stages ¶

Pipeline stages for the native Z-Image text-to-image path.

Classes¶

Trim padded Qwen states and materialize variable-length CFG streams.

ZImageDecodingStage(vae)

Apply the official latent transform and decode one image frame.

Source code in fastvideo/pipelines/basic/zimage/stages.py

def __init__(self, vae) -> None:
    self.vae = vae

ZImageDenoisingStage(transformer, scheduler)

Run the native Z-Image flow-matching loop.

Source code in fastvideo/pipelines/basic/zimage/stages.py

def __init__(self, transformer, scheduler) -> None:
    self.transformer = transformer
    self.scheduler = scheduler

Validate the image geometry and reproduce the official device RNG.

ZImageLatentPreparationStage(transformer)

Create the official fp32 image latents on the transformer device.

Source code in fastvideo/pipelines/basic/zimage/stages.py

def __init__(self, transformer) -> None:
    self.transformer = transformer

ZImageTimestepPreparationStage(scheduler)

Apply the native scheduler's zero endpoint and discrete schedule.

Source code in fastvideo/pipelines/basic/zimage/stages.py

def __init__(self, scheduler) -> None:
    self.scheduler = scheduler