Skip to content

Image generation

Add an image provider, and the assistant can make images in Chat and Code. You ask in your own words, the text model writes the prompt, and the image appears in the conversation while you watch it being made. Ask for a change and the model edits that image. In Code and in a chat with a folder it can save an image into the folder.

Image providers are separate from the text providers, and the same page adds video providers and audio providers. A text model calls the image model as a tool, so it works with any text model that can call tools, local ones included. Nothing changes until you add one: without an image provider, Atelier sends models exactly what it sent before.

Open Settings › Admin › Media. Servers running on this computer on their usual port show under Found on this computer, one click from added. Add lists every service.

After you add one, Test checks that it answers and lists its models. Make a test image makes one small real image. It’s a separate button because on a paid service it costs money.

On a server with several accounts, each person can also use an image server on their own computer: open Settings › Your models in the desktop app, and it works from sessions on the server while that computer is online.

Service Notes
ComfyUI Your models appear with no setup, and you can import your own workflows. Shows the picture forming as it works.
Forge, AUTOMATIC1111, SD.Next, reForge Start them with --api. Shows the picture forming.
SwarmUI Live progress works on a server without accounts. With accounts, you see the image when it’s done.
InvokeAI SD 1.x, SD 2.x and SDXL models. It shows that it’s working, but no step counts.
Draw Things Turn on the API Server in the app, with HTTP as the protocol. It uses the model open in the app, and shows no progress. For edits, sides must be multiples of 64.
KoboldCpp Start it with an image model loaded. It makes one image at a time.
stable-diffusion.cpp Use a recent sd-server. Shows live progress, but Stop can’t end a job on the server: it finishes in the background.
LocalAI

OpenAI, Google Gemini, Vertex AI, Black Forest Labs, Stability AI, xAI, Ideogram, Recraft, Mistral, Azure OpenAI and Amazon Bedrock connect with a key. If you’ve added OpenAI or Mistral as a text provider, Use the key from … saves pasting it again.

  • ChatGPT Plan uses the ChatGPT Plan you’ve already signed in to, and your plan’s limits. OpenAI doesn’t document this route, so it may stop working or not be open to every plan.
  • Amazon Bedrock takes a Bedrock API key, or ACCESS_KEY:SECRET with an optional :SESSION_TOKEN. It offers Stability’s models, in us-west-2.
  • Vertex AI uses express mode, and you type in the model ids.
  • Mistral has no image API of its own: a Mistral chat model calls its image tool, which runs Black Forest Labs’ FLUX, so each image also costs that model’s tokens. It takes the shape you ask for, but no size, seed or input image, and makes one image at a time. It hasn’t been tested against Mistral’s live service yet.

OpenRouter, fal.ai, Replicate, Together AI and Hugging Face give you many models with one key. For services that don’t list their models, type the model ids in the form. On Hugging Face an id can name a provider, as in org/model:replicate.

Any other server with /v1/images/generations, such as LiteLLM, works as OpenAI-compatible images.

Ask for it: “make a poster for the summer party, in a retro style”. The text model turns that into a full prompt and picks a size. A frame in the final shape appears straight away and says what’s happening, such as how many jobs are ahead of yours, or “Step 12 of 28”.

When it’s done, click the image to see it bigger and zoom in. Point at it for Copy, Download, Use as reference and Details, which shows the model, size, seed, time and cost. Copy needs https or localhost, so a browser on a plain http address doesn’t offer it. Images in the sidebar has everything you’ve made, and you can search it by prompt.

On local models you can ask for quality: high adds a second pass and a face fix on SDXL, and draft uses fewer steps.

  • Change an image. Say what to change: “make the sky darker”. This works on images you attach, too. The model needs an image model that can edit; if the one in use can’t, it says so.
  • Change just one area. Open an image, choose Change an area (the brush), paint over the part to change, say what it should become and press Generate. Atelier makes a mask the size of the picture and sends it with your words, and the model asks the image model to change only that part. The brush has a size, an eraser, undo and clear, and works with a mouse, a pen or a finger. With the keyboard, arrow keys move the brush, Space paints, and Shift with an arrow paints while it moves. It is on every image in a conversation, and on the Images page, where it sends the change to the conversation the image was made in. It needs an image model that takes a mask: OpenAI’s GPT Image models, Stability AI, Ideogram, Amazon Bedrock’s Stable Diffusion 3.5 Large, ComfyUI (SD 1.5, SDXL and the other checkpoint families it lists as able), Forge and AUTOMATIC1111, and SwarmUI. If the model in use can’t, you are told, and the assistant offers one that can; it never quietly changes the whole picture instead.
  • Save it. In Code or a chat with a folder, ask: “put it in public/hero.png”. That’s a file change like any other: it follows your permissions, shows in a chat’s output files, and Rewind undoes it in Code.
  • Choose the model. Set your default in Settings › General, or pick one for this conversation from the + menu under Image model. You can also just name one when you ask.

Use as reference on the Images page starts a new chat with that image attached.

Artifacts, inline visuals and replies can show an image the assistant made: see Images you made, in an artifact.

Text models that can’t see images can still make them: they get the image’s id instead of the picture, so they can still ask for changes.

Making an image touches nothing but the image provider, so it doesn’t interrupt you. Chat, Auto and Bypass run it without asking, and each turn can make at most 8 images, so text on a web page can’t run up a bill. Ask mode asks first, and Plan mode doesn’t make images. Saving to your folder is a file edit and follows your permissions.

A local image model and a local text model usually can’t sit in the GPU’s memory at once. If your image server shares a GPU with a local text provider, set Shares a GPU with on the image provider:

  • An image job waits its turn in the same queue as your text requests, and nothing else runs on that server meanwhile.
  • The text model is unloaded first, so the image model has the memory.
  • With Unload after each image, the image model leaves too, so the next text request has room.

The text model reloads for the next reply, which takes as long as loading it always does. Turn it off if your GPU can hold both.

Your checkpoints and models work without this: each appears as a model with a recipe that suits it. A model Atelier has no recipe for, such as a video model, is left out, and that’s what workflows are for (a workflow that saves video makes a video model).

Open the ComfyUI provider’s Workflows. Paste a workflow exported with Save (API), or drop a JSON file or a PNG made by ComfyUI, which carries its workflow. Atelier finds the prompt, seed, steps, size and input image, and shows you what it found so you can correct it. The workflow then appears as an image model.

A workflow saved with ComfyUI’s ordinary Save works too, unless it uses subgraphs. For those, use Save (API).

Atelier asks ComfyUI to keep images and input images in its temp folder, not your output folder. That folder empties when ComfyUI restarts.

Settings › Usage shows images and what they cost. A cost a service reports is exact; otherwise it comes from a list of prices and shows as an estimate. You can set your own price per model.

Admins can set limits under Settings › Admin › Media: images per turn, and images or spending per person per day. They can also hide a model from other accounts, which stops anyone else using it, even by name.

A generated image is stored like an attachment, with its conversation. Deleting the conversation deletes it. Images from a temporary chat stay out of the Images page. Exporting your data includes them. See Data and files.

Prompts and input images go to the image provider you chose. Images keep whatever the provider puts in them: ComfyUI writes the workflow into its PNGs, and some cloud services add a content credential.

What you see What to do
Start it with –api Forge and AUTOMATIC1111 only answer requests with that flag.
Missing model The workflow names a file that isn’t on the server.
The GPU ran out of memory Close other models, set Shares a GPU with, or use a smaller size.
The marked area can’t be changed The image model in use takes no mask. Choose another from the + menu under Image model, or ask the assistant to change the whole picture.
The provider refused the prompt That’s the provider’s filter, and it’s final. The assistant won’t reword it to get past it. A blurred or filtered image counts as refused, and isn’t kept.
No previews on a personal image server Update the desktop app.
ComfyUI isn’t starting jobs It’s stuck, usually because it ran out of memory. Restart it.
The first image takes minutes A local server is loading the model. The next ones are much faster while you stay with that model.