Everything else, answered.

It reads a picture and writes the prompt that would recreate it — naming the subject, lighting, composition, palette, medium and mood instead of leaving you to guess at adjectives. Geno goes further than a one-shot description: every attribute comes back as a separate field you can toggle, edit and reuse, and you can export the result as full text, an assembled prompt, or JSON.

A JSON prompt is the same instruction as a written one, with the parts named. Instead of a paragraph where subject, lighting, lens and palette are tangled into prose, each is its own field. That matters the moment you need a second image to match the first, because you can hold every field constant and change exactly one. Geno returns a JSON prompt from any reference image, alongside the plain-text and assembled forms — one analysis, three presentations, and nothing is regenerated when you switch between them.

How JSON prompts work

Three things prose cannot do. You can change one variable and know nothing else moved — same lighting, same lens, different wardrobe. You can diff two prompts and see exactly what changed, which two paragraphs will never tell you. And you can loop a script over a single field to produce a series, instead of hand-writing prompts you hope stay consistent. For one image, prose is fine. For the fifth image in a client campaign, prose is the reason the fifth one does not match.

See a real JSON prompt and the image it produced

It depends on how the model was trained. Models that take structured input read the fields directly; models trained on captions — Midjourney and the Stable Diffusion family among them — do not, so the braces and quotes arrive as tokens carrying no visual information, and you want the flattened text instead. This is exactly why Geno hands you all three forms from a single analysis: paste the JSON where it is parsed, the assembled text where it is not, and never re-run the decode just to change format.

Not on its own, and anyone promising otherwise is selling something. A model does not obey harder because you added brackets — run the same content as JSON and as well-written prose and you land in much the same place. What JSON changes is everything after the first image: it is a record you can version, review, hand to a colleague, and vary one field at a time. Judge it as a format for repeatability, not as a quality trick.

What structure actually buys you

Three different things. /describe gives you a one-shot wall of text. An image to text tool means OCR — it lifts written words off a scan or screenshot, which is a different job entirely and not what Geno does. Geno reads the visual qualities of a picture and returns named attributes you can toggle, reuse and build a library from, plus image generation, editing from reference shots, and every social and ad creative crop in the same workspace. Describe is a caption; Geno is a recipe.

Because turning an image into a prompt is one step out of five, and it is the only step those tools do — they hand you a caption and stop. Geno decodes the image into named, editable attributes, gives you the prompt as text, assembled or JSON, generates from it on nine models, crops the result to every social and ad size, and keeps the library tagged and searchable. Generation runs on paid models and every image is stored for you; that is what the subscription pays for. The free trial exists so you can prove the whole chain works on your own style before paying for any of it.

Every attribute comes back as its own editable field, so a wrong value is a two-second correction rather than a re-run: change it and the prompt updates. The analysis is also adapted to what kind of image it is — photographic, 3D render, illustration and graphic each draw on their own slice of a 96-field taxonomy — so a photograph is never scored on fields that only make sense for a vector poster. A free tool gives you one paragraph and no way into it.

That is what the structure is for. Decode one reference, hold the fields that carry the look — lighting, lens, palette, medium — change the subject, and image twelve still belongs with image one. Every decode is saved to a tagged, searchable library, so the next round of work for the same client starts from the saved prompt instead of from memory.

Prompts are written to be model-agnostic and work anywhere you paste them — Nano Banana, GPT Image, Midjourney, Flux and Stable Diffusion included. Image generation inside Geno runs on nine leading image models, selectable per job, and analysis runs on frontier vision models. No API key and no setup on your end.

Yes. Generation accepts reference images, so you can feed a finished shot back in to restyle it, extend it, or produce a variant that still reads as the same picture. How many references you can attach depends on the model — up to 16 on GPT Image 2, 14 on the Nano Banana models, fewer on some others — and the picker tells you the limit for the one you have selected.

Thirty crop presets: sixteen social placements across Instagram, Facebook, LinkedIn, YouTube, TikTok, X and Pinterest, plus fourteen Google Ads sizes. One generation, every placement, without leaving the workspace.

Neither. Generation and analysis run on Geno’s own model access — sign in with Google and upload an image. The prompt is the output, not the thing you have to be good at.

Any still image format your browser can open — JPEG, PNG, WebP, GIF, HEIC and the rest. There is no file-size limit to worry about, because images are resized in the browser to 2048px on the longest edge before anything is uploaded, so a 50-megapixel camera file is fine. You analyse one image at a time; when you attach reference images for generation, you can select several at once.

Your images and prompts live in your own private, per-user storage. Nobody else — including admins — can read your work.

You keep your rights in the images and prompts you provide, and you may use what you generate commercially, including in client deliverables. Two honest caveats, both in the Terms: your use is also subject to the terms of the underlying AI model providers, and because outputs come from third-party models we cannot warrant that any one of them is unique or non-infringing — similar outputs may be generated for other people.

Read the full terms

Freelance designers, marketers and creators who deliver AI visuals to clients or an audience — people who need the same style, repeatedly, in many formats. Not for one-off experimentation.

Each plan includes a monthly pool of credits that every image generation and analysis draws from. Run out and you can top up anytime; plan credits reset each month, top-ups never expire. Start with free trial credits, no card required.

Enough to decode a handful of references and generate your first images from them. The trial is for proving the decode works on your style — plans are for shipping with it.

Cancelling ends the subscription, not the account. Access continues to the end of the paid period, and your content and account data stay for as long as the account is active — so your library and saved prompts are still there if you come back. You can delete your history yourself at any time from Settings, or ask us to close the account and delete your personal data; some billing records are kept where the law requires.

Cancellation terms

Yes — we pay 10% of every payment, for as long as the person you referred stays subscribed.

See the affiliate programme