Troubleshooting Character LoRA: Why Art Style Collapses Despite Extensive Tuning

justbob9 · reddit · 2026-08-08

The author documents a frustrating journey of training a character LoRA for an anime model. Despite exhausting combinations of learning rates, batch sizes, and network ranks—and even using Claude for captioning—the results remained highly unstable.

Key issues include: 1. Resolution Sensitivity: Horizontal generations look fine, but vertical ones completely break down. 2. Poor Generalization: Good results only happen when exactly replicating dataset prompts; slight prompt changes turn the character into an unrecognizable mess. 3. Inconsistent Art Style: Difficulty in faithfully reproducing the original webtoon's style. The post sparked deep community discussion on dataset distribution and captioning strategies.

Original post →

More from Multimodal

Multimodal channel →