AnimeVideo-v3 Updates
Released on April 24, 2022, RealESRGAN-AnimeVideo-v3 includes significant improvements:- Better naturalness: More natural-looking results that preserve the original art style
- Fewer artifacts: Reduced visual artifacts and anomalies in upscaled output
- More faithful to original colors: Better color preservation and accuracy
- Better texture restoration: Improved detail and texture recovery
- Better background restoration: Enhanced background detail and clarity
Compared Methods
RealESRGAN-AnimeVideo-v3 is compared against the following anime upscaling methods:waifu2x
- Repository: nihui/waifu2x-ncnn-vulkan
- Test parameters:
tile=0,noiselevel=2 - One of the most popular anime upscaling tools
Real-CUGAN
- Repository: bilibili/ailab Real-CUGAN
- Version tested: 20220227 release
- Test parameters:
cache_mode=0,tile=0,alpha=1 - Developed by Bilibili’s AI Lab
RealESRGAN-AnimeVideo-v3
- Our latest anime video upscaling model
- Achieves better results with faster inference speed
Visual Comparisons
You may need to zoom in or click images to compare fine details. The images shown are resized and cropped patches from the original outputs.Full resolution inputs and outputs are available on Google Drive.
More Natural Results and Better Background Restoration
RealESRGAN-AnimeVideo-v3 produces more natural-looking results with superior background detail recovery compared to waifu2x and Real-CUGAN.| Input | waifu2x | Real-CUGAN | RealESRGAN AnimeVideo-v3 |
|---|---|---|---|
 |  |  |  |
 |  |  |  |
 |  |  |  |
Fewer Artifacts and Better Detailed Textures
RealESRGAN-AnimeVideo-v3 produces cleaner results with fewer artifacts and better preservation of fine textures and details.| Input | waifu2x | Real-CUGAN | RealESRGAN AnimeVideo-v3 |
|---|---|---|---|
 |  |  |  |
 |  |  |  |
 |  |  |  |
 |  |  |  |
Additional Comparisons
More examples demonstrating RealESRGAN-AnimeVideo-v3’s superior performance across various anime content types.| Input | waifu2x | Real-CUGAN | RealESRGAN AnimeVideo-v3 |
|---|---|---|---|
 |  |  |  |
 |  |  |  |
 |  |  |  |
 |  |  |  |
Performance Benchmarks
PyTorch Implementation
Inference speed comparison on NVIDIA V100 GPU. Times represent model inference only and exclude I/O operations.All measurements are frames per second (fps) for model inference time only. I/O time is not included.
| GPU | Input Resolution | waifu2x | Real-CUGAN | RealESRGAN-AnimeVideo-v3 |
|---|---|---|---|---|
| V100 | 1921 x 1080 (Full HD) | - | 3.4 fps | 10.0 fps |
| V100 | 1280 x 720 (HD) | - | 7.2 fps | 22.6 fps |
| V100 | 640 x 480 (SD) | - | 24.4 fps | 65.9 fps |
Key Takeaways
- Full HD (1080p): RealESRGAN-AnimeVideo-v3 is ~3x faster than Real-CUGAN
- HD (720p): RealESRGAN-AnimeVideo-v3 is ~3x faster than Real-CUGAN
- SD (480p): RealESRGAN-AnimeVideo-v3 is ~2.7x faster than Real-CUGAN
ncnn Implementation
Benchmarks for the ncnn (Neural Compute) implementation are planned for future releases. The ncnn version provides:- Optimized CPU inference
- Vulkan GPU acceleration support
- Mobile device compatibility
- Lower memory footprint