Character training

Character training creates a dedicated model of one character from its images, for the closest likeness in Generate Image. You start a training with one call, poll it until it finishes, and remove the model when you no longer need it. After training, Generate Image uses the trained model automatically whenever a prompt mentions that character.

Character training runs on Nodaro Cloud only; on self-hosted installs the routes do not exist and answer 404. On a self-hosted install, keep a character consistent with its approved portrait and reference images, as described in Characters. The routes take a bearer token and act only on your own characters. See Authentication.

The SDK wraps the routes as client.characters.train, trainingStatus and deleteLora. In the CLI, use nodaro characters train, training and delete-lora. An AI assistant connected over MCP starts a training with train_character and follows it on get_character. Removing a trained model is not an MCP tool.

With an OAuth app token, starting a training and removing a trained model need assets:write, the same scope as creating or deleting the character. Without it they answer 403 insufficient_scope before anything is reserved. The status read needs no scope. A session or a personal API token needs no scope.

Endpoints

Before you train

The character needs at least 4 different images. By default, Nodaro collects them from the character, in this order, removes duplicates, and trains on up to 20:

1. The approved portrait.
2. The reference photos.
3. The expressions, the poses, the head angles and the body angles.
4. The lighting variants.

To train on a subset, send { "selectedImageUrls": [...] } in the body: up to 200 URLs of the character's own images. The training then uses only those, and at least 4 must remain. An omitted or empty list means every eligible image.

Character sheets do not count, because their views repeat the angles. When a character has fewer than 4 images, generate a few angles and expressions first with POST /v1/generate-character-asset; see Characters.

Start a training

POST /v1/characters/:id/train reserves 1,650 credits and starts the training. It answers 202 with the training's job id, its id and the character's trigger word. You never need to type the trigger word: Nodaro adds it to prompts for you.

curl -X POST https://app.nodaro.ai/v1/characters/3f6c2a9e-8d41-4b7a-9c35-1e2f7a6b0d94/train \
  -H "Authorization: Bearer $NODARO_API_KEY"

# Or train on a subset of the character's images
curl -X POST https://app.nodaro.ai/v1/characters/3f6c2a9e-8d41-4b7a-9c35-1e2f7a6b0d94/train \
  -H "Authorization: Bearer $NODARO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"selectedImageUrls": ["https://cdn.nodaro.ai/characters/kira-portrait.png", "https://cdn.nodaro.ai/characters/kira-smile.png", "https://cdn.nodaro.ai/characters/kira-left.png", "https://cdn.nodaro.ai/characters/kira-right.png"]}'

import { createClient, StaticTokenAuth } from '@nodaro/sdk'

const client = createClient({
  baseUrl: 'https://app.nodaro.ai',
  auth: new StaticTokenAuth(process.env.NODARO_API_KEY!),
})

const training = await client.characters.train(characterId)
// Or a subset: client.characters.train(characterId, { selectedImageUrls: [...] })

{
  "jobId": "7a9c1e3b-5d2f-4b8a-9c6e-1f3d5b7a9c2e",
  "trainingId": "q4m8x2k6p1",
  "triggerWord": "TOK_kira_a1b2c3"
}

The training's job, jobId, ends completed, failed or cancelled with the training, so you can also follow it with GET /v1/jobs/:id/status.

A double click is safe: while a training is queued or running, another start answers 409 already_training_or_not_found and reserves nothing. Each token may start 3 trainings per minute.

Follow the training

GET /v1/characters/:id/training returns the training's state. A training takes about 15 minutes. Poll every 10 seconds or so. The SDK has no wait helper.

curl https://app.nodaro.ai/v1/characters/3f6c2a9e-8d41-4b7a-9c35-1e2f7a6b0d94/training \
  -H "Authorization: Bearer $NODARO_API_KEY"

let training = await client.characters.trainingStatus(characterId)
while (training.status === 'queued' || training.status === 'training') {
  await new Promise((resolve) => setTimeout(resolve, 10_000))
  training = await client.characters.trainingStatus(characterId)
}
if (training.status !== 'succeeded') console.warn(training.status, training.error)

{
  "status": "succeeded",
  "trainingId": "q4m8x2k6p1",
  "error": null,
  "trainedAt": "2026-09-20T11:42:08Z",
  "version": "b81f4c0e9d27",
  "triggerWord": "TOK_kira_a1b2c3",
  "imageCount": 12
}

Generation uses the trained model only while status is succeeded. trainedAt and version stay set through a re-training, a failed or cancelled one, or one that could not be started, but the model is not used in any of those states. A character that is not yours, or is archived, answers 404 not_found.

Generate images with the trained model

Once the status is succeeded, run Generate Image with a prompt that mentions exactly one trained character with @, and with that character wired into the node. Nodaro then uses the trained model instead of the selected model and the reference images:

• Automatic. The trigger word is put at the start of the prompt for you, and your @ mentions are removed from the text.
• 22 credits per image. The trained model's price replaces the selected model's price.
• One trained character at a time. When the prompt mentions two or more trained characters, Generate Image falls back to the selected model with reference images.
• Generate Image only. Other nodes, such as Modify Image and the video nodes, keep using the character's reference images.

See Run a single node for the Generate Image request, and Character training for the same feature in the editor.

Re-train or remove the model

• Re-train. Start a training again at any time after one has succeeded, failed or been cancelled. It costs 1,650 credits every time and replaces the previous model.
• Remove. DELETE /v1/characters/:id/lora cancels a training in progress and refunds its credits, deletes the trained model, and clears the character's training fields. The status then reads untrained. It returns { ok: true }, and it cannot be undone. Generations go back to reference images.
• Archive the character. Archiving the character also cancels a training in progress, refunds its credits and deletes the trained model.

curl -X DELETE https://app.nodaro.ai/v1/characters/3f6c2a9e-8d41-4b7a-9c35-1e2f7a6b0d94/lora \
  -H "Authorization: Bearer $NODARO_API_KEY"

Credits

See Credits for balances and transactions.

Errors

The refusals these routes write themselves use an older body shape: the code is a bare string, { "error": "<code>" }, sometimes with a top-level message. The scope, credit, rate-limit and job-policy refusals use the usual { "error": { "code", "message" } } envelope. The SDK reads both shapes into NodaroError.code and .message.