Turn an image into a short video with Cosmos 3 Edge on DGX Spark

Cosmos 3 Edge is NVIDIA's world model for generating or interpreting physical-world content from text, images, and actions. This guide follows one Generator path: image-to-video (I2V). Send a still image and a structured JSON prompt describing the continuation; the API returns an MP4. Reasoner requests for action prediction or video understanding are separate workloads and should not be compared with these video-generation times. Cosmos 3 Edge model card

DGX Spark is an ARM64 Linux system built around GB10. The official vLLM-Omni Cosmos3 image manifest includes linux/arm64, so this recipe selects that platform. The command binds the API only to 127.0.0.1; the request uploads the image directly. vLLM-Omni Cosmos3-Edge recipe

Prepare the NVIDIA driver and NVIDIA Container Toolkit on DGX Spark. For the default guardrails, first read the terms on the Cosmos Guardrail model page and obtain access for your own account. Then use the commands below to create a Python environment, sign in, and start the server. It is ready when the logs show Application startup complete. DGX Spark User Guide

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -U huggingface_hub requests
hf auth login
cosmos_cache="${HF_HOME:-$HOME/.cache/huggingface}"
mkdir -p "$cosmos_cache"
docker run --rm --name cosmos3-edge --platform linux/arm64 \
  --runtime nvidia --gpus all --ipc=host \
  -p 127.0.0.1:8000:8000 \
  -v "$cosmos_cache:/root/.cache/huggingface" \
  vllm/vllm-omni:cosmos3 \
  vllm serve nvidia/Cosmos3-Edge --omni --host 0.0.0.0 --port 8000 \
    --init-timeout 1800

Leave the server running and open a second terminal in the same working directory. Reactivate the environment and download only the sample image and JSON prompts. Inspecting the output video requires ffprobe, included with FFmpeg.

source .venv/bin/activate
hf download nvidia/Cosmos3-Edge --include "assets/*" --local-dir Cosmos3-Edge

Move into the downloaded sample folder and run the Python client below. It reads the model card's JSON prompt and negative prompt, then uploads the image as a multipart request to the local server. The request asks for 121 frames at 832×480 and 24 fps⁵; it writes the returned bytes to cosmos3_edge_i2v.mp4 in the current directory. I2V request format in the model card

cd Cosmos3-Edge
python - <<'PY'
import json, mimetypes, requests
from pathlib import Path
prompt = json.load(open('assets/example_i2v_prompt.json'))
negative = json.load(open('assets/negative_prompt.json'))
image = Path('assets/example_i2v_input.jpg')
data = {
    'prompt': json.dumps(prompt),
    'negative_prompt': json.dumps(negative),
    'size': '832x480',
    'num_frames': '121',
    'fps': '24',
    'num_inference_steps': '20',
    'guidance_scale': '6.0',
    'max_sequence_length': '4096',
    'flow_shift': '12.0',
    'seed': '0',
    'extra_params': json.dumps({
        'use_resolution_template': False,
        'use_duration_template': False,
        'guardrails': True,
    }),
}
mime = mimetypes.guess_type(image.name)[0] or 'image/jpeg'
with image.open('rb') as file:
    response = requests.post('http://127.0.0.1:8000/v1/videos/sync', data=data, files={'input_reference': (image.name, file, mime)}, headers={'Accept': 'video/mp4'}, timeout=1800)
response.raise_for_status()
output = Path('cosmos3_edge_i2v.mp4')
output.write_bytes(response.content)
print(f'Saved {output} ({len(response.content):,} bytes)')
PY
ffprobe -v error -select_streams v:0 -show_entries stream=width,height,nb_frames,r_frame_rate -of default=noprint_wrappers=1 cosmos3_edge_i2v.mp4

The server returns the video bytes and the client saves them. Afterward, use ffprobe to inspect the video dimensions and frame count and confirm the request produced an MP4 file.

On DGX Spark, the runtime affects how long the same image-to-video task takes. The table reports published generation of 189 frames. The shorter example above requests 121 frames, so keep the frame count alongside each timing.

DGX Spark · Generator I2V · 480p · 189 frames · one integrated platform · batch size⁶ 1. Conditions and results

If the server will not start, check docker info for the NVIDIA runtime. For a 422 response, inspect the server log and request fields, then confirm the JPEG and official JSON files are under Cosmos3-Edge/assets/. An MP4 stream reported by ffprobe confirms the output file was written.