packfix

packfix is a tool for verifying that AI-generated product imagery preserves the real packaging and for repairing images when it does not. It addresses the problem that image models can convincingly reproduce pose, lighting, and shadow but may subtly alter small print such as weight badges, certification roundels, or fine type along a seam. The tool replaces subjective visual inspection with a quantitative score.

The system has two separate components. The first, fidelity.py, measures whether the printed artwork in a generated frame corresponds to the real pack. It uses SIFT features on the real pack plate, matches them against the generated frame, solves a homography with RANSAC, and takes the inlier count as the score. A score above a gate indicates genuine correspondence; below it, the model invented print and the frame should be regenerated. The second component, packfix.py, repairs frames where the scene is correct but the print drifted. It warps the real pack into the quad found by the homography and transfers the generated pack's low-frequency shading onto the real artwork, so the true print sits in the scene's light. The alpha is eroded and feathered to preserve the generated pouch's edge and contact shadow.

The repository also includes export.py for cover-fitting to exact pixel sizes under a byte ceiling, qa.py for running a deliverable checklist as code, and cost.py for pricing runs from actual token counts. The plates are user-supplied; no packaging artwork is included. The project is MIT licensed.