Method __call__

diffsynth/pipelines/sd_video.py:144–269

(
        self,
        prompt,
        negative_prompt="",
        cfg_scale=7.5,
        clip_skip=1,
        num_frames=None,
        input_frames=None,
        ipadapter_images=None,
        ipadapter_scale=1.0,
        controlnet_frames=None,
        denoising_strength=1.0,
        height=512,
        width=512,
        num_inference_steps=20,
        animatediff_batch_size = 16,
        animatediff_stride = 8,
        unet_batch_size = 1,
        controlnet_batch_size = 1,
        cross_frame_attention = False,
        smoother=None,
        smoother_progress_ids=[],
        tiled=False,
        tile_size=64,
        tile_stride=32,
        seed=None,
        progress_bar_cmd=tqdm,
        progress_bar_st=None,
    )

    @torch.no_grad()
    def __call__(
        self,
        prompt,
        negative_prompt="",
        cfg_scale=7.5,
        clip_skip=1,
        num_frames=None,
        input_frames=None,
        ipadapter_images=None,
        ipadapter_scale=1.0,
        controlnet_frames=None,
        denoising_strength=1.0,
        height=512,
        width=512,
        num_inference_steps=20,
        animatediff_batch_size = 16,
        animatediff_stride = 8,
        unet_batch_size = 1,
        controlnet_batch_size = 1,
        cross_frame_attention = False,
        smoother=None,
        smoother_progress_ids=[],
        tiled=False,
        tile_size=64,
        tile_stride=32,
        seed=None,
        progress_bar_cmd=tqdm,
        progress_bar_st=None,
    ):
        height, width = self.check_resize_height_width(height, width)
        
        # Tiler parameters, batch size ...
        tiler_kwargs = {"tiled": tiled, "tile_size": tile_size, "tile_stride": tile_stride}
        other_kwargs = {
            "animatediff_batch_size": animatediff_batch_size, "animatediff_stride": animatediff_stride,
            "unet_batch_size": unet_batch_size, "controlnet_batch_size": controlnet_batch_size,
            "cross_frame_attention": cross_frame_attention,
        }

        # Prepare scheduler
        self.scheduler.set_timesteps(num_inference_steps, denoising_strength)

        # Prepare latent tensors
        if self.motion_modules is None:
            noise = self.generate_noise((1, 4, height//8, width//8), seed=seed, device="cpu", dtype=self.torch_dtype).repeat(num_frames, 1, 1, 1)
        else:
            noise = self.generate_noise((num_frames, 4, height//8, width//8), seed=seed, device="cpu", dtype=self.torch_dtype)
        if input_frames is None or denoising_strength == 1.0:
            latents = noise
        else:
            latents = self.encode_video(input_frames, **tiler_kwargs)
            latents = self.scheduler.add_noise(latents, noise, timestep=self.scheduler.timesteps[0])

        # Encode prompts
        prompt_emb_posi = self.encode_prompt(prompt, clip_skip=clip_skip, positive=True)
        prompt_emb_nega = self.encode_prompt(negative_prompt, clip_skip=clip_skip, positive=False)

        # IP-Adapter