LunarXuan Releases image-prompt-reverse Skill for Codex to Reverse-Engineer AI Image Prompts

LunarXuan released image-prompt-reverse, an open-source Codex skill that analyzes a reference image and generates high-fidelity AI image prompts. The skill breaks down composition, camera language, lighting, color, and materials, selects three to five key visual anchors, and outputs a 450–700 character Chinese prompt with an English version plus 10–15 English negative prompts. It supports photography, illustration, 3D, posters, logos, and chibi characters under a GPL-3.0 license.
Event version 1.
Key insights
Prompt generation from reference images
The skill analyzes uploaded reference images across composition, camera language, lighting, color, materials, background, spatial layers, mood, and post-processing style. It generates prompts intended for direct use in AI image generation tools.
Visual anchors and anti-fabrication
The skill prioritizes three to five visual anchors that most affect similarity and avoids fabricating details that are unclear in the image. It also excludes easily confused styles based on the target medium.
Default output format
Default output consists of a positive prompt of 450–700 Chinese characters in continuous natural language with an English equivalent, plus 10–15 English negative terms separated by commas.
Supported image categories
Supported types include photography such as portraits, products, documentary, food, animals, nature, and cityscapes; illustration styles including flat, anime, healing, hand-drawn, Chinese style, and cyberpunk; 3D styles; and typography, logos, posters, graphic design, IP characters, chibi characters, blind boxes, and multi-subject mixed scenes.
Installation and licensing
The repository is installed by copying its directory into the Codex personal skills folder at %USERPROFILE%\.codex\skills\image-prompt-reverse or by cloning it directly. The project is released under GPL-3.0 and has about 1,000 stars and 99 forks.