prev_prompt_ids.append(tokenizer(prev_prompt, truncation=True, padding="max_length", max_length=tokenizer.model_max_length, return_tensors="pt").input_ids.squeeze(0)) about storygen

[haoningwu3639]

For the first part, we aim to squeeze the 'batch' dimension, which may vary depending on the way of loading data. We will check the code again later.
And for the second part, in the second stage of training, the model needs to use the visual-language context module to calculate the diffusion features of the previous condition frames, so it may require larger GPU memory, that is, smaller batch size, but we still recommend you to use 'accelerate' that can load 'vae' and 'text_encoder' as 'fp16', which may be helpful for more efficient training.