquality_loras