gpu-container-nvidia-vsr

RunPod serverless container for NVIDIA Video Super Resolution upscaling.

How It Works

Calls the NVIDIA VFX SDK (nvvfx.VideoSuperRes) directly. The source video is streamed frame-by-frame with PyAV, each frame is super-resolved on the GPU, and the result is re-encoded with the hardware HEVC encoder (hevc_nvenc, falling back to libx265) and uploaded to Cloudflare R2.

Job Input

{
    "input": {
        "video_url": "https://...",
        "scale": 2,
        "quality": "HIGH"
    }
}

• video_url (required): http(s) URL of the source video.
• scale (optional, default 2): one of 1, 2, 3, 4. 1 denoises/deblurs at the same resolution.
• quality (optional, default HIGH): LOW, MEDIUM, HIGH, or ULTRA.

GPU Compatibility

Uses the NVIDIA VFX SDK, which supports datacenter GPUs (not limited to consumer RTX). Requires driver 570.190+.

GPU	Supported	Recommended
L40 / L40S	Yes	Primary target
RTX 6000 Ada	Yes	Good
A100 / H100	Yes	Overkill for upscaling
RTX 4090	Yes	Works

Quality Settings

Setting	Speed	Quality
LOW	Fastest	Basic
MEDIUM	Fast	Good
HIGH	Moderate	Very good
ULTRA	Slowest	Best

Build

docker build -t edreamai/gpu-container-nvidia-vsr:latest .

Algorithm

infinidream_algorithm: "nvidia-uprez"