Pipeline Settings

Pipeline Overview

Every frame passes through four stages in sequence. Each stage is optional and can be toggled independently. The pipeline executes in the following order:

Denoising

Denoising removes grain, noise, and compression artifacts from the source frame before it reaches the upscaling model. This prevents the AI model from amplifying unwanted noise. Recaster uses OpenCV Non-Local Means (NLMeans) denoising, which preserves edges while smoothing flat areas.

Strength guidelines:

• 1.0 – 2.0 — Light denoising for clean or lightly compressed sources
• 2.0 – 3.0 — Moderate denoising for typical video with compression artifacts
• 3.0 – 5.0 — Heavy denoising for noisy or grainy sources
• 5.0 – 10.0 — Very aggressive denoising for severely degraded material

Multi-Pass Upscaling

Multi-pass upscaling chains multiple 2x upscale passes to reach higher scale factors. Instead of using a single 4x model, multi-pass runs the 2x model twice in sequence. This produces significantly better detail preservation and reduces the "plastic" look that aggressive single-pass upscaling can cause.

Color Correction

AI upscaling models can sometimes shift colors slightly, especially in skin tones or saturated areas. The color correction stage uses LAB color space matching to preserve the original color palette of the source material. It adjusts the lightness and chrominance channels of the upscaled frame to match the original.

Color correction is a simple toggle (on or off) with no additional parameters. It is enabled by default in all presets except Fast.

Sharpening

The sharpening stage applies an unsharp mask to enhance edge definition in the upscaled output. This is particularly useful after denoising, which can soften edges slightly. The strength slider controls how much sharpening is applied.

Strength guidelines:

• 0.3 – 0.5 — Subtle sharpening that enhances edges without visible artifacts
• 0.5 – 1.0 — Moderate sharpening for general use
• 1.0 – 1.5 — Strong sharpening to compensate for heavy denoising
• 1.5 – 2.0 — Aggressive sharpening for severely softened material

Tile Settings

Both Real-ESRGAN and SwinIR models have an input resolution limit of 512x512 pixels. For larger frames, the upscaler automatically splits the image into overlapping tiles, processes each tile independently, and blends the results seamlessly. The tile settings control this behavior.

Audio Preservation

Recaster automatically preserves the audio track from your input video. The process works in three steps:

1. Detection — ffprobe examines the input file and identifies audio streams.
2. Processing — Video frames are upscaled independently without modifying the audio.
3. Muxing — ffmpeg combines the upscaled video with the original audio track into the final output file.

Audio preservation happens automatically and requires no configuration. If ffmpeg is not installed on your system, the upscaled video is still produced but without audio.

Preview Interval

During processing, Recaster displays live previews of upscaled frames. The preview interval controls how often these previews update. The default is every 30 frames. Lower values provide more frequent updates but add a small overhead to processing.

Settings Persistence

All pipeline settings are saved between sessions. When you close and reopen Recaster, your last-used configuration is restored automatically. You can reset to a clean state by selecting any quality preset, which overwrites all settings with the preset defaults.

Recommended Configurations

Here are starting-point configurations for common scenarios. Adjust values based on your specific source material.