Tag in your own language, copy out the English a model expects

Three clicks, and the English assembles on the right

The problem this solves is narrow and real. You know the look you want, but you do not know that models call it Tyndall effect, film grain or Fisheye Lens. The tool answers that by making those terms clickable rather than searchable, which is the difference between a thesaurus and a prompt editor.

The loop is three steps: pick a subject, pick a category, then click tags. The English prompt for your chosen model assembles in the right-hand panel as you click, and a Copy button hands it over. The library holds 5,040 bilingual tags across 16 subjects and 212 categories, covering lighting, composition, lenses, camera moves, art styles, materials and wardrobe, and the README stresses that the list was organised by hand rather than scraped into one flat list.

Model coverage is GPT-Image-2, Nano Banana, Midjourney, FLUX, Seedance, Veo and Kling, with compatibility for Stable Diffusion and DALL·E syntax as well. Nearly every tag carries a preview image, and clicking one opens a lightbox with zoom, rotate and download.

The housekeeping features are small and practical rather than clever: duplicate cleanup, a one-tap negative-prompt template, inline lighting and quality snippets, and a random colour swap.

The filter you are viewing is the URL you are looking at

Your selection is written into the address, with a query string that names the subject and attribute you filtered on. That single decision does two useful things: a bookmark reopens the same view, and a shared link drops the next person exactly where you were rather than at the top of a list.

Two display choices exist for long lists, and both persist and sync across tabs. Monochrome tags give you a plain grid. Multicolor paints the grid in blocks of eight drawn from a six-colour cycle, so your eye keeps its place when scrolling several hundred terms. A light, dark or system theme sits alongside them.

The same no-account approach covers state in general: there is no backend, and your selection lives in the URL and in your own browser and never reaches a server run by the project. That claim has one exception, and the README puts it in a section of its own rather than a footnote.

The translate box is the one thing that leaves your browser

Everything stays local except the free-text translation box. What you type there is sent to free public endpoints, using no API key of yours, and the built-in providers are tried in order: Google, Microsoft Edge, Youdao, then MyMemory, until one answers.

Those services rate-limit, and MyMemory refuses anything over 500 bytes. The app falls through the provider list, so a short phrase usually lands somewhere, but a long paragraph can still come back untranslated with nothing to tell you why.

The box is debounced at 1.5 seconds and does something more useful than translating: it suggests curated tags matching what you typed, which means it is a search entry point as much as a translation tool. That is the feature to try first, since a suggested tag is a local lookup while a translation is a round trip to somebody else's server.

Previews come from a CDN, even in the desktop build

The tag data ships inside the app, one JSON file per locale. The preview images do not. They are served from a CDN, so hover previews need a network connection in the browser and in the desktop build alike.

That split defines what the offline experience actually is. On an air-gapped machine you can still install, browse all 5,040 tags by subject and category, read their translations and assemble a prompt; what you lose is seeing what a style looks like before you commit to it.

The desktop builds exist for people who want that without a browser tab, and the formats match what each platform expects: an exe or msi on Windows, Intel and Apple Silicon builds for macOS, and deb, AppImage or rpm on Linux. Running it in a browser needs nothing installed at all, which is the shortest path to trying it.

Self-hosting is one container, because the build is a static export

The published image already contains the finished 18-language site, so there is nothing to build:

docker run -d -p 5666:5666 --name imgprompt ghcr.io/rockbenben/img-prompt:latest

Open http://localhost:5666 and it works on a NAS, a home server or a machine with no internet route out. A docker-compose.yml ships in the repo and is three settings: the same image, container name imgprompt, port 5666 on both sides, and restart unless-stopped.

The Dockerfile explains why the image is so thin. A first stage builds on node:24-alpine and runs yarn install with a frozen lockfile and a raised network timeout, then yarn build. A second stage is nginx:stable-alpine, which copies the nginx.conf and the static output from /app/out and serves it on port 5666. Nothing from the build stage ships except files.

There is a deliberate pin in that first line, with the reasoning in a Chinese comment in the file: the static output is architecture-independent, so the build is fixed to the native build platform to avoid running the same yarn build twice under QEMU, which doubles memory usage and invites out-of-memory failures. Because the result is a plain static export, any static host works as well as Docker, and the deployment guide covers Vercel.

Adding a tag means editing one JSON file per language

The tag library is open data with a shape you can read: src/app/data/prompt/prompt-<locale>.json, one file per language, each holding the localised terms and, for the entries sourced from GPT-Image-2 examples, per-entry author credits.

That is why tag contributions need no toolchain. Correcting a term in your own language, or adding a missing one, is a single file edit. For code changes the stack is Next.js 16 with React 19, TypeScript, Ant Design 6 and Tailwind CSS 4, with Node 20.9.0 as the declared floor, and the project asks for an issue first when a change is large.

The build scripts show how the locale data is handled rather than hand-written into pages. sliceData.js runs before both dev and build, flattenRscSegments.js runs after the build, and buildWithLang.js is a separate per-language build path. next-intl handles the interface, and next-themes the theme switching that syncs across tabs.

MIT covers the code, not the tag library

The licensing is split in a way worth reading before you do anything with the data. The project's own code is MIT. The prompt library is aggregated and curated from other open sources, and the README credits each one with its own terms.

The largest contributor is EvoLinkAI's awesome-gpt-image-2-prompts, holding GPT-Image-2 prompt examples under CC BY 4.0; entries from it were localized into 18 languages with the author credits preserved in the data file, which means attribution obligations travel with the data. The baseline tag taxonomy comes from Physton's sd-webui-prompt-all-in-one under AGPL-3.0, and promptoMANIA supplied structured keyword inspiration.

So the MIT licence on the repository does not make the JSON files freely reusable in isolation. If you fork the tag library and ship it, the CC BY 4.0 attribution and the AGPL-3.0 baseline both apply to the data, and that is a question for whoever owns your compliance process rather than for a licence file.

On maintenance, releases are tagged app-v3.1.0 on 2026-09-27, app-v3.0.0 on 2026-07-09 and app-v2.7.0 on 2026-04-25, with the last push on 2026-09-27.

Editorial conclusion

Adopt img-prompt if you can name the look you want in any language but not in the vocabulary Midjourney and FLUX were trained on, and read the licence note before you redistribute anything. Do not adopt it as an image-to-prompt reverse tool, because it only builds prompts from tags and never looks at a picture, and do not expect the translation box to handle a long paragraph, since one of its four free providers refuses input over 500 bytes. Verify first that the tag list in your language is complete enough to be useful, since each locale is a separate JSON file maintained by hand.

Frequently asked questions

Official sources

1. License: MIT
2. Project website
3. README
4. Releases
5. rockbenben/img-prompt on GitHub