Can Claude generate images? Not by itself — here's how to add it

What Claude can do with images right now

Claude is very good at reading images. Upload a screenshot, a chart, a whiteboard photo or a UI mock and it will describe it, extract the text, explain the diagram or critique the layout. In Claude Code you can drag an image into the terminal and ask it to build the UI you sketched.

Claude can also build certain visuals from scratch, as long as they are made of code:

• Diagrams and flowcharts (Mermaid, SVG)
• Charts from your data (HTML/JavaScript artifacts)
• Icons, logos and simple illustrations as SVG
• Slides, mockups and prototypes with Claude Design

What it cannot do is produce a raster image from a text prompt. There is no diffusion model inside Claude. Anthropic's own help center says it plainly: "Claude doesn't generate photos or illustrations the way image-generation tools do."

Why doesn't Claude have an image generator?

Anthropic has focused Claude on text, code, reasoning and agents rather than bundling a media model, and there is no announced plan to change that. That is not a limitation you have to live with, though. Claude is built to call external tools, and image generation is one of the tools it calls best, because it can look at the result and decide whether to try again.

How to generate images in Claude: connect an MCP server

MCP (Model Context Protocol) is an open standard that lets an AI assistant use outside tools. An MCP server is a small program that exposes a set of tools; the assistant reads the list, picks the right one, calls it and gets the result back in the conversation. Claude Code, Claude Desktop, Cursor, Windsurf and most other agent clients support it.

For image generation, the server has to do more than expose a "generate" button. A good one lets Claude:

1. Browse models and see what each is for (photorealism, anime, pixel art, video).
2. Read how to prompt each model, because a booru-tag model and a natural-language model want very different prompts.
3. Generate and get the finished image inside the tool result, so Claude can judge it and iterate.
4. Pay without a human copying API keys around.

Imference's MCP server does all four. It is free and open source (MIT, on npm); you only pay for the generations, from $0.003 per image, with no subscription.

Claude Code: one command

claude mcp add imference -e IMFERENCE_API_KEY=your-api-key -- npx -y imference-mcp

Get an API key and a few credits on the pricing page (1 credit = $0.001). Restart Claude Code and ask:

Generate a photorealistic banner of a lighthouse in a storm. Pick the best model for it and apply its prompting recommendations.

Claude will call list_models, read the chosen model's card with get_model, prepend the recommended quality prefix, pick a banner format, call generate, and show you the image. If it doesn't match, tell it what to change; it edits the prompt and runs again.

Claude Desktop, Cursor and other clients

Add this to your client's MCP configuration (Claude Desktop: Settings → Developer → Edit Config; Cursor: Settings → MCP):

{
  "mcpServers": {
    "imference": {
      "command": "npx",
      "args": ["-y", "imference-mcp"],
      "env": { "IMFERENCE_API_KEY": "your-api-key" }
    }
  }
}

Restart the client. Requires Node.js 18 or newer. If the first launch times out while npx downloads the package, clear the npx cache or install it globally with npm i -g imference-mcp and use imference-mcp as the command.

What it looks like in practice

The important part is the last step: Claude sees the image. That is what turns "generate an image" into "generate the image I asked for".

Which models can Claude use?

The catalog has 45+ image models and 3 video models and updates without touching the server, because model cards, parameters and prices come from the API. Families include SDXL, SD 1.5, FLUX.1, Z-Image, Chroma, Qwen-Image, Anima and Krea 2 Turbo for images, and Wan 2.2 and MiniMax-H3 (video with audio) for video. Ask Claude to "list the anime models" or "pick the cheapest photorealistic model" and it will.

Can Claude generate images for free?

Two honest options:

• Free and local: Imference Desktop is a free Windows/macOS app that runs seven model families on your own GPU (6 GB VRAM and up, or Apple Silicon). It is not driven by Claude, but if what you want is free images, it is the answer (see how to run Stable Diffusion locally).
• Through Claude, pay per image: the MCP route costs from $0.003 per image with no subscription. A $5 credit pack is more than a thousand images at the base price.

There is also a no-account option: set a wallet key instead of an API key and the server pays per generation in USDC on Base via the x402 protocol, with a hard spending cap (default $10 per payment). That is meant for autonomous agents that cannot sign up for anything.

Can Claude edit images?

Claude can look at an image and tell you what to change, but it cannot repaint pixels itself. With the MCP server connected, several models in the catalog accept an input image (image-to-image); ask Claude to check the model card with get_model to see whether the model you want supports it, then describe the edit.

Can Claude generate videos?

Same story as images: not natively, yes through the same MCP server. Wan 2.2 does text-to-video and image-to-video; MiniMax-H3 generates video with audio. Videos take longer, so Claude polls with check_status and shows the file when it is ready.

Claude vs ChatGPT for image generation

ChatGPT has image generation built in, and for a quick one-off picture that is simpler. The MCP route has three advantages once you generate regularly: you choose among dozens of specialised models instead of one, the agent reads each model's prompting guide so results are consistent, and you pay per image rather than per month. If you already work in Claude Code or Cursor, it also keeps images inside the workflow: Claude can generate a hero image and drop it straight into your project.

FAQ