flux_stages ¶

Bases: PipelineStage

Build CLIP pooled + T5 sequence + text_ids (and optional negative for true CFG).

Bases: PipelineStage

Unpack latents, apply VAE scaling/shift, decode to pixels (5D output B×3×1×H×W).

366
367
368

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

Bases: InputValidationStage

Require height/width divisible by 16 (VAE scale × 2 for FLUX packing).

Bases: TimestepPreparationStage

Flow Match with resolution-dependent mu from packed image sequence length.

31
32

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