Maybe there will be a way to visualize memories and dreams, by using Electroencephalography (EEG) and Neural Networks. So you can see what others think. Or see what others see, through their eyes.
Because it resumes generation from a few frames it will lose context. Imagine generating a paragraph and then the next one using only the last word you generated. Luckily images captured a lot of information so it's not that obvious. But for example you can't do a video that looks around 360 degrees is it's generated with two iterations. Very dreamlike.
Thank you for your video, great content as always! One question: in the video, you say that the video encoder is auto-regressive, so that it can be used on arbitrary number of video patches. But aren't standard transformer encoders already able to process inputs of arbitrary length? Usually the auto-regressive architecture is used in the decoder, because at inference time, we need it to generate the output causally. Am I missing something?
Thanks for this great question. Transformer sequence length is an interesting topic, which we've discussed here already: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-Xxts1ithupI.html Basically, even if it can generate / take in variable length input, it still has a predefined maximum input / output length due to practical limitations (compute time and memory). You are asking whether a causal model could not generate infinitely long video and -- for practical reasons -- the answer is no. Unmodified causal attention means that one attends to the whole generated past and for very long sequences. This means that the attention window increases linearly and computation time and memory increases quadratically. So because of limited compute time and memory, we cannot generate indefinitely, unless one applies such tricks as the Phenki authors with MaskGIT, to only attend to a small fraction of the tokens of the past generated output.
You are right, it is not a diffusion model. It's about content generation. 😅 I was more comfortable with it being in this playlist (especially as the last video in the row) rather than being nowhere close to it's fellow competition. But sure, I do not have the Paella video in the list, although Paella can be argued to be a diffusion model. I need to clean up.
I just wonder to the implementation level, these padding values as well as masking the tokens, did someone decide that we will fill these tensors with 0s? Does it matter what we are going to fill those vectors with? What if these padded/masked values of 0s overlap with actual data, how do we effectively instruct the model to disentangle masked values from 0s corresponding to the actual data?