I got sick of regenerating AI ads over one typo, so I made this

For years, I endured the monotony of AI-generated social media adverts, each one presented as a single image. When a typo or incorrect price surfaced, regeneration was the only recourse, hoping the good parts would remain intact. Frustrated by this cycle, I devised a solution.

Enter my creation, accessible via GitHub. This tool, compatible with Claude, ChatGPT, Codex, Cursor, and Gemini, operates differently from standard AI models. Instead of producing an image directly, it generates a layout file using Python. The script meticulously renders each element onto its own layer, scrutinizing for overlapping text, poor contrast, and elements extending beyond the canvas.

A Node script, sourced from ag-psd, then crafts a PSD file complete with editable text layers, masks, shadows, and organized folders. Photoshop or Photopea can open and edit this PSD.

The challenging aspect lay in the transformation of existing flat images into individual layers. Once merged into a JPG, undoing this process proved impossible. Thus, the AI retains each element as a distinct layer, eliminates extraneous objects using generative fill, and regenerates these remnants as cutouts. Before delivery, the model compiles a detailed list of every cutout on a grey background, ensuring accuracy.

To demonstrate its effectiveness, I've included before/after examples from real client projects in the repository, complete with the PSD files. This allows users to inspect the layers themselves. While I'm not an AI expert, I've encountered this issue often and hope this solution can alleviate similar struggles. Feedback and criticism are most welcome.