[sb2702]

Adds two 2x models, WebSR GAN-S and WebSR GAN-M, and a new architecture, WebSR UNet.

What they are. Small UNets distilled from RealESR General x4 v3, designed for fast inference on integrated graphics cards.

Key features

1. A UNet design, which replicates the receptive field of the original RealESRGaN x4v3 model with far fewer layers (7 vs 32+), which minimizes tensor bandwidth, often the primary bottleneck on integrated graphics and apple silicon, not arithmetic density.

2. Uses conditional convolutions and weight banks to handle different kinds of content more effectively in one inference pass

Performance

Uses RealESRGANx4v3 as the teacher, the WebSR models run ~6x faster with ~39 dB of accuracy and an LPIPS score of 0.01 (largely imperceptible to the human eye). Performance on an M6 Macbook Pro:

Model	Precision	Output	FPS	Real PSNR (dB)	Real LPIPS
WebSR gan-s	fp16	2×	61.5	38.672	0.0111
WebSR gan-m	fp16	2×	42.2	39.549	0.0092
Real-ESRGAN x4v3	fp16	4×	10.4	reference	reference

Architecture

RealESRGAN:

WebSR - GAN:

Where it's coming from

These are models built for free.upscaler.video (~800k Monthly Active Users), as part of the WebSR SDK.

If you just Google "Video Upscaler", free.upscaler.video is the first result, ahead of Topaz or whatever. More info here and here.

Files

• Files: https://huggingface.co/katana-video/websr (onnx/gan-s_image.onnx, onnx/gan-m_image.onnx), MIT
• Teacher model: Real-ESRGAN realesr-general-x4v3 (BSD-3-Clause)

Let me know if you need more info