How many images do you need?

Start with the concept, not a target number

There is no image count that guarantees a useful LoRA across every trainer. A small focused concept and a broad visual style need different evidence. Follow the selected trainer’s actual requirements once available, and use the local workspace to inspect what each image contributes.

Add variation that answers a question

For a character, ask whether different viewpoints and expressions still show the same identity. For a style, vary the subject while keeping the visual language coherent. For a product, show the parts and materials that distinguish the object from its surroundings. More nearly identical images do not answer those questions.

Remove avoidable noise

Remove corrupt files, unrelated images and repeated downloads. Be cautious with screenshots containing interface chrome or visible labels you do not want learned. Consistent framing can be useful, but repeated backgrounds may unintentionally become part of the concept. Use files you are authorized to train on.

Know what the workspace checks

The image preparation tool accepts JPG, PNG and WebP, up to 30 images, 15 MB each and 100 MB total. It reads actual dimensions and rejects files that cannot be decoded. These are local browser limits. It does not score sharpness, verify rights, detect visual near-duplicates or guarantee provider acceptance.

Keep a dataset revision record

Save the original files separately from the JSON draft. Record which images you added or removed and why. When comparing a future training run, change one major variable at a time. A clear revision record is often more useful than expanding the collection without a reason.

Put the note into practice

Read next