Can Claude Generate Images? Yes, With a Connector

Not by itself, but yes once you add an image connector. Claude's own help article, Can Claude produce images?, says Claude doesn't make photos or illustrations the way image tools do: it reads images you upload and builds diagrams, charts and interactive visuals. Add a custom connector such as Visualizee (our tool) and Claude can call a real image model from the chat, show you the result and keep going. We made the four images below that way on October 29, 2026, from inside Claude, in 21 to 103 seconds each.

Quick answer: Claude has no built-in image generator. In Claude on the web or Desktop, go to Customize → Connectors → + Add → Add custom connector, paste https://visualizee.app/api/mcp, sign in to Visualizee once, then ask for a render in any chat. Custom connectors work on every Claude plan, Free included (one connector). A Pro-quality render took 21 to 40 seconds and 150 credits in our test; an Ultra render took 103 seconds and 300 credits.

All setup guides for Claude and 13 other agents live on our MCP server page. This post is the Claude walkthrough with real results.

What Claude Can and Can't Do With Images

Out of the box, Claude reads images but doesn't draw photos. Upload a floor plan, a room photo or a screenshot and it can describe it, extract text, critique a layout or write code from it; ask for a photo of a living room and it will offer a description, an SVG sketch or a chart instead.

Anthropic's help article says it plainly: Claude doesn't generate photos or illustrations the way image-generation tools do. The diagrams and charts it builds are code it writes, not pictures from an image model. The way around it is official: Claude supports custom connectors using remote MCP, and an image service that speaks MCP becomes a tool Claude can call.

How to Add an Image Connector to Claude

It takes about two minutes and no API key. You add the server address once, sign in once, and switch the connector on in the chats where you want images.

1. Open the connectors settings. In Claude on the web or in Claude Desktop, go to Customize → Connectors, click + Add and choose Add custom connector.
2. Name it and paste the URL. Call it Visualizee and paste the server URL: https://visualizee.app/api/mcp. Click Continue.
3. Leave the advanced settings empty. Visualizee signs you in with OAuth, so you need no API key or request headers. Click Add.
4. Sign in to Visualizee. Choose to sign in now (or later, with Connect in the connectors list). A Visualizee page opens; sign in or create a free account and approve access. Claude never sees your password.
5. Turn it on in a chat. In a conversation, click +, choose Connectors and switch Visualizee on. Then ask for an image in plain words.

Which plans: Anthropic lists custom connectors on the Free, Pro, Max, Team and Enterprise plans; Free users are limited to one custom connector. On Team and Enterprise an Owner adds the connector under Organization settings → Connectors and each member then signs in from Customize → Connectors. We checked the help center on October 29, 2026; menus move, so if a name differs, look for "Connectors".

Visualizee isn't listed in Anthropic's connector directory, so you add it by URL rather than from the directory, the same custom-connector route any remote MCP server uses.

Four Images Made Inside Claude

All four jobs worked from a Claude chat on the first run: 21 to 103 seconds and 150 to 300 credits each, 1,050 credits in total. Three kept every element we named; the room lost a wall socket. Each example below is a real call from a Claude chat with the Visualizee connector on, one run each, no retries. Under each one: what you would type, the exact instruction Claude sent to the connector, how long the render took on the server and what it cost. Every render was saved in a project called "Blog posts" and a chat named after this post, so they can be reopened and edited in the app.

1. A SketchUp-style model screenshot to a photoreal exterior

This is the same job as sketch to render.

What to type in Claude: "Make this model screenshot photoreal, late afternoon. Keep the house exactly as modeled. Use the highest quality."

What Claude sent: "Turn this model screenshot into a photoreal exterior render on a clear late afternoon. Keep exactly: the camera angle, the house shape and proportions, the flat roof and its overhang, every window and the front door in the same place and size, the cantilevered wood-clad upper floor, the carport on the right and the path. Change only: real materials (white smooth render on the lower floor, vertical larch boards on the upper floor, dark gray window frames, real glass with reflections), a real lawn and low planting along the path, a real sky. Don't add or remove any window, door, column or floor."

Result: Ultra quality, 103 seconds, 300 credits. Every window, the door, the carport and the path stayed where the model put them; the render added trees and a hedge behind the house, which the instruction left open. Because "highest quality" maps to the slowest model, this one came back as pending first and was collected about a minute later (more on that in the section on slow renders below).

2. Furnishing a photo of an empty room

What to type in Claude: "Furnish this empty room as a calm Scandinavian living room. Don't touch the windows, radiator, door, floor or ceiling light."

What Claude sent: "Furnish this empty room as a calm, lived-in Scandinavian living room: a light gray three-seat sofa facing the windows, a round oak coffee table, a wool rug, one lounge chair, a floor lamp, a low sideboard against the back wall, a large plant and two framed prints. Keep exactly: the camera, the walls, both windows and their frames, the radiator, the doorway to the hallway, the oak floor, the ceiling and the light fitting. Don't add or move any window or door, and don't change the floor or wall color."

Result: Pro quality, 34 seconds, 150 credits. The camera, both windows, the radiator, the doorway, the floor and the bare bulb are all where they were. One slip: the wall socket next to the doorway disappeared, part of the walls the instruction kept. For the full staging workflow, see virtual staging with AI.

3. A new colorway for a branded product

What to type in Claude: "Make a studio packshot of a white insulated bottle with the logo TARNWICK. Then make the same bottle in terracotta with an off-white logo."

What Claude sent (second step): "Make a new colorway of this exact bottle: change only the powder-coated body color from chalk white to a deep terracotta orange, with the "TARNWICK" logo printed in off-white instead of black. Keep exactly: the bottle shape and proportions, the brushed steel cap and carry loop, the logo's wording, size, font and position, the matte finish, the camera angle, the background, the light and the shadow. Nothing else written anywhere."

Result: two Pro renders, 21 seconds for the white base (made from text) and 27 seconds for the colorway, 150 credits each. The shape, cap, loop and logo position held and the logo is still spelled TARNWICK. Tarnwick is a brand we made up for this test; a web search found no product by that name. To run this on your own sample photo, see colorways from one sample photo.

4. A car wrap on a parking-lot photo

What to type in Claude: "Wrap this car in satin forest green. Keep everything else as it is."

What Claude sent: "Wrap this car in a satin forest green vinyl wrap, covering every painted body panel. Keep exactly: the car's body shape, every panel line and shut line, the headlights, grille, mirrors, windows and trim, the wheels and tires, the ride height, the camera angle, the parking lot, the hedge and the light. Change nothing except the body color and finish. No badges, logos or text."

Result: Pro quality, 40 seconds, 150 credits. Wheels, lights, grille, the blank plate and the background stayed put. We then sent the identical request again a minute later, as you might after a dropped connection: it returned the same image with "no new charge", and the server logged one render, not two. For more wrap colors, see car wrap colors.

Total for the four examples: 5 renders, 1,050 credits, all on the quality presets Claude picks by default; "highest quality" was the only time we asked for more.

Claude With a Connector vs ChatGPT's Built-In Images

ChatGPT can make images by itself; Claude needs a connector. For design work the deciding question is not who draws, but whether your design survives the edit. OpenAI's help article Images in ChatGPT says ChatGPT Images is available on all tiers, can edit an uploaded image, and warns that selections in its editor "are not always precise", so edits may spread past the area you picked.

Here is how the two setups differ, from what each vendor documents and what we measured above. We have not run ChatGPT on these four inputs; that side-by-side is the next post in this series.

If you want a quick illustration in a chat, ChatGPT's built-in images are the shorter path. If you are showing a client their own house, room, product or car, the job is to keep the design and change one thing, and that is where a design-focused tool earns its place. Model choice matters as much as the chat you use: our best AI model for architectural rendering hub runs seven models on six design jobs, and the Nano Banana 2.1 vs Pro test shows how much a cheaper model moves. For what AI rendering is and where it stops, see AI rendering.

What Happens When a Render Is Slow

In our test every Pro render finished in under 45 seconds and Claude showed it right away. Slower ones, like the 103-second Ultra exterior above, come back as "still rendering" with a receipt number, and Claude collects the image a moment later without you doing anything.

In plain words: Claude waits up to about 45 seconds for an image. If the model needs longer, the connector answers "pending, ready in about 55 seconds" and gives Claude an id. Claude then asks for that id (it waits up to another 45 seconds each time) until the image is ready. In our test that took two check-ins, and the render was charged once.

Two rules make this safe. Collecting a pending render costs nothing. And if a connection drops and Claude repeats exactly the same request within 10 minutes, Visualizee returns the render already in progress instead of starting and charging a new one, which is what happened with the car above. If an old chat shows "server isn't responding" after a render request, the image is likely in your Visualizee project already.

What It Can't Do

The connector gives Claude an image tool, not a new skill set. These are the limits we hit or know of.

• Claude doesn't see the result the way you do. It gets a small preview and the link, so it can describe the image and spot obvious misses, but check fine details yourself: in the room above, a wall socket vanished.
• Models add what you leave open. The exterior gained trees and a hedge nobody asked for. Name every element that must stay, and check additions as well as losses.
• Ultra is slow. 103 seconds here, against 21 to 40 seconds on Pro. Explore on Pro or Base and save Ultra for the final shot.
• No layers, no exact sizes. A render is a picture: it has no dimensions, no editable layers and no guarantee that a sofa would fit the room. Sizes come from your drawings.
• It isn't in Anthropic's directory. You add Visualizee by URL as a custom connector, on the web or in Desktop. We haven't tested the Claude mobile apps for this post.
• One run each, one day. Another run can come out better or worse. This shows what to expect, not a guarantee.

Frequently Asked Questions

Can Claude generate images on its own?

No. Claude's help article says it doesn't generate photos or illustrations the way image tools do; it reads images you upload and draws diagrams, charts and interactive visuals. For photoreal images inside a Claude chat, add an image tool as a custom connector, such as Visualizee's MCP server at https://visualizee.app/api/mcp.

Is generating images in Claude free?

Adding a custom connector is free on every Claude plan, Free included (one custom connector). The images are paid for by the image service: Visualizee's Free plan gives 5 renders, or 4 if you use your one Pro render, with no card. On the $40 Pro plan a Pro-quality render costs about $0.31. See pricing.

Which Claude plans can use a custom connector?

All of them. Anthropic lists custom connectors using remote MCP on Free, Pro, Max, Team and Enterprise, in Claude on the web and in Claude Desktop; Free is limited to one custom connector. On Team and Enterprise an Owner adds the connector for the organization and members then sign in to it.

What happens to my photos and prompts?

Claude sends the connector only what the render needs: the instruction it writes, the image links and the project and chat names. Renders and uploads are saved in your Visualizee workspace, in the project and chat Claude names. The image files sit at long, unguessable links so Claude can show them, so anyone you give a link to can open that file. Read the privacy policy before you upload client work.

Which image models does Claude use through Visualizee?

Claude picks a quality, not a model: Base (Grok Imagine 2.0, 80 credits at 1K), Pro (GPT Image 2.5 Flare, 150 credits, the default) or Ultra (GPT Image 2.5 Sunburst, 300 credits). Ask for a specific model by name and Claude can pass it instead, for example Nano Banana Pro or Seedream 5.0 Pro.

Does this work in Claude Code?

Yes. Run claude mcp add --transport http visualizee https://visualizee.app/api/mcp once in your terminal, then run /mcp inside a session, pick visualizee and sign in. The same tools then work while you code, for example for textures, UI art or product shots for a site.

The Short Version

Claude can't draw photos by itself, and Anthropic says so. Add Visualizee as a custom connector (https://visualizee.app/api/mcp, any Claude plan, two minutes) and Claude can make and edit photoreal images in the chat: four design jobs here took 21 to 103 seconds and 150 to 300 credits each, slow renders were collected automatically, and a repeated request wasn't charged twice.

---

Try it in your own Claude chat. Add the connector from our MCP server page, or start in the app with a model screenshot or a photo and tell Vizzy what to keep and what to change: try Visualizee free, 5 renders (4 if you use your one Pro render), no card.