> ## Documentation Index
> Fetch the complete documentation index at: https://docs.omnichar.org/llms.txt
> Use this file to discover all available pages before exploring further.

# Finetune a character in Omnichar Cloud

> Change the references a character is built from, rebuild it, and test the result against a continuity score, all in Omnichar Cloud.

A character that comes back wrong in some shots usually needs better reference photos rather than
a better prompt. **Finetune** is where you change the photos a character is built from, rebuild it,
and test the rebuild against the same score the rest of the app uses.

Open it from **Finetune** in the sidebar, from a character tile, or from the **Finetune this
character** link that appears under any take that scored below 70.

## What the page shows

Pick a character from the switcher at the top. Everything below it applies to that character.

| Card | What it is for |
| - | - |
| **Face** | Front and three-quarter views, no sunglasses. |
| **Body & Clothes** | Full-length photos, and the outfit either worn or laid flat. |
| **Voice** | About 30 seconds of clean speech, uploaded or recorded in the browser. |
| **Description** | The locked description that goes out with every generation. |
| **Recent scores** | How closely recent takes matched, with the weakest signal named. |
| **Test** | Run a shot through the character and score it without leaving the page. |

Each reference card carries the score for the part of the character it covers, and the weakest
card is flagged with a note telling you which photo to add. If the face averages lowest, the Face
card says so.

> [!NOTE]
>
> Characters built before Omnichar recorded which photo filled which slot show one **References**
> card instead of two. Rebuilding one records the split, and the two cards appear from then on.

## Change the references and rebuild

**1. Add or remove photos**

Use the **+** on a card to add a photo, or the **x** on a thumbnail to drop one. A character
needs at least one reference picture to build.

**2. Rebuild**

The header switches to **Unsaved changes** as soon as anything differs from the saved
character. Press **Rebuild** to apply it. A rebuild is priced the same way a build is, by the
time it takes, which recently came to 2-3 credits.

**3. Test it**

New takes use the new references from the moment the rebuild lands. The Test card below is the
quickest way to see the difference.

Discard puts the cards back to the character as it stands. Nothing is applied until you rebuild, so
you can try a set of photos and walk away from it.

## Test the character

The Test card runs a real generation against the references currently saved.

* Pick one of the four suggested shots, or type your own.
* Choose image or video, the renderer, an aspect ratio, a resolution, and 1 or 3 takes.
* The price shows before you run. Three takes are three runs, each priced and scored on its own.
* Double-click a result to open it full size. Click outside it or press Escape to close.

Every test lands in a single session named **Finetune**, so testing does not leave a new session
behind every time. If you close the tab while a test is rendering, the placeholders are still there
when you come back.

## Reading the score history

**Recent scores** charts the last 20 takes this character was cast in, oldest on the left. The
dashed line marks 70, where a score counts as strong, and the lowest bar is capped in white.
Switch between Overall, Face, Subject and Wardrobe to find which signal is dragging the average
down, then add photos to the card that covers it.

See [Continuity scoring](https://docs.omnichar.org/cloud/scoring.md) for what each number measures.

## Next steps

> [card] [Continuity scoring](https://docs.omnichar.org/cloud/scoring.md)
>
> What the score measures, and where it is blind.

> [card] [The .char file](https://docs.omnichar.org/characters.md)
>
> What a character holds and which models take one.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.