Why generated images don't always match the character

Worth understanding before you spend a long time troubleshooting, because some of this is a genuine limitation rather than a fault.

The image prompt is written from limited context. A language model does write the prompt for the image, but it only has the most recent message and the character's avatar to work from. It doesn't read your whole conversation. So the picture can miss details that seem obvious to you from earlier in the scene.

It works from a single reference image. The result will drift from your character if the avatar is low resolution, shows them at an unusual angle, or has unusual colouring.

Our older model has known biases. If you're using Legacy Mode, the underlying model tends towards certain body types and skin tones, which means male characters sometimes come out looking female and colouring isn't always accurate. This is a limitation of that model, not a settings problem.

What helps

Turn off Legacy Mode and use our newer models — see article 5.4.

Use the Qwen models. They're much better at matching an avatar.

Edit the prompt. Select the pencil icon at the bottom right of a generated image to adjust the prompt directly. The AI uses your version as its reference.

Reference the important details in your message, since the last message is what the prompt is built from. If you want a specific outfit or setting in the image, mentioning it in the message you send helps.

Use a clear, front-facing, good-quality avatar as your character's reference image.