https://github.com/AsuharietYgvar/AppleNeuralHash2ONNX/issues/1#issuecomment-902977931 https://github.com/AsuharietYgvar/AppleNeuralHash2ONNX/issue... I posted

https://github.com/AsuharietYgvar/AppleNeuralHash2ONNX/issues/1#issuecomment-902977931 https://github.com/AsuharietYgvar/AppleNeuralHash2ONNX/issue... I posted some examples that look like totally normal images, they're no harder to produce, you just need to noise-shape the gradient descent so that the introduced noise has a spectrum similar to the image. E.g. just feeding back a gaussian highpassed version of the error signal is sufficient.