modelscope
Question about krea realtime
Hi, krea realtime is an autoregressive model with causal attention. Does Diffsynth implement causal attention?
@unrealMJ No. This model is quite special. Causal attention is a technique used for video continuation. Although the model itself supports generating videos of unlimited length, it suffers from noticeable visual degradation. Therefore, we have not yet implemented this capability for this model. Currently, in DiffSynth-Studio, this model is only used as an accelerated text-to-image model for inference and training.
krea realtime是通过self-forcing蒸馏得到的模型，我认为似乎不能使用flow matching的方式继续微调，而且这里微调使用的依然是双向注意力 https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/full/krea-realtime-video.sh