Remote Upscaling

Overview

Remote upscaling lets you process videos on cloud GPU instances instead of your local machine. This is especially useful for long videos, high scale factors (4x or 8x), or when your local GPU is not powerful enough. Cloud instances with modern GPUs like the RTX 3090 or RTX 4090 can process frames at 5 to 10 FPS, making even large upscaling jobs practical.

Instance Setup

Before you can upscale remotely, you need a running cloud instance of the correct type. Recaster uses Vast.ai for cloud GPU provisioning, and upscaling uses a dedicated UPSCALE instance type.

Three-Phase Workflow

Remote upscaling follows a three-phase workflow: upload, process, and download. Each phase shows progress in the interface and can be canceled at any point.

Phase 1: Upload

Your video file is uploaded to the cloud instance via rsync. The progress bar shows upload percentage in real time. Upload speed depends on your internet connection and the file size.

Phase 2: Process

The cloud instance runs the upscaling pipeline on each frame. Three standalone scripts are automatically deployed to the instance on first use:

• upscale_server.py — Main processing server with CLI (approximately 500 lines)
• swinir_upscaler.py — Standalone SwinIR implementation (approximately 320 lines)
• real_esrgan_upscaler.py — Standalone Real-ESRGAN implementation (approximately 320 lines)

These scripts have no dependencies on Recaster itself — they are standalone ONNX Runtime + OpenCV implementations that run directly on the cloud instance. Models are downloaded automatically on demand.

During processing, the remote server outputs progress messages that Recaster parses in real time:

PROGRESS:100/1000:10%     # Frame progress
PROGRESS:200/1000:20%
PROGRESS:audio:0%         # Audio muxing started
PROGRESS:audio:100%       # Audio muxing complete
RESULT:/workspace/output.mp4  # Final result path

Phase 3: Download

Once processing completes, the upscaled video is downloaded back to your local machine via rsync. The progress bar shows download percentage. The final file is saved to your chosen output path.

Enabling Remote Mode

To switch the upscaling widget to remote mode:

1. Open the Video Upscaling interface by clicking Upscale in the sidebar.
2. Toggle the Remote switch at the top of the widget. This switch is only visible with a Studio license.
3. If you have one UPSCALE instance running, it auto-selects. If you have multiple instances, a selection dialog appears.
4. If no UPSCALE instances are running, you will be prompted to launch one from the Remote panel.

Performance Estimates

The table below shows typical total processing times for common scenarios on a modern cloud GPU (RTX 3090 or RTX 4090). Times include upload, processing, and download phases.

Processing rate is approximately 5 to 10 FPS depending on the model, scale factor, and GPU. SwinIR is slightly slower than Real-ESRGAN. Multi-pass upscaling takes proportionally longer (roughly 2x for 4x scale, 3x for 8x scale).

Remote Server CLI

Recaster automatically constructs the remote command from your pipeline settings. For reference, the full set of CLI arguments supported by the remote server is listed below:

Cancellation

You can cancel remote upscaling at any point by clicking the Stop button. Recaster terminates the active worker (upload, processing, or download) with a graceful 3-second shutdown timeout. The rsync and SSH processes on the remote instance are properly terminated.

Troubleshooting