Upload a selfie, ask for a cartoon, and the first result usually looks great. Then you ask for slightly smaller eyes. Then a warmer background. By the fourth round, the jawline has shifted, the freckles are gone, and the cartoon looks more like your cousin than you.
That drift is the real problem with turning photos into cartoon art inside an AI chat, and most how-to guides skip it. Dedicated AI image to cartoon tools let you redraw one region of a picture instead of the whole thing. Chat assistants that call image models for you have made generating easier and keeping a likeness harder. Here’s how the workflow runs in Claude today, and how to stop the face from wandering off.
What Can Claude Do With Your Photo on Its Own?
Out of the box, Claude reads images. It can describe your photo’s lighting, point out the features that make a face recognizable, suggest a style for your audience, and write a detailed generation prompt. It can also build diagrams and simple SVG graphics from code.
It doesn’t paint pixels. Anthropic’s help center says Claude doesn’t produce photos or illustrations the way image generators do. For a long time that meant a two-tab routine: write the prompt in Claude, copy it, paste it into an image tool, upload the photo again, and repeat.
Connectors ended that routine. Through the Model Context Protocol, Claude can call outside tools from inside the chat, and image models are among them.
Setting Up the Claude AI Image Generator Connector
OpenArt’s Claude AI image generator connector plugs its image models into Claude. OpenArt says setup takes about a minute:
- Open Settings in the Claude app or on claude.ai, go to Connectors, and choose Add custom connector.
- Name it OpenArt and paste https://mcp.openart.ai/mcp into the URL field.
- Click Connect and sign in with your OpenArt account. You don’t need an API key.
OpenArt’s guide suggests setting tool permissions to Always Allow. You might keep approvals on for the first few sessions so you can watch each call before it spends credits.
From there, you upload a photo in the chat and describe the change you want. OpenArt lists GPT Image 2.5, Nano Banana 2 Lite, Seedream 5.0, Wan 2.7 Image and Grok Image 2.0 as available models. You can name one, or let Claude choose. Finished images appear in the chat and save to your OpenArt library.
On cost, OpenArt reports that the connector carries no extra fee. Images draw from your regular OpenArt credits, MCP access comes with plans from Starter up, and new accounts get 40 free credits. Prices change often, so check OpenArt’s pricing page before you plan a big batch.
Why Faces Drift When an Agent Holds the Pencil
Here’s the catch. When Claude runs the job, it writes the prompt, may pick the model, and decides whether your request got done. Its target is your latest message. Your face stays safe only as long as the instructions protect it.
Take that “smaller eyes” request. Claude rewrites the prompt with the new detail and sends off a fresh generation. The model redraws the whole portrait, so anything you didn’t mention is up for grabs: the slight tilt of your smile, the cowlick, the small scar on your eyebrow. Ask for a new look, and Claude may switch models as well, which brings a fresh interpretation of your features.
OpenArt’s own cartoon page describes this exact risk, and its answer on the website is inpainting, which repaints just one area. The chat route moves fast and handles broad changes well. Precise touch-ups need more care.
Claude isn’t alone here. The newer class of image agents that rewrite prompts makes the same trade-off between speed and control. Skipping the check costs more later, as the freelancers now hired for fixing AI-generated designs can tell you.
How to Keep the Likeness Locked
Small wording changes make a big difference in what the model treats as fixed.
| What you want | Vague ask | Ask that protects the face |
| Smaller eyes | “Make the eyes smaller” | “Edit only the eyes. Keep face shape, hairline, and skin tone identical to the last image.” |
| A new style | “Try it as anime” | “Same model as before, anime style. Keep my jaw, nose and freckles from the original photo.” |
| A new background | “Change the background” | “Swap the background for a plain teal wall and leave the subject untouched.” |
A few habits help even more:
- Lock the model. Once a result looks right, name that model in every follow-up.
- Change one thing per message. If the face shifts, you’ll know which request caused it.
- Put Claude’s eyes to work. Download the cartoon, upload it beside your original photo, and ask Claude to list every difference in face shape, hair and distinguishing marks. You stop noticing drift after the tenth look. A side-by-side check doesn’t.
- Finish small fixes on the website. Results sync to your OpenArt library, so open the keeper there and use inpainting for the last corrections.
A Prompt That Puts Likeness First
Most cartoon prompts spend their words on style. Lead with identity instead:
Turn the attached portrait into a clean 2D cartoon for a profile picture. Keep these exactly: oval face, center-parted dark hair, small scar above the left eyebrow, slightly lopsided smile. Style: bold outlines, flat color, soft cel shading, plain warm background. Square crop, head and shoulders.
Let Claude draft this part. Ask it to study your photo and list the five features a friend would recognize you by, then build the prompt around them. Reading an image is something Claude does well, and the list gives the image model something concrete to protect.
The source photo still matters. OpenArt recommends front-facing portraits in decent light, uploaded as JPEG, PNG or WEBP. Its cartoon page says group shots and pet photos work too.
Which Style Suits Which Job?
| Use | Style that tends to fit | Watch for |
| Social avatar | Flat-color classic toon | Detail that disappears at small sizes |
| Blog or tutorial art | Comic or ink illustration | Matching the look across a series |
| Character concepts | Anime or Pixar-like 3D | Proportions changing between poses |
| Gifts and prints | Watercolor or 3D | Resolution, so upscale before printing |
OpenArt lists Western cartoon, anime, comic book, watercolor, and pixel art among its styles, along with Pixar-like 3D. For a recurring series, a reusable character built on OpenArt keeps the same face from scene to scene.
Before You Upload Someone Else’s Face
A face photo is personal data. In this workflow, it travels from your Claude chat to OpenArt and stays in your OpenArt library. Read both services’ privacy terms, and get permission before you cartoonize a friend, a coworker, or a child.
Commercial use raises separate questions. OpenArt says it makes no ownership claim on generated images and allows commercial use on Plus plans and above. It also offers an IP Safety Check, while noting the check flags risk and doesn’t count as legal clearance. If a cartoon will help sell a product, talk to a lawyer about the rules where you operate.
So, Chat or Editor?
Use Claude when the hard part is deciding what you want: the style, the audience, the features worth protecting. Move to OpenArt’s editor when the hard part is a precise fix. Plenty of people end up doing both in one sitting.
Next time a cartoon comes back looking almost like you, scroll up and find the round where it stopped being you. That message usually shows which tool should finish the job.
Related: AI Image Generation vs Editing: What’s the Difference?
