FilterFusion
A lightweight model for real-time video denoising. It combines a low-resolution controller with a full-resolution, interpretable filter bank, achieving real-time 1080p denoising on a single CPU thread.
Timeline (last 90 days)
FilterFusion is a video denoising model with only 109K parameters that can process 1080p video at up to 55.9 FPS on a single CPU thread.
FilterFusion achieves 34.72 dB PSNR and 0.9330 SSIM on the CRVD dataset.
The FilterFusion network itself has a computational cost of 0.18 GFLOPs.
Under the conditions of 1080p resolution and a single CPU thread, FilterFusion's inference speed is 17.9 ms per frame.
FilterFusion's architecture uses a low-resolution controller plus a full-resolution filter bank, placing expensive decisions at low resolution and cheap synthesis at the original resolution.
FilterFusion's low-resolution controller explicitly injects two prior cues: temporal variance and spatial variance.
FilterFusion's filter bank includes classic operations such as Gaussian blur, median filtering, and weighted summation.
FilterFusion's synthesis path is implemented by compiling in C and then executes efficiently on the CPU, without relying on GPU acceleration or specialized inference frameworks.
FilterFusion's inference speed is 122 to 252 times faster than lightweight CNN baselines that require more than 2 seconds to process a single frame.
FilterFusion is currently validated mainly on the CRVD dataset, and whether the filter bank's performance remains robust in more complex real-world scenarios still needs further observation.
All Facts (10)
FilterFusion is a video denoising model with only 109K parameters that can process 1080p video at up to 55.9 FPS on a single CPU thread.
50%UnverifiedFilterFusion achieves 34.72 dB PSNR and 0.9330 SSIM on the CRVD dataset.
50%UnverifiedThe FilterFusion network itself has a computational cost of 0.18 GFLOPs.
50%UnverifiedUnder the conditions of 1080p resolution and a single CPU thread, FilterFusion's inference speed is 17.9 ms per frame.
50%UnverifiedFilterFusion's architecture uses a low-resolution controller plus a full-resolution filter bank, placing expensive decisions at low resolution and cheap synthesis at the original resolution.
50%UnverifiedFilterFusion's low-resolution controller explicitly injects two prior cues: temporal variance and spatial variance.
50%UnverifiedFilterFusion's filter bank includes classic operations such as Gaussian blur, median filtering, and weighted summation.
50%UnverifiedFilterFusion's synthesis path is implemented by compiling in C and then executes efficiently on the CPU, without relying on GPU acceleration or specialized inference frameworks.
50%UnverifiedFilterFusion's inference speed is 122 to 252 times faster than lightweight CNN baselines that require more than 2 seconds to process a single frame.
50%UnverifiedFilterFusion is currently validated mainly on the CRVD dataset, and whether the filter bank's performance remains robust in more complex real-world scenarios still needs further observation.
50%