Method generate

wan/animate.py:297–648

r""" Generates video frames from input image using diffusion process. Args: src_root_path ('str'): Process output path replace_flag (`bool`, *optional*, defaults to False): Whether to use character replace. clip_len

(
        self,
        src_root_path,
        replace_flag=False,
        clip_len=77,
        refert_num=1,
        shift=5.0,
        sample_solver='dpm++',
        sampling_steps=20,
        guide_scale=1,
        input_prompt="",
        n_prompt="",
        seed=-1,
        offload_model=True,
    )

        return bg_images, mask_images

    def generate(
        self,
        src_root_path,
        replace_flag=False,
        clip_len=77,
        refert_num=1,
        shift=5.0,
        sample_solver='dpm++',
        sampling_steps=20,
        guide_scale=1,
        input_prompt="",
        n_prompt="",
        seed=-1,
        offload_model=True,
    ):
        r"""
        Generates video frames from input image using diffusion process.

        Args:
            src_root_path ('str'):
                Process output path
            replace_flag (`bool`, *optional*, defaults to False):
                Whether to use character replace.
            clip_len (`int`, *optional*, defaults to 77):
                How many frames to generate per clips. The number should be 4n+1
            refert_num (`int`, *optional*, defaults to 1):
                How many frames used for temporal guidance. Recommended to be 1 or 5.
            shift (`float`, *optional*, defaults to 5.0):
                Noise schedule shift parameter. 
            sample_solver (`str`, *optional*, defaults to 'dpm++'):
                Solver used to sample the video.
            sampling_steps (`int`, *optional*, defaults to 20):
                Number of diffusion sampling steps. Higher values improve quality but slow generation
            guide_scale (`float` or tuple[`float`], *optional*, defaults 1.0):
                Classifier-free guidance scale. We only use it for expression control. 
                In most cases, it's not necessary and faster generation can be achieved without it. 
                When expression adjustments are needed, you may consider using this feature.
            input_prompt (`str`):
                Text prompt for content generation. We don't recommend custom prompts (although they work)
            n_prompt (`str`, *optional*, defaults to ""):
                Negative prompt for content exclusion. If not given, use `config.sample_neg_prompt`
            seed (`int`, *optional*, defaults to -1):
                Random seed for noise generation. If -1, use random seed
            offload_model (`bool`, *optional*, defaults to True):
                If True, offloads models to CPU during generation to save VRAM

        Returns:
            torch.Tensor:
                Generated video frames tensor. Dimensions: (C, N, H, W) where:
                - C: Color channels (3 for RGB)
                - N: Number of frames
                - H: Frame height 
                - W: Frame width 
        """
        assert refert_num == 1 or refert_num == 5, "refert_num should be 1 or 5."

        seed_g = torch.Generator(device=self.device)
        seed_g.manual_seed(seed)