used to OOM on my iPhone. This time Opus suggested to skip onnxruntime all together and wrote the whole LaMa model in custom WebGPU shaders.
-> 5x faster and works in <1s on my 5 year old iphone 13 mini AND on android (not even tested last time, yes I bought a phone just for this). All in-browser, the photo never leaves the device.
posted this before: https://news.ycombinator.com/item?id=49120119
used to OOM on my iPhone. This time Opus suggested to skip onnxruntime all together and wrote the whole LaMa model in custom WebGPU shaders.
-> 5x faster and works in <1s on my 5 year old iphone 13 mini AND on android (not even tested last time, yes I bought a phone just for this). All in-browser, the photo never leaves the device.