Lora Overfitting Causes Prompt Conflicts in Krea
Review lora performance and disable overfitted ones to avoid prompt conflicts.
Test each lora on a variety of prompts and disable those that overfit.
Summary
The post explains that many popular loras, especially character loras and realism/spicy/de‑censor models, are over‑fitted, leading to prompt bleeding and keyword mixing. When a lora is over‑fitted, Krea may ignore parts of the prompt, causing the model to produce images that blend unrelated elements. Users often discover that turning off the lora resolves generation issues, as the base model regains its intended behaviour. The article notes that over‑fitted loras lobotomise the base model when used beyond their training scope, resulting in inconsistent results across prompts. It advises testing lora performance before deployment and disabling those that over‑fit. The author refrains from naming specific loras to avoid targeting creators.
Overall, the post highlights the importance of evaluating lora fit and suggests disabling over‑fitted ones to maintain prompt fidelity.
Key changes
- Overfitted loras can cause prompt bleeding and keyword mixing.
- Krea may ignore parts of prompt due to overfitting.
- Popular loras often overfit, showing great results only on training images.
- Turning off lora often resolves generation issues.
- Character loras and realism/spicy/de‑censor models are worst offenders.
- Loras lobotomize base model when used beyond training scope.
- Users should test lora performance before deployment.
- Overfitting leads to inconsistent results across prompts.