When working with Flux, prioritising quality over quantity is crucial. A small set of 8 high-quality, well-curated images often yields better results than 100+ poorly selected ones.
For captioning, strike a balance between:
- How you typically prompt (to align with your usual generation style).
- Excluding elements you don’t want the model to learn (e.g., artefacts, unwanted details).
It may seem counterintuitive, but less is sometimes more—if uncertain, leaving images uncaptioned and relying on trigger words can be a safer approach.