# Output spec

## Files
- `ad-4x5.png` 1080 x 1350 (default). On request `ad-1x1.png` 1080 x 1080 and `ad-9x16.png` 1080 x 1920 (each from its own art).
- Art is generated at 1088 x 1360, 1088 x 1088 and 1088 x 1936 and cover-cropped down by the script, so nothing is upscaled.
- PNG from the script. Convert to JPEG or WebP for upload if the platform asks.

## Fonts
Default text: Georgia (headline) and Helvetica (body) on Mac and Windows; on Linux fontconfig falls back to DejaVu, so the look and line breaks differ. The script prints the font it asked for. For exact brand type pass `--font-file <your .ttf>` (bold cut) and `--body-font-file`. Check the output, not the command.

## Checks before you return it
1. Photoreal: believable light, materials, contact shadows; no melted geometry or smeared glass.
2. Fidelity: the product matches the input photo (shape, colour, label wording and position, cap).
3. Check every format you made, not only the first. Exact text: headline, sub-line and button read letter for letter as approved. Read the final image, do not trust the command line.
4. The text sits in the empty area, clear of the product, readable at 320 px wide.
5. 9:16: nothing important in the top 14 % or bottom 20 %.
6. No extra logos, watermarks, invented claims or prices.

## If a check fails
Max 3 attempts. Change one thing per attempt and say what: the label wording in `--product`, the scene, or the model. View each result beside the input. Then return the best one and say what still differs (for example "label letter-spacing is slightly wider").