Describe a picture and watch it come out of the noise

Imagine is image generation inside Carpathian AI chat. Type /imagine and a description, or ask a model for a picture in plain words, and an open source image model running on hardware we own draws it into the conversation. A free account is all it takes, and every account starts with an allowance of images before anything is billed.

Create a Free AccountOpen the Chat
carpathian.ai / chat / imagine

This preview draws in your browser so you can watch the steps. The image models are in /chat.

Inside the chat
/imagine on any model, or ask in plain words
Priced per image
An allowance first, then the published rate
Hardware we own
Open source image models on our own compute
On the API
OpenAI-compatible images endpoint

Two ways to ask for a picture

There is no separate image app to open. A generation is a turn in the conversation, so the picture sits under the message that asked for it, and the thread is there to go back to with everything else in your history.

Use /imagine when you know what you want and want it fast. Ask in plain words when the picture is part of a longer back and forth, and let the model write the prompt from the conversation so far.

  • Image generation inside the same chat as the text models
  • Up to four pictures from one prompt
  • Your choice of size, in multiples of 16
  • Every generation kept with the prompt and settings it ran with
  • A gallery that pages through your whole history
  • An allowance of images on every account before anything is billed
  • A failed generation never counts against the allowance
Create a Free Account
carpathian.ai / chat / asking
  • /imagine a lighthouse at dusk
    The fastest route. It works on every model, because the chat model is not called at all and the prompt goes straight to the image model.
  • /imagine a lighthouse at dusk 768x512
    A size at the end of the prompt picks the shape. Width and height are multiples of 16, and anything over the model's ceiling falls back to 512x512.
  • Draw me four versions of that lighthouse
    On a model that can use tools, plain words are enough. The model sends the request to the image model itself and picks up a size or a count if you mention one.
  • The reply holds a tile per picture
    Each tile is the size you asked for, so the image lands in the box already on screen. A progress bar counts the model's steps while it draws.

What the steps are doing

An image model does not paint from left to right. It starts from a field of random noise the size of the picture and removes some of it at every step, steering what is left toward the prompt. The broad shapes settle first, then the edges, then the fine detail.

That is why the chat shows a step count while a picture draws rather than a guess at the time left. The starting noise comes from a seed, so the same prompt on the same model with the same seed draws the same picture, and a new seed draws a new one.

Drag the slider to walk one picture from noise to finished. The preview on this page draws in your browser to show the idea; the models in the chat do the work on our hardware.

carpathian.ai / imagine / steps
Step9 of 30

What it costs

Every image is billed at its model's per-image rate, published on the AI model pricing page, so a prompt that returns four pictures costs four. An account includes 2 images every 7 days before anything is billed. The allowance is a rolling window, so the next image frees up as your oldest generation ages out, and a generation that fails does not count.

AI Model Pricing

Questions about image generation

Where do I generate images?
In the chat. Type /imagine followed by a description on any model, or ask a model that can use tools for a picture in plain words and it sends the request to the image model itself. The pictures land in the conversation under the reply, with the prompt beside them.
Do I need an account?
Yes, a free one. Text chat works signed out, image generation does not, and the generation endpoint refuses a signed-out request on the server rather than relying on the page to hide the button. Signing up takes an email address.
How much does it cost?
Every account includes an allowance of images before anything is billed. Past that, each image is billed at its model's published per-image rate, so a prompt that returns four pictures is billed as four. The allowance is a rolling window, so the next image frees up when your oldest generation ages out rather than on a calendar date.
Does a failed generation use up my allowance?
No. A generation that fails is not counted. Queued and running generations are counted while they are open, so a burst of requests cannot get ahead of the count while the first is still drawing.
What sizes can I ask for?
The default is 512x512. Put a size at the end of the prompt, for example /imagine a lighthouse at dusk 768x512, to pick another. Width and height are multiples of 16, each model has a ceiling on total pixels, and a size above the ceiling falls back to the default rather than failing.
How many images can one prompt return?
Up to four. Each one counts separately against the allowance and against billing.
Can I call it from my own code?
Yes. Create an AI instance bound to an image model and post to /images/generations on the AI base URL with that key. The request and response follow the OpenAI images format, the images come back as b64_json, and the call returns when the pictures are ready.
Where are my images kept?
In your organization's storage, listed in your gallery. Images are kept for a retention window and removed after it, so download anything you want to keep for good.
What am I allowed to generate?
The Acceptable Use Policy covers it, and section 4 governs the AI services. Adults are treated as adults and the line sits at what is illegal. Adult material may be generated but not hosted on Carpathian cloud, storage, or compute, and anything that depicts or sexualizes a minor is prohibited everywhere, including prompts that ask for it.

The rest of the AI work

The chat Imagine lives in is free for text, with no account and no message cap. The models page covers which open weights we run and why, and the API page covers the keys, limits, and logs that come with building on them.

Free AI ChatOpen Source AI ModelsInference API

Make an account and ask for a picture.

Sign up with an email address, open the chat, and type /imagine. The allowance covers your first pictures before anything is billed.

Create a Free AccountContact Us