In neural style transfer, we train the pixels of an image, and not the parameters of a network.

That's correct. In neural style transfer, we don't train the parameters of a network. Instead, we start with an input image and modify it to minimize the difference in content and style between the input image and the content and style images, respectively.

Here are the steps involved in neural style transfer:

1. Content Image: This is the image we want to transfer the style onto.

2. Style Image: This is the image we use to extract the artistic style.

3. Generated Image: This is the output image that starts as a random noise image and eventually becomes the final image with the content of the content image and the style of the style image.

4. Model: We use a pre-trained convolutional neural network. VGG19, a variant of VGG model, is commonly used.

5. Layers: We need to choose layers from the model for content and style representations. Lower layers are usually chosen for content and a combination of layers are chosen for style.

6. Loss Function: We define a loss function to measure how different the content and style of the generated image are from the content of the content image and the style of the style image. This loss function is minimized using gradient descent.

7. Optimization: We perform gradient descent to minimize the loss function. Instead of updating the weights of the network, we update the pixel values of the generated image.

So, in essence, we are training the pixels of an image to match the content of one image and the style of another.