Method interpolate

diffsynth/extensions/RIFE/__init__.py:171–197

(self, images, scale=1.0, batch_size=4, num_iter=1, progress_bar=lambda x:x)

    @torch.no_grad()
    def interpolate(self, images, scale=1.0, batch_size=4, num_iter=1, progress_bar=lambda x:x):
        # Preprocess
        processed_images = self.process_images(images)

        for iter in range(num_iter):
            # Input
            input_tensor = torch.cat((processed_images[:-1], processed_images[1:]), dim=1)

            # Interpolate
            output_tensor = []
            for batch_id in progress_bar(range(0, input_tensor.shape[0], batch_size)):
                batch_id_ = min(batch_id + batch_size, input_tensor.shape[0])
                batch_input_tensor = input_tensor[batch_id: batch_id_]
                batch_input_tensor = batch_input_tensor.to(device=self.device, dtype=self.torch_dtype)
                flow, mask, merged = self.model(batch_input_tensor, [4/scale, 2/scale, 1/scale])
                output_tensor.append(merged[2].cpu())
            
            # Output
            output_tensor = torch.concat(output_tensor, dim=0).clip(0, 1)
            processed_images = self.add_interpolated_images(processed_images, output_tensor)
            processed_images = torch.stack(processed_images)

        # To images
        output_images = self.decode_images(processed_images)
        if output_images[0].size != images[0].size:
            output_images = [image.resize(images[0].size) for image in output_images]
        return output_images


class RIFESmoother(RIFEInterpolater):