I don't know if it's possible yet, but the technology behind (guided diffusion), allows it. But you will probably have to train it with a dataset that shows mul

I don't know if it's possible yet, but the technology behind (guided diffusion), allows it. But you will probably have to train it with a dataset that shows multiple associated images together. Denoising diffusions have been successfully applied to produce coherent videos. In fact, there are even some benefits to train on video because images share information between them, which make context extraction easier. Counterintuitively, because of the shared information between frames, their computations can also be partially shared, it can be almost free to generate multiple frames at the same time : A pair of RGB image can be considered as a 6D image, a 5-frame sequence a 15D image, and the neural network can learn to compress the various image channel together. The subsequent neural network layers can be kept 256D (for example) the same size as they were for a single frame. Of course this trick won't work if the pair of image are too different but you can always stack images side by side to have a bigger film-roll image and you use a standard attention mechanism to attend to various parts.