# Character training

> Train a custom model on a character's pictures in Character Studio for the closest likeness in Generate Image, for 1,500 credits in about 15 minutes.

Source: https://nodaro.ai/docs/guides/character-training

**Character training** teaches a custom image model one specific character, from the character's own pictures. After training, [Generate Image](https://nodaro.ai/docs/nodes/image/generate-image) draws that character with its trained model, which gives the closest likeness Nodaro offers. Training costs 1,500 credits and takes about 15 minutes.

Character training runs on Nodaro Cloud only. On self-hosted editions, the **LoRA** page does not appear in Character Studio.

## Reference pictures or a trained model

Without training, Generate Image sends the character's approved pictures to the model you selected, as references. After training, Generate Image runs the character's own trained model, a Flux LoRA, instead. The trained model has learned the character, so it needs no reference pictures.

| | Reference pictures (every character) | Trained model |
| --- | --- | --- |
| **Nodes** | Every image and video node that accepts references | Generate Image only |
| **Model** | The model you select on the node | The character's trained model |
| **Price per image** | The selected model's price | 20 credits |
| **Several characters in one image** | Yes | No. The node falls back to reference pictures |
| **Setup** | None | 1,500 credits and about 15 minutes |

Train a character once it has a good spread of on-model pictures and you need its closest likeness in many images. For video, the character's reference pictures remain the method; read [Consistent characters](https://nodaro.ai/docs/guides/consistent-characters).

## What training needs

- **At least 4 pictures.** Training uses the character's portrait, reference photos, expressions, poses, head angles, body angles and lighting variations. Motion clips and sheets are not used.
- **At most 20 pictures.** When you select more, Nodaro takes the first 20 in the order above.
- **Clean, on-model pictures.** The model learns whatever you give it, including a face that drifted. Unselect any picture that does not look like the character.

Build the character in [Character Studio](https://nodaro.ai/docs/guides/character-studio) first. The portrait, the five head angles and a few expressions give training the face from several sides.

## Train a character

### Open the LoRA page

Open the character in Character Studio and choose **LoRA** under **Resources**.

### Choose the training images

Under **Training images**, every picture the character has is selected. Click a picture to leave it out, or use **Select all** and **Clear**. The counter shows how many are selected, and **Train high-fidelity model** stays disabled until you select at least 4.

### Start the training

Click **Train high-fidelity model**. Nodaro reserves 1,500 credits and starts training.

### Wait about 15 minutes

The page shows **Training… (~15 min)**, and the node on the canvas shows **Training…**. You can close the studio and keep working.

### Check the result

When training succeeds, a green **Trained** badge appears on the LoRA page, on the Character Asset node and in the `@` mention list. The LoRA page also shows the trigger word and the training date.

## Use the trained character

Connect the Character Asset node to the **Assets** input of a Generate Image node, and mention the character in the prompt.

```
@maya:1 reading on a night train, soft window light
```

When the node uses this one trained character, Nodaro runs the trained model instead of the model selected under **Provider**.

- **The price is 20 credits per image.** It replaces the selected model's price.
- **The trigger word is added for you.** Nodaro puts it at the start of the prompt and removes the `@` mention tokens. Do not type it.
- **Reference pictures are not sent.** The trained model works from the prompt alone, so the node's other reference images are not used for that run.
- **Two or more characters use references.** When the image uses more than one character, the node runs the selected model with every character's reference pictures.
- **Other nodes are unchanged.** [Modify Image](https://nodaro.ai/docs/nodes/image/modify-image) and every video node always use the character's reference pictures.

Workflow: A trained character mentioned in the prompt makes Generate Image run its trained model; the result then becomes the first frame of a video.

- Character Asset → Generate Image (assets)
- Text → Generate Image (prompt)
- Generate Image → Generate Video (start frame)

To bring the trained likeness into video, make the still with Generate Image first, then use it as the start frame of [Generate Video](https://nodaro.ai/docs/nodes/video/generate-video).

## Re-train, cancel or remove

| Action | When | What happens |
| --- | --- | --- |
| **Cancel** | While the model trains | Training stops and the credits are refunded. Click **Train high-fidelity model** to start again. |
| **Re-train** | After a training succeeds | Trains again on the selected images and replaces the current model. Costs 1,500 credits each time. |
| **Remove** | After a training succeeds | Deletes the trained model. Generate Image goes back to reference pictures. |
| **Try again** | After a training fails | Starts a new training. The page shows why the last one failed, and a failed training is refunded. |

Deleting the character also cancels a training in progress, refunds it, and deletes the trained model.

## Credits

| Action | Credits |
| --- | --- |
| Training | 1,500 |
| Re-training | 1,500 |
| Each image from the trained model | 20, instead of the selected model's price |
| A training that fails or is cancelled | Refunded |

See [Credits](https://nodaro.ai/docs/concepts/credits) for how Nodaro reserves and refunds credits.

## Tips

- **Curate before you train.** Unselect drifted or blurry pictures; the model learns them too.
- **Cover the angles.** The portrait, the five head angles and a few expressions teach the face from several sides.
- **Keep one character per image.** The trained model runs only when the image uses a single character.
- **Keep references ready for video.** Training changes Generate Image only, so keep the character's pictures on-model for video work.

## From code and agents

The REST API starts a training, reads its status and removes the trained model. The status is one of `untrained`, `queued`, `training`, `succeeded`, `failed` or `cancelled`. Starting a training is limited to 3 requests per minute per token. See [Character training](https://nodaro.ai/docs/developers/api/character-training).

A `POST /v1/generate-image` call runs the trained model when its prompt mentions a single trained character that is connected to the node. See the [MCP tools](https://nodaro.ai/docs/mcp/tools) for generating images from an AI assistant.

## Frequently asked questions

### How much does character training cost?

Training costs 1,500 credits, and each re-training costs 1,500 credits again. Each image made with the trained model costs 20 credits, instead of the price of the model selected on the node. A failed or cancelled training is refunded.

### How many pictures do I need to train a character?

At least 4. Nodaro trains on up to 20 of the pictures you select, taken in this order — the portrait, reference photos, expressions, poses, head angles, body angles and lighting variations.

### How long does character training take?

About 15 minutes. You can close the studio and keep working. The Character Asset node shows Training… until the model is ready.

### Which nodes use the trained model?

Only Generate Image. Other image and video nodes, such as Modify Image and Generate Video, keep using the character's reference pictures.

### Can I use two trained characters in one image?

Not with their trained models. When a Generate Image node uses two or more characters, it runs the selected model with the characters' reference pictures instead.

### Can I train a character on a self-hosted install?

No. Character training runs on Nodaro Cloud only, and the LoRA page is hidden in Character Studio on other editions.