# SAM 3 Video

> Segment and track objects through a video from a text prompt — name several objects separated by commas and SAM 3 keeps each one masked across every frame.

## Overview

- **Model ID**: `sam-3-video`
- **Workflow**: `sam-3/video`
- **Provider**: Meta
- **Category**: Video → Video (video)
- **Output**: A video URL.
- **Tags**: segmentation, object tracking, video input
- **Unified endpoint**: `POST https://api.picsart.com/workflows/v1/models/execute` (model id in the body)
- **Compatibility endpoint**: `POST https://api.picsart.com/workflows/sam-3/video/execute` (vendor's own parameters)
- **Execution**: long-running — use the async submit/poll pair or the SDK, which polls for you
- **Model page**: https://picsart.com/api-platform/models/sam-3-video

## Pricing

Rates are in USD. Every row below is an exact rate for the settings it names — the columns are the settings that change the price for this model. Every response reports what was actually charged in `response.usage.credits`.

| Rate |
| --- |
| $0.011 / sec |

Published rates. Volume pricing, invoicing and annual agreements: https://picsart.com/api-platform/contact-sales

## Authentication

The SDK and the REST API authenticate with a Picsart API key sent as a bearer token. Create a key in the console at https://picsart.com/api-platform/dashboard/api-keys, keep it in a server-side environment variable, and never ship it in browser code or commit it.

```bash
export PICSART_API_KEY="YOUR_KEY"
```

```http
Authorization: Bearer $PICSART_API_KEY
```

The CLI and the MCP server do not use an API key: `gen-ai login` signs in through the browser once per machine and stores a token locally.

## Interfaces

- **SDK** — `@picsart/ai-sdk` (TypeScript, Node 20+, ESM). Type-safe `ai.generate()` / `ai.generateText()` with per-model parameter types; polls long jobs for you.
- **REST** — the API gateway at `https://api.picsart.com`. Sync `/execute`, or async `/submit` + `/{taskId}/result`.
- **CLI** — `@picsart/gen-ai` (`gen-ai`). Generate from a terminal, quote cost before spending, emit JSON for pipelines.
- **MCP** — the same models inside Claude Code, Cursor, Windsurf and other MCP hosts. Setup: https://picsart.com/gen-ai-mcp/

```bash
npm install @picsart/ai-sdk      # SDK
npm install -g @picsart/gen-ai   # CLI (also drives the MCP server)
```

## Unified API (recommended)

One parameter set, used by the SDK and by REST alike: the names below are what `ai.generate('sam-3-video', …)` takes, and what goes inside the inner `params` of a REST call. Over REST the model id travels in the body, not the path — one workflow, `v1/models`, fronts the whole catalogue.

### Execution modes

- **Sync** — `POST /workflows/v1/models/execute` returns the result in one call. Only for jobs that finish within ~20 seconds; otherwise it times out.
- **Async** — `POST /workflows/v1/models/submit` returns a task id; poll `GET /workflows/v1/models/{taskId}/result` until `response.status` is `COMPLETED` or `FAILED`. Task statuses: `ACCEPTED` · `IN_PROGRESS` · `COMPLETED` · `FAILED`.
- **SDK** — `ai.generate()` always awaits the finished result; polling is hidden.

Video jobs routinely outrun the sync limit: use async over REST, or the SDK, which polls for you.

### Input parameters

- **`prompt`** (`string`, _required_): Prompt
  - e.g. "person, cloth"
- **`videoUrl`** (`string`, _required_): Source Video
  - Accepts: video URL
- **`detectionThreshold`** (`number`, _optional_): Detection Threshold
  - Default: `0.5`
  - Range: 0–1 (step 0.05)
- **`applyMask`** (`boolean`, _optional_): Apply Mask
  - Default: `true`
- **`outputFormat`** (`string`, _optional_): Format
  - Default: `"mp4"`
  - Options: `mp4`, `webm`

### Input JSON Schema

The inner `params` object. Over REST it is wrapped as `{ "params": { "model": "<id>", "params": <this> } }`; the SDK takes it directly.

```json
{
  "type": "object",
  "required": [
    "prompt",
    "videoUrl"
  ],
  "properties": {
    "prompt": {
      "type": "string",
      "description": "Prompt"
    },
    "videoUrl": {
      "type": "string",
      "format": "uri",
      "description": "Source Video"
    },
    "detectionThreshold": {
      "type": "number",
      "minimum": 0,
      "maximum": 1,
      "multipleOf": 0.05,
      "default": 0.5,
      "description": "Detection Threshold"
    },
    "applyMask": {
      "type": "boolean",
      "default": true,
      "description": "Apply Mask"
    },
    "outputFormat": {
      "type": "string",
      "enum": [
        "mp4",
        "webm"
      ],
      "default": "mp4",
      "description": "Format"
    }
  },
  "additionalProperties": false
}
```

### Required parameters example

The required parameters plus the defaults that shape the output, as a REST body:

```json
{
  "params": {
    "model": "sam-3-video",
    "params": {
      "prompt": "A serene mountain lake at golden hour, ultra detailed",
      "videoUrl": "https://cdn.example.com/input.mp4"
    }
  }
}
```

### Response

The body is the gateway's status envelope: the output is at `response.result`, next to the task id, its status and the credits charged. Asset models return `url` plus an `items` array of results.

```json
{
  "status": "success",
  "response": {
    "id": "task_abc123",
    "status": "COMPLETED",
    "result": {
      "url": "https://cdn.picsart.com/…/result",
      "items": [
        {
          "url": "https://cdn.picsart.com/…/result"
        }
      ],
      "model": "sam-3-video"
    }
  }
}
```

What `ai.generate()` / `ai.generateText()` resolve to (the envelope already unwrapped):

```json
{
  "items": [
    {
      "url": "https://cdn.picsart.com/…/result"
    }
  ],
  "model": "sam-3-video"
}
```

### Usage examples

#### TypeScript (SDK)

```typescript
import { createClient } from '@picsart/ai-sdk';

const ai = createClient({
  apiKey: process.env.PICSART_API_KEY,
  apiUrl: 'https://api.picsart.com',
});

const result = await ai.generate('sam-3-video', {
  prompt: "A serene mountain lake at golden hour, ultra detailed",
  videoUrl: "https://cdn.example.com/input.mp4"
});

console.log(result.url); // video URL
```

#### cURL (sync)

```bash
curl -X POST 'https://api.picsart.com/workflows/v1/models/execute' \
  -H 'Authorization: Bearer $PICSART_API_KEY' \
  -H 'Content-Type: application/json' \
  -d '{
  "params": {
    "model": "sam-3-video",
    "params": {
      "prompt": "A serene mountain lake at golden hour, ultra detailed",
      "videoUrl": "https://cdn.example.com/input.mp4"
    }
  }
}'
```

#### cURL (async: submit, then poll)

```bash
# 1. Submit the job
curl -X POST 'https://api.picsart.com/workflows/v1/models/submit' \
  -H 'Authorization: Bearer $PICSART_API_KEY' \
  -H 'Content-Type: application/json' \
  -d '{
  "params": {
    "model": "sam-3-video",
    "params": {
      "prompt": "A serene mountain lake at golden hour, ultra detailed",
      "videoUrl": "https://cdn.example.com/input.mp4"
    }
  }
}'
# → { "status": "success", "response": { "id": "task_abc123" } }

# 2. Poll until response.status is COMPLETED
curl 'https://api.picsart.com/workflows/v1/models/task_abc123/result' \
  -H 'Authorization: Bearer $PICSART_API_KEY'
```

#### CLI

```bash
gen-ai login  # once per machine
gen-ai models info sam-3-video  # this model's parameters
gen-ai pricing sam-3-video  # quote before spending credits
gen-ai generate -m sam-3-video \
  -p "A serene mountain lake at golden hour"
```

## Compatibility API (vendor parameters)

Alongside the unified API, the `sam-3/video` workflow takes Meta's own parameters exactly as the vendor defines them — nothing renamed or reshaped. It is the shortest path onto Picsart for request bodies already written against the vendor; the trade is that a request is written for one vendor, and the result comes back in the vendor's own shape. Same authentication.

- **Endpoint**: `POST https://api.picsart.com/workflows/sam-3/video/execute` (sync) · `POST …/submit` + `GET …/{taskId}/result` (async)
- **SDK**: `ai.apis.run('sam-3/video', params, { mode })` — params forwarded verbatim

### Vendor parameters

| Parameter | Type | Required | Default | Details |
| --- | --- | --- | --- | --- |
| `prompt` | `string` | no | — | Text prompt for segmentation. Use commas to track multiple objects (e.g., 'person, cloth'). · e.g. `person` |
| `apply_mask` | `boolean` | no | `true` | Apply the mask on the video. |
| `text_prompt` | `string` | no | — | [DEPRECATED] Use 'prompt' instead. Kept for backward compatibility. · _deprecated_ |
| `video_url` | `string` | yes | — | The URL of the video to be segmented. |
| `box_prompts` | `BoxPromptBase[]` | no | `[]` | List of box prompt coordinates (x_min, y_min, x_max, y_max). |
| └ `x_max` | `integer` | no | — | X Max Coordinate of the box |
| └ `y_min` | `integer` | no | — | Y Min Coordinate of the box |
| └ `y_max` | `integer` | no | — | Y Max Coordinate of the box |
| └ `x_min` | `integer` | no | — | X Min Coordinate of the box |
| └ `object_id` | `integer` | no | — | Optional object identifier. Boxes sharing an object id refine the same object. |
| `video_output_type` | `string` | no | `X264 (.mp4)` | The output type of the generated video. · Options: `X264 (.mp4)`, `VP9 (.webm)` |
| `detection_threshold` | `number` | no | `0.5` | Detection confidence threshold (0.0-1.0). Lower = more detections but less precise.  · 0.1–1 |
| `point_prompts` | `PointPromptBase[]` | no | `[]` | List of point prompts |
| └ `y` | `integer` | no | — | Y Coordinate of the prompt |
| └ `x` | `integer` | no | — | X Coordinate of the prompt |
| └ `object_id` | `integer` | no | — | Optional object identifier. Prompts sharing an object id refine the same object. When a text prompt is also given, the id selects which detected object the points refine. |
| └ `label` | `integer` | no | — | 1 for foreground, 0 for background · Options: `0`, `1` |
| `options` | `object` | no | — | Platform-specific options for safety checks and Picsart Drive integration. |
| └ `safety_checks` | `object` | no | — | Safety check configuration. |
| └   `enabled` | `boolean` | no | `true` | Enable or disable text safety checks. Defaults to true. |
| └ `drive` | `object` | no | — | Picsart Drive storage options. When provided, the result will be saved to the user's Drive. |
| └   `name` | `string` | yes | — | Filename to save the result as in Picsart Drive. |
| └   `attributes` | `object` | no | — | Additional file metadata attributes. |
| └   `folder` | `object` | no | — | Destination folder in Picsart Drive. |

### Example request body

```json
{
  "params": {
    "prompt": "person",
    "video_url": "https://cdn.example.com/input.mp4"
  }
}
```

### Usage examples

#### TypeScript (`ai.apis.run`)

```typescript
import { createClient, ApiRunMode } from '@picsart/ai-sdk';

const ai = createClient({
  apiKey: process.env.PICSART_API_KEY,
  apiUrl: 'https://api.picsart.com',
});

// Calls the 'sam-3/video' workflow directly — params are sent as-is.
const { result, usage } = await ai.apis.run('sam-3/video', {
  prompt: "person",
  video_url: "https://cdn.example.com/input.mp4"
}, {
  mode: ApiRunMode.SYNC,
});

console.log(result); // workflow-specific output
console.log(usage?.credits); // amount charged, in $
```

#### cURL (sync)

```bash
curl -X POST 'https://api.picsart.com/workflows/sam-3/video/execute' \
  -H 'Authorization: Bearer $PICSART_API_KEY' \
  -H 'Content-Type: application/json' \
  -d '{
  "params": {
    "prompt": "person",
    "video_url": "https://cdn.example.com/input.mp4"
  }
}'
```

#### cURL (async: submit, then poll)

```bash
# 1. Submit the job
curl -X POST 'https://api.picsart.com/workflows/sam-3/video/submit' \
  -H 'Authorization: Bearer $PICSART_API_KEY' \
  -H 'Content-Type: application/json' \
  -d '{
  "params": {
    "prompt": "person",
    "video_url": "https://cdn.example.com/input.mp4"
  }
}'
# → { "status": "success", "response": { "id": "wf_abc123" } }

# 2. Poll until response.status is COMPLETED
curl 'https://api.picsart.com/workflows/sam-3/video/wf_abc123/result' \
  -H 'Authorization: Bearer $PICSART_API_KEY'
```

### Response

Over HTTP the vendor-shaped result arrives inside the gateway envelope at `response.result`:

```json
{
  "status": "success",
  "response": {
    "id": "wf_abc123",
    "status": "COMPLETED",
    "result": {
      "boundingbox_frames_zip": {
        "file_name": "z9RV14K95DvU.png",
        "url": "https://cdn.picsart.com/…/result.mp4",
        "content_type": "image/png",
        "file_size": 4404019
      },
      "video": {
        "file_name": "z9RV14K95DvU.png",
        "url": "https://cdn.picsart.com/…/result.mp4",
        "content_type": "image/png",
        "file_size": 4404019
      }
    }
  }
}
```

`ai.apis.run()` unwraps that envelope and resolves to `{ result, usage }`:

```json
{
  "result": {
    "boundingbox_frames_zip": {
      "file_name": "z9RV14K95DvU.png",
      "url": "https://cdn.picsart.com/…/result.mp4",
      "content_type": "image/png",
      "file_size": 4404019
    },
    "video": {
      "file_name": "z9RV14K95DvU.png",
      "url": "https://cdn.picsart.com/…/result.mp4",
      "content_type": "image/png",
      "file_size": 4404019
    }
  }
}
```

## Additional resources

- Model page: https://picsart.com/api-platform/models/sam-3-video
- Try it: https://picsart.com/api-platform/models/sam-3-video/playground
- SDK: https://picsart.com/api-platform/models/sam-3-video/sdk
- REST API (unified): https://picsart.com/api-platform/models/sam-3-video/api
- Compatibility API (vendor parameters): https://picsart.com/api-platform/models/sam-3-video/advanced
- Pricing: https://picsart.com/api-platform/models/sam-3-video/pricing
- This document: https://picsart.com/api-platform/models/sam-3-video/llms.txt
- All models: https://picsart.com/api-platform/models
- Docs — quickstart https://picsart.com/api-platform/docs/quickstart, authentication https://picsart.com/api-platform/docs/authentication, SDK https://picsart.com/api-platform/docs/sdk, REST https://picsart.com/api-platform/docs/api-reference, CLI https://picsart.com/api-platform/docs/cli, MCP https://picsart.com/api-platform/docs/mcp
- CLI reference: https://picsart.github.io/picsart-mcp-cli-docs/guide/introduction