🎞️ Zero2x Video Upscaler
Zero2x is a work-in-progress video upscaling tool that uses deep learning models to enhance your videos frame by frame.
This app leverages region-based difference detection to speed up processing and reduce unnecessary computation.
✨ Features
- Multiple Upscaling Models: Choose from a selection of pre-trained models for different styles and quality.
- Region-Based Upscaling: Only upscale parts of the frame that have changed, making processing faster and more memory-efficient.
- Full Frame Upscaling: Optionally upscale every frame in its entirety for maximum quality.
- Customizable Settings: Fine-tune thresholds, padding, and region detection for your specific needs.
- Progress Tracking: See estimated time remaining and per-frame progress.
- Downloadable Results: Download your upscaled video when processing is complete.
🧑🔬 Technique
This app uses the Segmented Absolute Differences (SAD) (Created by me) program to compare each frame with the previous one.
If only small regions have changed, only those regions are upscaled using the selected model.
If the whole frame is different, the entire frame is upscaled.
This hybrid approach balances speed and quality.
🚧 Work In Progress
- More models and settings will be added soon.
- Some features may be experimental or incomplete.
- Feedback and suggestions are welcome!
- The quality of the upscaled video may vary depending on the model and settings used.
Tip: If you encounter CUDA out-of-memory errors, try increasing the segment grid size or lowering the region count. Note: The reason i named this project Zero2x is because i was inspired by Video2x, but i wanted my own version with a different approach. It is running on HuggingFace's ZeroGPU hardware, which is why i came up with the name.
- Use Regions: When enabled, only changed areas between frames are upscaled. This is faster but may miss subtle changes.
- Threshold: Controls how sensitive the difference detection is. I found high values to introduce unmatching regions, be careful.
- Padding: Adds extra pixels around detected regions to include out of bounds pixels.
- Min Percentage: If the similarity between frames is above this value, only regions are upscaled; otherwise, the full frame is upscaled.
- Max Rectangles: Limits the number of regions to process per frame for performance.
- Segment Rows/Columns: Controls the grid size for region detection. More segments allow finer detection but may increase processing time. Uses less Vram when used.
- Full Frame Interval: Forces a full-frame upscale every N frames. Set to 1 to always upscale the full frame. This is to prevent regions from glitching out.
- Motion Threshold: Controls how sensitive the motion detection is. Upscaling motion frames increases faulty regions. Lower = More strict
Use regions to upscale only the different parts of the video (⚡️ Experimental, Faster)