Method generate

wan/text2video.py:114–271

r""" Generates video frames from text prompt using diffusion process. Args: input_prompt (`str`): Text prompt for content generation size (tupele[`int`], *optional*, defaults to (1280,720)): Controls video resolution, (width,he

(self,
                 input_prompt,
                 size=(1280, 720),
                 frame_num=81,
                 shift=5.0,
                 sample_solver='unipc',
                 sampling_steps=50,
                 guide_scale=5.0,
                 n_prompt="",
                 seed=-1,
                 offload_model=True)

        self.sample_neg_prompt = config.sample_neg_prompt

    def generate(self,
                 input_prompt,
                 size=(1280, 720),
                 frame_num=81,
                 shift=5.0,
                 sample_solver='unipc',
                 sampling_steps=50,
                 guide_scale=5.0,
                 n_prompt="",
                 seed=-1,
                 offload_model=True):
        r"""
        Generates video frames from text prompt using diffusion process.

        Args:
            input_prompt (`str`):
                Text prompt for content generation
            size (tupele[`int`], *optional*, defaults to (1280,720)):
                Controls video resolution, (width,height).
            frame_num (`int`, *optional*, defaults to 81):
                How many frames to sample from a video. The number should be 4n+1
            shift (`float`, *optional*, defaults to 5.0):
                Noise schedule shift parameter. Affects temporal dynamics
            sample_solver (`str`, *optional*, defaults to 'unipc'):
                Solver used to sample the video.
            sampling_steps (`int`, *optional*, defaults to 40):
                Number of diffusion sampling steps. Higher values improve quality but slow generation
            guide_scale (`float`, *optional*, defaults 5.0):
                Classifier-free guidance scale. Controls prompt adherence vs. creativity
            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 (81)
                - H: Frame height (from size)
                - W: Frame width from size)
        """
        # preprocess
        F = frame_num
        target_shape = (self.vae.model.z_dim, (F - 1) // self.vae_stride[0] + 1,
                        size[1] // self.vae_stride[1],
                        size[0] // self.vae_stride[2])

        seq_len = math.ceil((target_shape[2] * target_shape[3]) /
                            (self.patch_size[1] * self.patch_size[2]) *
                            target_shape[1] / self.sp_size) * self.sp_size

        if n_prompt == "":
            n_prompt = self.sample_neg_prompt
        seed = seed if seed >= 0 else random.randint(0, sys.maxsize)
        seed_g = torch.Generator(device=self.device)