The prompt TII demonstrates
The model card example asks for a long paragraph about a picture: the prompt is User:<image> followed by the instruction, then Falcon: to open the reply. Following the published format is the single most reliable way to get a relevant caption rather than a generic one.
Getting detail rather than a summary
- Ask for what you want described, not just "describe this".
- Request a length, since the model will otherwise pick one for you.
- Ask for specific attributes: count, position, colour, text, material.
- Run it twice with different instructions if a single pass is too coarse.
Resolution is the ceiling
The model supports dynamic high-resolution encoding with tiling up to about 1008 pixels because small objects and text disappear when a large image is downscaled. If your captions miss detail that is clearly visible to you, the input pipeline is the first thing to check.
Related pages
Sources
falcon2.lol is an independent third-party site. It is not affiliated with, endorsed by, sponsored by or operated by the Technology Innovation Institute, and it does not speak for TII. Every number on this page is attributed to the source listed above.