turbodiffusion_i2v_pipeline

TurboDiffusion I2V (Image-to-Video) Pipeline Implementation.

This module contains an implementation of the TurboDiffusion I2V pipeline for 1-4 step image-to-video generation using rCM (recurrent Consistency Model) sampling with SLA (Sparse-Linear Attention).

Key differences from T2V: - Uses dual models (high/low noise) with boundary switching - sigma_max=200 (vs 80 for T2V) - Mask conditioning with encoded first frame

Classes

fastvideo.pipelines.basic.turbodiffusion.turbodiffusion_i2v_pipeline.TurboDiffusionI2VPipeline

TurboDiffusionI2VPipeline(*args, **kwargs)

Bases: LoRAPipeline, ComposedPipelineBase

TurboDiffusion I2V pipeline for 1-4 step image-to-video generation.

Uses RCM scheduler, SLA attention, and dual model switching for high-quality I2V generation.

Source code in fastvideo/pipelines/lora_pipeline.py

def __init__(self, *args, **kwargs) -> None:
    super().__init__(*args, **kwargs)
    self.device = get_local_torch_device()
    # build list of trainable transformers
    for transformer_name in self.trainable_transformer_names:
        if (transformer_name in self.modules and self.modules[transformer_name] is not None):
            self.trainable_transformer_modules[transformer_name] = (self.modules[transformer_name])
        # check for transformer_2 in case of Wan2.2 MoE or fake_score_transformer_2
        if transformer_name.endswith("_2"):
            raise ValueError(
                f"trainable_transformer_name override in pipelines should not include _2 suffix: {transformer_name}"
            )

        secondary_transformer_name = transformer_name + "_2"
        if (secondary_transformer_name in self.modules and self.modules[secondary_transformer_name] is not None):
            self.trainable_transformer_modules[secondary_transformer_name] = self.modules[
                secondary_transformer_name]

    logger.info(
        "trainable_transformer_modules: %s",
        self.trainable_transformer_modules.keys(),
    )

    for (
            transformer_name,
            transformer_module,
    ) in self.trainable_transformer_modules.items():
        self.exclude_lora_layers[transformer_name] = (transformer_module.config.arch_config.exclude_lora_layers)
    self.lora_target_modules = self.fastvideo_args.lora_target_modules
    self.lora_path = self.fastvideo_args.lora_path
    self.lora_nickname = self.fastvideo_args.lora_nickname
    self.training_mode = self.fastvideo_args.training_mode
    if self.training_mode and getattr(self.fastvideo_args, "lora_training", False):
        assert isinstance(self.fastvideo_args, TrainingArgs)
        if self.fastvideo_args.lora_alpha is None:
            self.fastvideo_args.lora_alpha = self.fastvideo_args.lora_rank
        self.lora_rank = self.fastvideo_args.lora_rank  # type: ignore
        self.lora_alpha = self.fastvideo_args.lora_alpha  # type: ignore
        logger.info(
            "Using LoRA training with rank %d and alpha %d",
            self.lora_rank,
            self.lora_alpha,
        )
        if self.lora_target_modules is None:
            self.lora_target_modules = [
                "q_proj",
                "k_proj",
                "v_proj",
                "o_proj",
                "to_q",
                "to_k",
                "to_v",
                "to_out",
                "to_qkv",
                "to_gate_compress",
            ]
            logger.info(
                "Using default lora_target_modules for all transformers: %s",
                self.lora_target_modules,
            )
        else:
            logger.warning(
                "Using custom lora_target_modules for all transformers, which may not be intended: %s",
                self.lora_target_modules,
            )

        self.convert_to_lora_layers()
    # Inference
    elif not self.training_mode and self.lora_path is not None:
        self.convert_to_lora_layers()
        self.set_lora_adapter(
            self.lora_nickname,  # type: ignore
            self.lora_path,
        )  # type: ignore

Functions

fastvideo.pipelines.basic.turbodiffusion.turbodiffusion_i2v_pipeline.TurboDiffusionI2VPipeline.create_pipeline_stages

create_pipeline_stages(fastvideo_args: FastVideoArgs) -> None

Set up pipeline stages with proper dependency injection.

Source code in fastvideo/pipelines/basic/turbodiffusion/turbodiffusion_i2v_pipeline.py

def create_pipeline_stages(self, fastvideo_args: FastVideoArgs) -> None:
    """Set up pipeline stages with proper dependency injection."""

    self.add_stage(stage_name="input_validation_stage", stage=InputValidationStage())

    self.add_stage(stage_name="prompt_encoding_stage",
                   stage=TextEncodingStage(
                       text_encoders=[self.get_module("text_encoder")],
                       tokenizers=[self.get_module("tokenizer")],
                   ))

    self.add_stage(stage_name="conditioning_stage", stage=ConditioningStage())

    self.add_stage(stage_name="timestep_preparation_stage",
                   stage=TimestepPreparationStage(scheduler=self.get_module("scheduler")))

    self.add_stage(stage_name="latent_preparation_stage",
                   stage=LatentPreparationStage(scheduler=self.get_module("scheduler"),
                                                transformer=self.get_module("transformer", None)))

    # I2V: Encode initial image to latent space
    self.add_stage(stage_name="image_latent_preparation_stage",
                   stage=ImageVAEEncodingStage(vae=self.get_module("vae")))

    self.add_stage(stage_name="denoising_stage",
                   stage=DenoisingStage(transformer=self.get_module("transformer"),
                                        transformer_2=self.get_module("transformer_2", None),
                                        scheduler=self.get_module("scheduler"),
                                        vae=self.get_module("vae"),
                                        pipeline=self))

    self.add_stage(stage_name="decoding_stage", stage=DecodingStage(vae=self.get_module("vae"), pipeline=self))

Functions