Class LTXVTiledVAEDecode

tiled_vae_decode.py:11–211

    name="LTXVTiledVAEDecode",
)
class LTXVTiledVAEDecode:

    @classmethod
    def INPUT_TYPES(s):
        return {
            "required": {
                "vae": ("VAE",),
                "latents": ("LATENT",),
                "horizontal_tiles": ("INT", {"default": 1, "min": 1, "max": 6}),
                "vertical_tiles": ("INT", {"default": 1, "min": 1, "max": 6}),
                "overlap": ("INT", {"default": 1, "min": 1, "max": 8}),
                "last_frame_fix": ("BOOLEAN", {"default": False}),
            },
            "optional": {
                "working_device": (["cpu", "auto"], {"default": "auto"}),
                "working_dtype": (["float16", "float32", "auto"], {"default": "auto"}),
            },
        }

    RETURN_TYPES = ("IMAGE",)
    RETURN_NAMES = ("image",)

    FUNCTION = "decode"

    CATEGORY = "latent"

    def decode(
        self,
        vae,
        latents,
        horizontal_tiles,
        vertical_tiles,
        overlap,
        last_frame_fix,
        working_device="auto",
        working_dtype="auto",
    ):
        # Get the latent samples
        samples = latents["samples"]

        if last_frame_fix:
            # Repeat the last frame along dimension 2 (frames)
            # samples: [batch, channels, frames, height, width]
            last_frame = samples[
                :, :, -1:, :, :
            ]  # shape: [batch, channels, 1, height, width]
            samples = torch.cat([samples, last_frame], dim=2)

        batch, channels, frames, height, width = samples.shape
        time_scale_factor, width_scale_factor, height_scale_factor = (
            vae.downscale_index_formula
        )
        image_frames = 1 + (frames - 1) * time_scale_factor

        # Calculate output image dimensions
        output_height = height * height_scale_factor
        output_width = width * width_scale_factor