Can Claude Generate Images? Not on Its Own, but You Can Add an Image Model in Two Minutes

What can Claude do with images today?

A lot of confusion comes from the fact that Claude is clearly good with images. It reads screenshots, describes photos, pulls numbers out of charts and spots problems in a UI mockup. It just cannot paint. Anthropic's developer docs put it plainly: Claude is an image understanding model and cannot generate, edit or manipulate images.

The custom visuals feature is useful and easy to mistake for image generation. When you ask for a flowchart or a comparison chart, Claude writes HTML and SVG and renders it in the reply. You can copy a snapshot or download the SVG. It is a whiteboard, though, not an image model, and it falls apart the moment you ask for a cat in a spacesuit.

How do you generate images in Claude?

Claude supports connectors, which are built on the Model Context Protocol. A connector gives Claude new tools. Add one that wraps an image model and Claude can call it whenever you ask for a picture, then show you the result in the chat.

We built the 3D AI Studio connector for exactly this, and for the step after it: the same connection also makes 3D models. Here is the setup in Claude on the web or desktop.

1. 1

   Open Settings and then Connectors

   In Claude on the web or the desktop app. On Team and Enterprise plans an owner may have to add the connector for the organisation first.

2. 2

   Add a custom connector

   Name it 3D AI Studio and paste https://mcp.3daistudio.com/mcp as the remote MCP server URL. Leave the advanced fields empty.

3. 3

   Connect and sign in

   A 3D AI Studio page opens. Sign in, check what Claude will be allowed to do, and click Authorize. There is no API key to copy.

4. 4

   Ask for an image by name

   Start a new chat and say "use 3D AI Studio to make an image of...". Naming the tool stops Claude from offering an SVG instead.

claude mcp add --transport http --scope user 3daistudio https://mcp.3daistudio.com/mcp

Can Claude edit images and photos?

With a connector, yes, and this is where it gets more useful than a plain image generator. You attach an image, describe the change in normal words, and Claude passes both to the model. Two real edits:

“Add armor to the character”

“Turn me into a stylized character”

Current edit models keep the parts you did not mention, which is the thing older image tools got wrong. The face, pose and colours survive; only the requested change happens. You can also hand over several images at once, for example a character and a background, and ask Claude to combine them. GPT Image 2.5 takes up to 10 input images and Nano Banana 2.1 up to 14.

How Claude gets your image to the model

One quirk worth knowing: Claude can see an image you attach, but it cannot pass the file itself to a connector. So 3D AI Studio shows a small upload card in the chat. You drop the file in, say "done", and Claude carries on. In Claude Code you get a short upload link instead.

Which image model should Claude use?

You do not have to pick. If you just ask for an image, Claude chooses a sensible default. When you care about the result, these are the differences that matter:

• Text in the image (a sign, a label, a poster headline): ask for GPT Image 2.5. It spells reliably.
• The same character in five poses: Nano Banana 2.1 is the one we reach for. It keeps a design steady across a long series of images.
• Reference art you will turn into 3D: either works. Ask for a plain background and even lighting, because shadows end up painted into the texture.

What are the other ways to get images in Claude?

3D AI Studio is not the only route, and depending on what you already pay for, another one might fit better. A fair comparison:

Our view, with the obvious bias declared: for most people the connector route wins because it needs no API key, works in the Claude apps rather than only the desktop app, and keeps every image in a dashboard where you can find it later. The community servers are good if you are a developer with a Gemini key already, but they run locally, so they do not reach Claude on the web or on your phone.

Can Claude turn an image into a 3D model?

This is the part a plain image connector cannot do. Once Claude has made an image you like, ask it to "turn this into a 3D model". The connector sends the image to a 3D model, Prism 3.1 by default, and a couple of minutes later a textured mesh appears in an interactive viewer in the chat. You can export it as GLB, FBX, OBJ, STL, USDZ or 3MF.

The models below went through that exact process: an image in, a textured 3D model out. Pick an input and drag the model to look around it.

Choose an input image

Prompts to try in Claude

Real prompts and the images they produced. Start your message with "Use 3D AI Studio to..." so Claude reaches for the connector.

Generate

Generate

Generate

Generate

Generate

Edit

Getting better images out of Claude

• Describe the subject first, then the style, then the framing. "A red fox in a scarf, children's book illustration, centred, plain background" beats a paragraph of adjectives.
• Let Claude write the prompt. Describe what you want in your own words and ask it to expand that into a detailed image prompt before generating. It is very good at this.
• Iterate with edits instead of starting over. "Same image, but at night" keeps everything you liked.
• Say the aspect ratio when it matters: 16:9 for a banner, 9:16 for a phone wallpaper, 1:1 for an icon.

> make 4 item icons for the inventory: potion, key, scroll, gem. flat background, same style
● 3daistudio · generate_image_gpt (4 prompts, 1:1)
  └ 4 images generated
Saved them to assets/icons/: potion.png, key.png, scroll.png, gem.png.
Want transparent backgrounds? I can run remove_background on all four.

Frequently asked questions