Skip to navigation

Running Inference

View as MarkdownOpen in Claude

VSR NIM exposes a bidirectional streaming gRPC API. The quickstart bundle includes a reference client and JSON request configurations for gRPC, RTP/UDP, and Cloud File workflows.

Run the gRPC Sample

Download the VSR compressed-video quickstart resource from NGC and extract it:

mkdir -p vsr-quickstart
ngc registry resource download-version \
nim/nvidia/vsr-nim-compressed-video-quickstart:1.0.10 \
--dest vsr-quickstart
tar xzf vsr-quickstart/vsr-nim-compressed-video-quickstart-*.tar.gz \
-C vsr-quickstart
cd vsr-quickstart/quickstart

From the extracted VSR quickstart folder, build the reference client image:

docker build -t vsr-client:1.0.0 .

The packaged quickstart includes a sample H.264 elementary stream for the bundled gRPC and RTP/UDP clients:

ls -lh media/input.h264

To use your own MP4 input instead, convert it to H.264 Annex B before updating input_video_file in config/client_config.sh:

ffmpeg -y -i my_input.mp4 -map 0:v:0 -c:v copy -bsf:v h264_mp4toannexb -an my_input.h264

Run the gRPC sample:

chmod +x vsr_client.sh
./vsr_client.sh

By default, the sample client does the following:

  1. Connects to 127.0.0.1:8001.
  2. Reads /host/media/input.h264.
  3. Reads samples/client/vsr/grpc_vsr.json.
  4. Uses the input codec, target resolution, and quality preset from the JSON file.
  5. Writes the upscaled result to output_vsr_grpc.h264.

For best results, choose an input resolution lower than the target resolution, such as 720p input for 1440p or 4K output.

RTP/UDP Sample

To use the RTP/UDP sample:

  1. Start VSR NIM with the RTP/UDP launch command in Getting Started.
  2. In config/client_config.sh, set sample_app to rtp_udp.
  3. If the client is not running on the Docker host, edit samples/client/vsr/rtp_udp_vsr.json.
  4. Run ./vsr_client.sh.

The RTP/UDP workflow is controlled by the gRPC API, but media is transferred over RTP/UDP ports. The Dockerfile-built RTP/UDP sample client is intended for H.264 and H.265 RTP streams from Ubuntu-packaged GStreamer plugins. Use the default H.264 sample config and packaged media/input.h264 for the quickstart path. AV1 RTP requires rtpav1pay and rtpav1depay GStreamer Rust RTP elements, which are not available from stock Ubuntu 24.04 packages.

Cloud File Sample

To use the Cloud File sample:

  1. Set sample_app to cloud_file in config/client_config.sh.

  2. Upload an input video to object storage.

  3. Create a readable presigned URL for the input object and a writable presigned URL for the output object.

  4. Export both URLs and run ./vsr_client.sh.

    export INPUT_PRESIGNED_URL='https://storage.example.com/input.mp4?presigned-query'
    export OUTPUT_PRESIGNED_URL='https://storage.example.com/output.mp4?presigned-query'
    ./vsr_client.sh

    The wrapper passes these values to cloud_file_vsr_client as --input_presigned_url and --output_presigned_url, overriding the placeholder URIs in samples/client/vsr/cloud_file_vsr.json.

For container output, use an input container supported by the Support Matrix. If you upload the packaged media/input.h264, use an elementary-bitstream output object such as output.h264.

Build a Custom gRPC Client

Protocol buffer definitions live in samples/protos/video_processing/. Generate language stubs with your own protoc toolchain, and invoke nvidia.video_processing.v1beta1.VideoSuperResolutionService/ProcessVideo against the VSR NIM gRPC endpoint.

For example, to generate Python stubs with grpcio-tools:

python3 -m pip install grpcio grpcio-tools
python3 -m grpc_tools.protoc \
-I samples \
--python_out=. --grpc_python_out=. \
samples/protos/video_processing/vsr.proto \
samples/protos/video_processing/common.proto

API Summary

The VSR gRPC API uses the following:

FieldValue
Packagenvidia.video_processing.v1beta1
ServiceVideoSuperResolutionService
MethodProcessVideo(stream StreamingVSRRequest) returns (stream StreamingVSRResponse)
Proto filessamples/protos/video_processing/

The gRPC, RTP/UDP, and Cloud File examples all use the same VSR request configuration model.