Helper function to get sorted image files from a directory.

Helper function to get sorted image files from a directory.

Downsamples and upsamples the mask to remove fine details.

Paint points on image Args: input_image: np.ndarray, (H, W, 3) input_points: np.ndarray, (N, 2), [x, y] coordinates

Using Sobel to compute gradient. Return the magnitude.

Loads all necessary models into memory. Args: base_model_path (str): Path to the base Stable Video Diffusion model.

Applies a random sequence of available augmentations.

Converts a mask to its bounding box representation.

내부에 중첩된 구멍과 도형까지 모두 완벽하게 보존하는 다각형 마스크를 생성합니다.

Applies a diverse set of temporal augmentations to randomly selected mask frames. For each frame selected for augmentation, one of the follow

Predict mask from prompts Args: prompts: Dictionary with point_coords, point_labels, mask_input prom

Runs the core inference process on a sequence of conditioning and mask frames. Args: cond_frames (list[Image.Image]): Li

Save frames as video file

Compute padding tuple.

Converts a list of PIL images to a normalized video tensor.

Resizes an image maintaining its aspect ratio. The longest side of the image is scaled to the maximum of target_width and target_height.

Encodes a video tensor into the VAE's latent space.

Introduces instability to a mask by adding salt-and-pepper noise to imitate segmentation artifacts like holes and stray pixels. Args:

Paint mask on image with transparency Args: input_image: np.ndarray, (H, W, 3) input_mask: np.ndarray, (H, W), binary ma

Track object through video using SAM2 Args: frames: List of numpy arrays, [(H,W,3)]*n, uint8 RGB frames

Converts all parts of a mask to simplified polygons, preserving all disconnected components. The level of simplification is controlled by

Applies a diverse set of temporal augmentations to randomly selected mask frames. For each frame selected for augmentation, one of the follow

Creates the additional time IDs for conditioning the UNet.

Applies geometric augmentations to a static foreground image.

Scans directories listed in the config file to find fg/bg assets.

Resizes a tensor with anti-aliasing for CLIP input, mirroring k-diffusion. This is a direct copy of the helper function from your ori

Converts a mask to a simplified polygon and back to a mask. The level of simplification is controlled by `simplification_tolerance`.

Remove output files older than max_age_minutes to prevent storage overflow This runs periodically to clean up abandoned files

Remove old output videos to save storage space

Expands the input channels of the `conv_in` layer of the UNet and updates the config. The original weights are preserved. The new weights fo

Extract frames from video file Args: video_path: Path to video file max_frames: Maximum number of frames to extract (def

Get mask for first frame only (for preview) Args: frame: np.ndarray, (H, W, 3), uint8 RGB frame poin

Initialize SAM2 and VideoMaMa models

Loads a sequence of images and corresponding masks from folders.

Load SAM2 video tracker with pretrained weights Args: device: Device to run on Returns: SAM2VideoTracker in

Load VideoMaMa pipeline with pretrained weights Args: device: Device to run on Returns: VideoInferencePipel

r""" Sets the attention processor to use to compute attention. Parameters: processor (`dict` of `AttentionProcessor` or o

Run VideoMaMa inference on video frames with mask conditioning Args: pipeline: VideoInferencePipeline instance frames_np

Generates a single synthetic video clip and its corresponding masks. This method contains the core logic from `generate_video`.

Initializes the on-the-fly dataset generator. Args: config_file (str): Path to the YAML config file specifying fg/bg sou