image-reading

Files (1)

• SKILL.md

SKILL.md (first lines)

# Read Images

Run `img2txt` from the `gsv` target. Use `man img2txt` for current syntax.

Choose the mode based on the task:

- `caption`: Describe the image. Use `--length short|normal|long`.
- `query`: Answer a specific visual question. Requires `--prompt`.
- `ocr`: Extract text. Optionally use `--prompt` to narrow the extraction.
- `point`: Return normalized coordinates for instances of an object. Requires `--target`.
- `detect`: Return normalized bounding boxes for instances of an object. Requires `--target`.

Pass images directly using a local path, `gsv:/path`, `target:/path`, or
`[target-with-colons]:/path`. Target images stream directly and do not need to
be copied into GSV first.

For `query` and `ocr`, use `--response-format` for JSON, XML, Markdown, or CSV.
Use `--schema` with JSON when the result must have a specific structure.

Use `--reasoning --json` when query reasoning or grounding is useful.
Use `--stream` only for unstructured caption, query, or OCR text; it cannot be
combined with JSON envelopes, reasoning, or structured output.

`point` and `detect` produce JSON by default.

From deathbyknowledge/gsv, MIT.

More from deathbyknowledge/gsv