> ## Knowledge Base Index
> Fetch the complete knowledge base index at: https://help.atlabs.ai/sitemap.xml
> Use this file to discover available pages before exploring further.
> Pure-Markdown content can be obtained by appending a '.md' suffix to the content URLs listed in the sitemap (without the trailing slash).

# Scenes: Generate vs. Edit Image

## 🎨 Should I use Generate or Edit Image to fix my scene's visual?

The Visual tab has two tabs at the top: **Generate** and **Edit Image.** Picking the wrong one for a small fix is one of the most common reasons a character ends up looking completely different than intended.

![](https://storage.crisp.chat/users/helpdesk/website/-/4/1/a/c/41accd17b39fe400/screenshot-2026-09-23-at-82722_1oiwr85.png)

#### The rule

* **Use Generate** when you want to replace the scene's image entirely: a new prompt, a different composition, a different moment. Tag your characters, objects, and locations with '@' and write the full shot from scratch.
* **Use Edit Image** when you want a targeted, isolated change: a shirt color, a removed object, a camera angle nudge, without touching anything else in the frame.

🚨 A common mistake: going to Generate and writing something like *"change his shirt to blue."* Generate re-runs the entire image from your prompt; it doesn't know to leave everything else untouched, so the whole shot can come out different, not just the shirt. For a small change like that, use Edit Image instead.

#### Fixing a character that's drifted from their reference

If a character has drifted from how they're supposed to look, anchor the fix to your locked character reference:

![](https://storage.crisp.chat/users/helpdesk/website/-/4/1/a/c/41accd17b39fe400/screenshot-2026-09-23-at-91135_1bup8p9.png)

1. Open the scene, go to **Visual → Edit Image.**
2. Write a prompt referencing both the current frame and the correct character, for example: *"Update the current frame, make the actor look like @current\_frame, and match @Leo."*
3. This gives Atlabs two references to work from: the scene's current composition and the correct character appearance, rather than one.

💡 **Tip:** Image and motion models are probabilistic; the same prompt can occasionally produce different results even when tagged correctly. If a fix doesn't land the first time, retrying with the same prompt and switching the model can help.