Docs menu
On this page

Image Generation

The Image Generation node generates an image from a text prompt, and optionally a reference image. Use it for thumbnails, scenes, product shots, or any still image your workflow needs.

Inputs and outputs

Inputs: Image ref (one or several reference images; everything you wire in is used, unless you let an Agent choose, see below), Prompt, and a generic Flow input.

Output: an Image handle. Most runs produce one image, but if you pick a higher count on a model that supports it (see Settings below), the run generates that many.

Writing a prompt

Type your prompt into the box at the bottom of the card ("Describe the image..."). If Prompt is connected from another node, this box is disabled and shows "Prompt connected from input" instead.

You can connect more than one LLM, Text or Agent node to Prompt. Their texts are joined into one prompt, in the order you connected them.

The generated image appears in the preview area at the top of the card, with an animated effect while it's generating. If the run made more than one image, small dots sit at the bottom of the picture, one per image. Click a dot to see that image. Click Download in the settings bar to save the image on screen, this works even when the image is hosted elsewhere.

Click Save to Media, next to Download, to keep a copy in your Media Library, in the folder named after your workflow. It saves the image the card is showing, and the icon turns into a check once it's saved. If your storage is full, the card tells you and nothing is saved. If the node made several images in one run, use the dots to pick the one you want first.

Choosing a model

Click the model picker to search and choose a model. Options are grouped by provider, with prices shown where available.

The default model is Nano Banana Pro, with a default aspect ratio of 9:16.

If you connect or upload a reference image, the model picker automatically filters to models that support image-to-image generation. If your current model can't do that, the app switches you to one that can.

Settings on the card

  • Model picker
  • Aspect ratio chip
  • Resolution chip (only shown if the selected model supports it)
  • Seed chip (only shown if the selected model supports seeds): leave it empty and every run gives you a different image. Type a number, or hit the dice to get one, and runs with the same prompt and settings tend to land on the same image, so you can change one word at a time and see what it does. Models are not perfectly repeatable, and switching model or settings changes the result, so treat a seed as a way to stay close to a look rather than a guarantee. Clear the field to go back to random.
  • Images chip (only shown if the selected model can make more than one per run): pick how many images to generate, from 1 up to the model's limit
  • References chosen by Agent: an on/off button with the agent's icon (see below)
  • Download, Save to Media, Inputs / Outputs, Rename, Delete

Chips only appear when the selected model supports them, so swapping models changes what you see on the bar.

Letting the Agent choose the references

Normally this node uses every picture wired into Image ref. When an Agent writes this node's prompt, you can let it choose instead. Click the button with the agent's icon in the settings bar. Hover over it and it reads "References chosen by Agent: on" or "off". It is off on a new node.

With it on:

  • The Agent is told what is wired into Image ref, by name, and picks what this image needs. Only those are used. It can also pick none, and then the image uses no reference.
  • It needs an Agent connected to Prompt. If there is none, you are told before the run starts.
  • If the Agent has not chosen for this image yet, for example because you switched this on after it last ran, the image stops before it is made and asks you to run the Agent again.

What the Agent picked is on the Agent's card, under the prompt it wrote for this node. How it chooses: Choosing the references for each image.

Models available

This node runs on Atlas Cloud, like every node. You can choose from OpenAI (GPT Image 2.5 Sunburst, GPT Image 2.5 Flare, GPT Image 2), Microsoft (MAI Image 2.6), xAI (Grok Imagine 2.0), ByteDance (Seedream 5.0 Pro), Google (Nano Banana 2, Nano Banana 2 Lite, Nano Banana Pro), Qwen (Qwen Image 3.0), and Black Forest Labs (FLUX.2 Flex, FLUX.2 Pro). Prices marked "from" are the lowest tier: higher quality or resolution settings cost more. Check the model picker for what's currently available and its price.

Running the node

Select the node to reveal its Run button. A standalone node with no inputs runs itself; run it again to also run everything downstream. If its inputs already have results, running it runs everything downstream too. If inputs aren't ready yet, you'll be asked whether to run the whole workflow, or run everything up to and including this node.

Click the Inputs / Outputs button at any time to see exactly what data went in and came out.

PreviousScreenwriterNextVideo Generation