2.1 KiB
2.1 KiB
Upscale Module
The Upscale module raises video resolution using traditional FFmpeg scaling or AI-based Real-ESRGAN (ncnn).
Status
- AI upscaling is wired through the Real-ESRGAN ncnn backend.
- Traditional scaling is always available.
- Filters and frame rate conversion can be applied before AI upscaling.
AI Upscaling (Real-ESRGAN ncnn)
Requirements
realesrgan-ncnn-vulkaninPATH.- Vulkan-capable GPU recommended.
Pipeline
- Extract frames from the source video (filters and fps conversion applied here if enabled).
- Run
realesrgan-ncnn-vulkanon extracted frames. - Reassemble frames into a lossless MKV with the original audio.
AI Controls
- Model Preset
- General (RealESRGAN_x4plus)
- Anime/Illustration (RealESRGAN_x4plus_anime_6B)
- Anime Video (realesr-animevideov3)
- General Tiny (realesr-general-x4v3)
- 2x General (RealESRGAN_x2plus)
- Clean Restore (realesrnet-x4plus)
- Processing Preset
- Ultra Fast, Fast, Balanced (default), High Quality, Maximum Quality
- Presets tune tile size and TTA.
- Upscale Factor
- Match Target or fixed 1x/2x/3x/4x/8x.
- Output Adjustment
- Post-scale multiplier (0.5x–2.0x).
- Denoise
- Available for
realesr-general-x4v3(General Tiny).
- Available for
- Tile Size
- Auto/256/512/800.
- Output Frames
- PNG/JPG/WEBP for frame extraction.
- Advanced
- GPU selection, threads (load/proc/save), and TTA toggle.
Notes
- Face enhancement requires the Python/GFPGAN backend and is currently not executed.
- AI upscaling is heavier than traditional scaling; use smaller tiles for low VRAM.
Traditional Scaling
- Algorithms: Lanczos, Bicubic, Spline, Bilinear.
- Target: Match Source, 2x/4x, or fixed resolutions (720p → 8K).
- Output Quality: Lossless (CRF 0), Near-lossless (CRF 16, default), High (CRF 18).
Filters and Frame Rate
- Filters configured in the Filters module can be applied before upscaling.
- Frame rate conversion can be applied with or without motion interpolation.
Logging
- Each upscale job writes a conversion log in the
logs/folder next to the executable.