[Integrations](https://spiral.ai/integrations) › [Hermes](https://spiral.ai/integrations/hermes) › [Sa2VA 4B video segmentation](https://spiral.ai/integrations/hermes/sa2va-4b-video-segmentation)

# How to Use Sa2VA 4B video segmentation in Hermes

Sa2VA 4B returns a text answer and segmentation mask videos for a supplied clip. Use it to describe visible actions and isolate objects across video rather than in a single still image.

|  |  |
| --- | --- |
| Model | Sa2VA 4B video segmentation |
| Available through | fal |
| Category | Data |
| Price | Up to $3.12 per run at the default inputs for 60 seconds of input. |
| Returns | text |
| Runs | In the background |
| Setup | About 2 minutes, no API key |

## Prompts to try with Sa2VA 4B video segmentation in Hermes

- Each morning, use Sa2VA 4B video segmentation on newly supplied Big Buck Bunny excerpts for the animation lesson. Request an action summary and matching character masks for each excerpt.
- Every Wednesday, use Sa2VA 4B video segmentation on the supplied Blender Foundation Sintel trailer excerpts. Request scene descriptions with masks for the objects discussed and prepare an editor's review queue.
- Every Monday at 9am: use Sa2VA 4B video segmentation to look this up and give me the result as a table.
- Pull the latest numbers from Sa2VA 4B video segmentation and tell me what changed since last week.
- Check Sa2VA 4B video segmentation and summarize what you find in three bullet points.

## How to set up Sa2VA 4B video segmentation in Hermes

Pick the way that suits you. Each takes about two minutes.

### Use the Spiral skill

1. **Send Hermes this message**: Your agent reads Spiral's skill and sets itself up.

```
set up https://spiral.ai/SKILL.md
```

2. **Sign in or paste a key**: Approve Spiral in the browser, or give the agent a key from the console.

### Use Spiral's MCP server

1. **Add the server and sign in**: Hermes opens your browser: sign in to Spiral and click Authorize. No API key. When Hermes asks Enable all tools? \[Y/n/select], answer Y: n cancels without saving the server, and select hides the tools you leave out.

```bash
hermes mcp add spiral \
  --url https://api.spiral.ai/mcp \
  --auth oauth \
  --connect-timeout 300
```

2. **Check it**: In a chat that was already open, run `/reload-mcp`.

```bash
hermes mcp list
hermes mcp test spiral
```

### Use the Spiral CLI

1. **Install and sign in**: Installs the Spiral CLI, signs you in and adds its skill to your coding agents.

```bash
npx -y spiral-cli@latest init
```

### Call the Spiral API

1. **Create an API key**: In the console, under API keys.

2. **Run Sa2VA 4B video segmentation**: One request; the result comes back by polling or a webhook.

```bash
curl https://api.spiral.ai/v1/runs \
  -H "Authorization: Bearer $SPIRAL_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: $(uuidgen)" \
  -d '{
    "provider": "fal",
    "model": "fal-ai/sa2va/4b/video",
    "input": {
      "prompt": "Could you please give me a brief description of the video? Please respond with interleaved segmentation masks for the corresponding parts of the answer.",
      "video_url": "https://drive.google.com/uc?id=1iOFYbNITYwrebBBp9kaEGhBndFSRLz8k"
    }
  }'
```

Then paste one of the prompts above: Hermes finds Sa2VA 4B video segmentation on Spiral, shows the price and runs it.

Setup guide in the docs: https://spiral.ai/docs/quickstarts/hermes

## Sa2VA 4B video segmentation tools and prices

- [Sa2VA 4B Video · Vision](https://spiral.ai/tools/fal/fal-ai/sa2va/4b/video) (fal): $0.052 per second.

## Settings you can ask for

| Setting | Allowed values | Default | Description |
| --- | --- | --- | --- |
| num_frames_to_sample | 5 to 100 | — | Number of frames to sample from the video. Defaults to 32 if not provided. Minimum 5, maximum 100. |

## What can you do with Sa2VA 4B video segmentation in Hermes?

Prepare a recurring review queue of media descriptions and object masks from newly supplied assets. A scheduled agent can ask Sa2VA the same visual questions for each delivery and collect the results for an editor.

### Draft visual descriptions

Ask Sa2VA to describe a product photo or demonstration clip, then use the answer as a draft for a media caption.

### Prepare annotation drafts

Request object descriptions with corresponding masks for an image or video dataset. Give annotators the text and masks together for review.

### Catalog product media

Describe visible products, colors, and surroundings in catalog photos and demonstration videos. Use the descriptions to draft searchable asset labels.

### Prepare masks for editing

Name the object an editor needs to isolate and request its segmentation mask. Review the returned image mask or mask video before using it in an editing workflow.

### Review generated media

Ask specific questions about objects and their relationships in generated stills or clips. Compare the answers and masks with the creative brief during review.

### Create visual teaching material

Use licensed illustrations or animation clips to draft object descriptions and visual questions. Request matching masks to highlight the regions discussed in a lesson.

## Overview

- [fal](https://spiral.ai/providers/fal)
- [All Hermes guides](https://spiral.ai/integrations/hermes)

## Sa2VA 4B video segmentation in other agents

- [Amp](https://spiral.ai/integrations/amp/sa2va-4b-video-segmentation)
- [ChatGPT](https://spiral.ai/integrations/chatgpt/sa2va-4b-video-segmentation)
- [Claude](https://spiral.ai/integrations/claude-desktop/sa2va-4b-video-segmentation)
- [Claude Code](https://spiral.ai/integrations/claude-code/sa2va-4b-video-segmentation)
- [Cline](https://spiral.ai/integrations/cline/sa2va-4b-video-segmentation)
- [Codex](https://spiral.ai/integrations/codex/sa2va-4b-video-segmentation)
- [Cursor](https://spiral.ai/integrations/cursor/sa2va-4b-video-segmentation)
- [Devin](https://spiral.ai/integrations/devin/sa2va-4b-video-segmentation)
- [dots](https://spiral.ai/integrations/dots/sa2va-4b-video-segmentation)
- [Factory Droid](https://spiral.ai/integrations/factory-droid/sa2va-4b-video-segmentation)
- [Gemini](https://spiral.ai/integrations/gemini/sa2va-4b-video-segmentation)
- [Gemini CLI](https://spiral.ai/integrations/gemini-cli/sa2va-4b-video-segmentation)
- [Google Antigravity](https://spiral.ai/integrations/antigravity/sa2va-4b-video-segmentation)
- [goose](https://spiral.ai/integrations/goose/sa2va-4b-video-segmentation)
- [Grok](https://spiral.ai/integrations/grok/sa2va-4b-video-segmentation)
- [Grok Build](https://spiral.ai/integrations/grok-build/sa2va-4b-video-segmentation)
- [Instinct](https://spiral.ai/integrations/instinct/sa2va-4b-video-segmentation)
- [Manus](https://spiral.ai/integrations/manus/sa2va-4b-video-segmentation)
- [Muse](https://spiral.ai/integrations/muse/sa2va-4b-video-segmentation)
- [Muse Code](https://spiral.ai/integrations/muse-code/sa2va-4b-video-segmentation)
- [n8n](https://spiral.ai/integrations/n8n/sa2va-4b-video-segmentation)
- [Notion](https://spiral.ai/integrations/notion/sa2va-4b-video-segmentation)
- [OpenClaw](https://spiral.ai/integrations/openclaw/sa2va-4b-video-segmentation)
- [opencode](https://spiral.ai/integrations/opencode/sa2va-4b-video-segmentation)
- [Perplexity](https://spiral.ai/integrations/perplexity/sa2va-4b-video-segmentation)
- [Raycast](https://spiral.ai/integrations/raycast/sa2va-4b-video-segmentation)
- [VS Code](https://spiral.ai/integrations/vscode/sa2va-4b-video-segmentation)
- [Windsurf](https://spiral.ai/integrations/windsurf/sa2va-4b-video-segmentation)
- [Zapier](https://spiral.ai/integrations/zapier/sa2va-4b-video-segmentation)

## More Hermes guides

- [Sa2VA 4B image segmentation](https://spiral.ai/integrations/hermes/sa2va-4b-image-segmentation)
- [Sa2VA 8B image segmentation](https://spiral.ai/integrations/hermes/sa2va-8b-image-segmentation)
- [SAM 3 image embedding](https://spiral.ai/integrations/hermes/sam-3-image-embedding)
- [Sa2VA 8B video segmentation](https://spiral.ai/integrations/hermes/sa2va-8b-video-segmentation)
- [Moondream 2 object coordinates](https://spiral.ai/integrations/hermes/moondream-2-object-coordinates)
- [fal NSFW image check](https://spiral.ai/integrations/hermes/fal-nsfw-image-check)

## Questions about Sa2VA 4B video segmentation in Hermes

**Can Hermes use Sa2VA 4B video segmentation?**

Yes. Add Spiral to Hermes as an MCP server, and Hermes can find and run Sa2VA 4B video segmentation like any built-in tool.

**How much does Sa2VA 4B video segmentation cost in Hermes?**

Up to $3.12 per run at the default inputs for 60 seconds of input. You pay from your Spiral balance as you go, with no subscription, and a failed run is not charged. Your Hermes plan is separate.

**Do I need a fal account to use Sa2VA 4B video segmentation in Hermes?**

No. Spiral runs Sa2VA 4B video segmentation on its own fal account. You sign in to Spiral once and never paste a key.

**How long does it take to set up Sa2VA 4B video segmentation in Hermes?**

About two minutes: add the server and sign in, then check it.

**Does Hermes wait while Sa2VA 4B video segmentation runs?**

Sa2VA 4B video segmentation runs in the background. Hermes starts the run, checks on it and gives you the text when it's ready.

**What does Sa2VA return besides a description?**

Its documented outputs include generated text and segmentation masks. Image analysis returns image masks; video analysis returns mask video files.

**Can Sa2VA fetch media behind a login?**

Externally hosted media URLs must be publicly accessible. A link that requires a website login falls outside that documented boundary.

**How do the published 4B and 8B results compare?**

ByteDance reports MMBench scores of 81.8 for 4B and 84.4 for 8B. On the MeVIS video segmentation benchmark, it reports 55.9 and 58.9 respectively; these are benchmark scores, not accuracy percentages for your media.

**How should I ask for segmentation masks?**

Explicitly request masks for the objects discussed in the answer. ByteDance's example asks for interleaved segmentation masks corresponding to parts of the description.

**Which scenes need closer review?**

The Sa2VA paper identifies occlusion and long referring descriptions as challenges. It also calls out longer videos with camera motion, so review those results closely before using the masks as annotations.

