Remove step distillation LoRA from the pipeline's transformer. This function subtracts the LoRA weights from the transformer, restor

Wrap VAE with tiling support for high resolution processing This implementation matches LightX2V's tiled_encode/tiled_decode exactly:

Replace pipeline VAE with tiny VAE for faster decoding. Args: pipeline: WanI2VLongPipeline instance model_type: "wan2.1"

Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with the multistep UniPC.

Decode latents with tiling for high resolution (matches LightX2V exactly) Key differences from simple implementation:

Encode video with tiling for high resolution (matches LightX2V exactly) Key differences from simple implementation:

Estimate N-D similarity transformation with or without scaling. Parameters ---------- src : (M, N) array Source coordinates.

r""" Reconstruct video tensors from patch embeddings. Args: x (List[Tensor]): List of patchified features

Apply transform to boxes Copy from nanodet/data/transform/warp.py

Function invoked when calling the pipeline for generation. Args: Examples: Returns:

Function invoked when calling the pipeline for generation. Args: Examples: Returns:

Function invoked when calling the pipeline for generation. Args: Examples: Returns:

Function invoked when calling the pipeline for generation. Args: Examples: Returns:

Args: original_scheduler: The original scheduler from the pipeline denoising_step_list: List of timesteps for step di

r""" Initialize the diffusion model backbone. Args: model_type (`str`, *optional*, defaults to 't2v'): Mo

Initialize the one euro filter.