thumbnailer-upscaler
ONNX exports of an upscaling model for the browser
1 min readupdated
This model is designed to upscale images right on the user's device using WebGPU or WebAssembly. It is based on the official Real-ESRGAN weights (v0.2.5.0 release) and exported to the ONNX format without changing the weights. It runs entirely locally in the browser without uploading images to a server. This is useful for web developers and creators of client-side apps for graphics processing.
Key features
- Runs entirely on the visitor's device via WebGPU or WebAssembly
- Two weight variants: fp16 for WebGPU and fp32 for CPU
- Upscales image resolution by 4x
- Full accuracy of the original Real-ESRGAN model without retraining
What you can do
- Embed image upscaling into the client side of a JavaScript web app
- Process photos locally in the browser without sending data to a server
- Use WebGPU acceleration with fp16 support for fast performance
Example
Input data is fed in float RGB format within the range [0, 1] with dynamic width and height, and the output is an image upscaled by 4x.
Read more
DetailsNumber of parameters: 1,213,296
Files: realesr-general-x4v3.fp16.onnx (2.4 MB) and realesr-general-x4v3.fp32.onnx (4.9 MB)
Based on Real-ESRGAN (Xintao Wang et al., ICCVW 2021)
License: BSD 3-Clause










Comments