How to turn an image back into a prompt

The fastest way to learn what a good prompt looks like is to read one written about a picture you already like. Describe does that: hand it an image and a vision model writes the prompt, which you can then keep as a reusable tile, edit, and fire straight back into the generator. It also runs on batches, up to eighty images in a single request, which is what makes it useful for captioning a whole shoot or a training set.

What Describe is for

Describe lives on /image, next to Generate. You point it at one image or a folder of them, pick a vision model, pick a description style, and it writes a caption per image. The captions are kept — they become prompt tiles you can edit, reorder and reuse — and there is a "generate all" action that fires the whole set back through the image models as prompts. That round trip is the actual feature: picture in, prompt out, new pictures back.

Three jobs it does well. Working out why a picture you like works, by reading a machine's description of it rather than guessing. Captioning a training set before a LoRA run, where every image needs a consistent description. And building a prompt library out of reference material you already have, instead of writing fifty prompts from a blank page.

Which vision model, and what it costs

The three describe models
ModelRuns onPer imageWallet
Aeon 32BOur own GPU0.3 credits⚡ green or ✦ gold
GPT-4o VisionOpenAI0.3 credits✦ gold only
Claude SonnetA subscription we ownNo per-image feeSpends a daily call allowance

Aeon is the one to reach for on a free account, because it is the only row green credits can pay for. GPT-4o Vision costs the same 0.3 but bills a provider per call, so it comes out of gold — a free account has none and will get a refusal rather than a surprise charge.

Which plan: Claude has no per-image fee, and it is not therefore free. Every Claude call that describes an image or enhances a prompt counts against one fixed daily ceiling on the subscription we own: 5 a day on Free, 20 on Starter, 50 on Creator, 150 on Pro, 500 on Ultra. Buying credits does not raise it — it is the one limit here that money cannot move. Claude in chat is metered separately, out of that section's own daily reply pool.

That ceiling is worth planning around rather than discovering. Five calls a day on a free account is enough to describe a handful of images and nothing more, and if you have already spent them in chat there is nothing left for Describe. For a batch of any size, Aeon at 0.3 credits an image is both cheaper and the only one that scales.

Describing a batch

  1. Open Describe. Go to /image and switch to the Describe tool. It is the second entry in the tool rail, next to Generate.
  2. Add the images. Drop in one picture or a whole folder. A single request accepts up to eighty files, and they are captioned in parallel rather than one after another.
  3. Pick the model and the style. Aeon on the green lane, GPT-4o Vision on gold, Claude against the daily allowance. The style presets change what the caption is for — a literal inventory of the frame reads very differently from a prompt written to be re-generated.
  4. Run it and read one caption properly. Before captioning eighty images, do one and check the style is the one you want. Changing your mind after a full batch costs the whole batch again.
  5. Edit the tiles, then generate from them. Captions land as editable prompt tiles. Trim the parts that describe the specific photo rather than the idea, then use generate-all to run the whole set through your chosen image models.

Getting captions you can actually reuse

The description style changes everything

The style preset is not decoration. A description written as an inventory — subject, objects, colours, position — is what a training set wants, because consistency between captions matters more than any single caption being good. A description written as a prompt is a different document: it names framing, lens feel, light direction and mood, and it is meant to be pasted back into a generator and re-rolled.

Picking the wrong one is the most common way to waste a batch. Eighty captions in the wrong register are eighty captions you will rewrite by hand, and the credits are already spent. Run one image, read the caption properly, and only then commit the set.

The style also decides how far a caption drifts when you generate from it. A tight, literal description reproduces something close to the original composition. A loose, interpretive one gives you variations on a theme, which is usually what you want when the point is to build a prompt library rather than to copy a picture.

The limits, plainly

Eighty images is the per-request cap, so a thousand-image set is thirteen requests rather than one. Every caption on Aeon or GPT-4o Vision is charged at 0.3 credits whether or not you keep it, because the model ran either way. And Aeon shares our single GPU with every generation job on the platform, so a large batch during a busy period will wait its turn — the queue is real, and it is the same queue the image models sit in.

There is no free tier of this. 0.3 credits is small, but eighty images is ⚡24, which is more than a free account's entire daily refill. Plan a large batch against the wallet you actually have rather than against the per-image number.

How many images can I describe at once?

Eighty per request, captioned in parallel. Larger sets are split into several requests, and each is charged per image at the model's rate.

Is Claude the free option?

No. It has no per-image credit fee, but every Claude call across chat, enhance and describe counts against a fixed daily ceiling on our own subscription — 5 a day on Free — and credits cannot raise it. For anything beyond a few images, Aeon is the cheaper and more scalable choice.

Can I use Describe on a free account?

Yes, on Aeon, which is payable in green credits at 0.3 an image. GPT-4o Vision needs gold, which a free account does not have.

Will the caption reproduce the picture if I generate from it?

It will produce something in the same territory, not a copy. A description is a lossy encoding of an image — useful for learning what makes a picture work and for seeding variations, not for reconstruction.

Do the descriptions stay attached to my images?

Yes. They are saved alongside the image and appear with it in the library, so a set captioned once stays captioned.