Method visual

wan/modules/clip.py:527–542

(self, videos)

            clean='whitespace')

    def visual(self, videos):
        # preprocess
        size = (self.model.image_size,) * 2
        videos = torch.cat([
            F.interpolate(
                u.transpose(0, 1),
                size=size,
                mode='bicubic',
                align_corners=False) for u in videos
        ])
        videos = self.transforms.transforms[-1](videos.mul_(0.5).add_(0.5))

        # forward
        with torch.cuda.amp.autocast(dtype=self.dtype):
            out = self.model.visual(videos, use_31_block=True)
            return out