Image Prompt Reverse reads a reference image and writes a prompt to recreate it, and has 448 GitHub stars in its first 26 days

Matching a reference image with an AI image tool usually means staring at the picture and guessing what words describe it. Image Prompt Reverse is a new open-source Codex skill that reads the picture once and returns a prompt an image generator can use to reproduce the look, and its author says the same skill runs against photography, illustration, 3D renders, logos, IP characters and abstract graphics.

The skill lives at LunarXuan/image-prompt-reverse on GitHub, published under the GPL-3.0 licence. The repository was created on 2026-09-01 and has 448 stars and 34 forks in its first 26 days, with the most recent push on 2026-09-06.

What the skill returns

The README says the default output has two parts. A positive prompt of 450 to 700 Chinese characters in a single natural-language paragraph, followed by an English version of the same prompt. Then a negative prompt of 10 to 15 English terms separated by commas, listing the things a user does not want in the output. Both are written to be pasted straight into a text-to-image tool.

The author says the skill picks three to five visual anchors that most affect similarity to the source image before it writes the prompt, so the output leads with what matters most to the match. The README also states that the skill will not invent details it cannot see, and that any text, logos or instructions inside the source image are treated as visual content.

Which images it is written for

The supported list on the README covers many categories. For photography it names portraits, product shots, documentary images, food, animals, nature and cityscapes. For illustration it names flat, anime, therapeutic, hand-drawn, Chinese-inspired and cyberpunk. For 3D it names realistic, semi-realistic, cartoon, clay, designer-toy and product-render styles. It also covers typography, logos, posters, graphic design, abstract graphics, IP-inspired characters, chibi characters and blind-box figures.

That range is why this is a workflow tool with a broad audience. A marketer trying to match the look of a competitor's product photograph, a designer working from a mood board, a founder building a brand asset from a reference sketch, all get to the same first draft prompt from the same skill without knowing image-generation vocabulary.

The tool runs as a Codex skill package. Anyone already using Codex adds the skill once, invokes it with a slash command and passes it a reference image. The skill runs inside Codex, using the same account and API key the user has already set up.

Source

• Repository and README: LunarXuan/image-prompt-reverse on GitHub