modular_blocks_krea2.py

# Copyright 2026 Krea AI and The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import torch

from diffusers.utils import logging
from diffusers.modular_pipelines.modular_pipeline import AutoPipelineBlocks, SequentialPipelineBlocks
from diffusers.modular_pipelines.modular_pipeline_utils import InputParam, InsertableDict, OutputParam
from .before_denoise import (
    Krea2CreateMaskLatentsStep,
    Krea2PrepareLatentsStep,
    Krea2PrepareLatentsWithStrengthStep,
    Krea2RoPEInputsStep,
    Krea2SetTimestepsStep,
    Krea2SetTimestepsWithStrengthStep,
)
from .decoders import (
    Krea2AfterDenoiseStep,
    Krea2DecoderStep,
    Krea2InpaintProcessImagesOutputStep,
    Krea2ProcessImagesOutputStep,
)
from .denoise import (
    Krea2DenoiseStep,
    Krea2InpaintDenoiseStep,
)
from .encoders import (
    Krea2InpaintProcessImagesInputStep,
    Krea2ProcessImagesInputStep,
    Krea2TextEncoderStep,
    Krea2VaeEncoderStep,
)
from .inputs import (
    Krea2AdditionalInputsStep,
    Krea2TextInputsStep,
)


logger = logging.get_logger(__name__)


# ====================
# 1. TEXT ENCODER
# ====================


# auto_docstring
class Krea2AutoTextEncoderStep(AutoPipelineBlocks):
    """
    Text encoder step that encodes the text prompt into a text embedding. This is an auto pipeline block.
       - `Krea2TextEncoderStep` (text_encoder) is used when `prompt` is provided.
       - if `prompt` is not provided, step will be skipped.

      Components:
          text_encoder (`Qwen3VLModel`): The text encoder to use tokenizer (`Qwen2Tokenizer`): The tokenizer to use
          guider (`ClassifierFreeGuidance`)

      Inputs:
          prompt (`str`, *optional*):
              The prompt or prompts to guide image generation.
          negative_prompt (`str`, *optional*):
              The prompt or prompts not to guide the image generation.
          max_sequence_length (`int`, *optional*, defaults to 512):
              Maximum sequence length for prompt encoding.

      Outputs:
          prompt_embeds (`Tensor`):
              The prompt embeddings.
          prompt_embeds_mask (`Tensor`):
              The encoder attention mask.
          negative_prompt_embeds (`Tensor`):
              The negative prompt embeddings.
          negative_prompt_embeds_mask (`Tensor`):
              The negative prompt embeddings mask.
    """

    model_name = "krea2"
    block_classes = [Krea2TextEncoderStep()]
    block_names = ["text_encoder"]
    block_trigger_inputs = ["prompt"]

    @property
    def description(self) -> str:
        return (
            "Text encoder step that encodes the text prompt into a text embedding. This is an auto pipeline block.\n"
            " - `Krea2TextEncoderStep` (text_encoder) is used when `prompt` is provided.\n"
            " - if `prompt` is not provided, step will be skipped."
        )


# ====================
# 2. VAE ENCODER
# ====================


# auto_docstring
class Krea2InpaintVaeEncoderStep(SequentialPipelineBlocks):
    """
    This step is used for processing image and mask inputs for inpainting tasks. It:
       - Resizes the image to the target size, based on `height` and `width`.
       - Processes and updates `image` and `mask_image`.
       - Creates `image_latents`.

      Components:
          image_mask_processor (`InpaintProcessor`) vae (`AutoencoderKLQwenImage`)

      Inputs:
          mask_image (`Image`):
              Mask image for inpainting.
          image (`Image | list`):
              Reference image(s) for denoising. Can be a single image or list of images.
          height (`int`, *optional*):
              The height in pixels of the generated image.
          width (`int`, *optional*):
              The width in pixels of the generated image.
          padding_mask_crop (`int`, *optional*):
              Padding for mask cropping in inpainting.
          generator (`Generator`, *optional*):
              Torch generator for deterministic generation.

      Outputs:
          processed_image (`Tensor`):
              The processed image
          processed_mask_image (`Tensor`):
              The processed mask image
          mask_overlay_kwargs (`dict`):
              The kwargs for the postprocess step to apply the mask overlay
          image_latents (`Tensor`):
              The latent representation of the input image.
    """

    model_name = "krea2"
    block_classes = [Krea2InpaintProcessImagesInputStep(), Krea2VaeEncoderStep()]
    block_names = ["preprocess", "encode"]

    @property
    def description(self) -> str:
        return (
            "This step is used for processing image and mask inputs for inpainting tasks. It:\n"
            " - Resizes the image to the target size, based on `height` and `width`.\n"
            " - Processes and updates `image` and `mask_image`.\n"
            " - Creates `image_latents`."
        )


# auto_docstring
class Krea2Img2ImgVaeEncoderStep(SequentialPipelineBlocks):
    """
    Vae encoder step that preprocess and encode the image inputs into their latent representations.

      Components:
          image_processor (`VaeImageProcessor`) vae (`AutoencoderKLQwenImage`)

      Inputs:
          image (`Image | list`):
              Reference image(s) for denoising. Can be a single image or list of images.
          height (`int`, *optional*):
              The height in pixels of the generated image.
          width (`int`, *optional*):
              The width in pixels of the generated image.
          generator (`Generator`, *optional*):
              Torch generator for deterministic generation.

      Outputs:
          processed_image (`Tensor`):
              The processed image
          image_latents (`Tensor`):
              The latent representation of the input image.
    """

    model_name = "krea2"

    block_classes = [Krea2ProcessImagesInputStep(), Krea2VaeEncoderStep()]
    block_names = ["preprocess", "encode"]

    @property
    def description(self) -> str:
        return "Vae encoder step that preprocess and encode the image inputs into their latent representations."


class Krea2AutoVaeEncoderStep(AutoPipelineBlocks):
    model_name = "krea2"
    block_classes = [Krea2InpaintVaeEncoderStep, Krea2Img2ImgVaeEncoderStep]
    block_names = ["inpaint", "img2img"]
    block_trigger_inputs = ["mask_image", "image"]

    @property
    def description(self):
        return (
            "Vae encoder step that encode the image inputs into their latent representations.\n"
            + "This is an auto pipeline block.\n"
            + " - `Krea2InpaintVaeEncoderStep` (inpaint) is used when `mask_image` is provided.\n"
            + " - `Krea2Img2ImgVaeEncoderStep` (img2img) is used when `image` is provided.\n"
            + " - if `mask_image` or `image` is not provided, step will be skipped."
        )


# ====================
# 3. DENOISE (input -> prepare_latents -> set_timesteps -> prepare_rope_inputs -> denoise -> after_denoise)
# ====================


# assemble input steps
# auto_docstring
class Krea2Img2ImgInputStep(SequentialPipelineBlocks):
    """
    Input step that prepares the inputs for the img2img denoising step. It:
       - make sure the text embeddings have consistent batch size as well as the additional inputs (`image_latents`).
       - update height/width based `image_latents`, patchify `image_latents`.

      Components:
          pachifier (`Krea2Pachifier`)

      Inputs:
          num_images_per_prompt (`int`, *optional*, defaults to 1):
              The number of images to generate per prompt.
          prompt_embeds (`Tensor`):
              text embeddings used to guide the image generation. Can be generated from text_encoder step.
          prompt_embeds_mask (`Tensor`):
              mask for the text embeddings. Can be generated from text_encoder step.
          negative_prompt_embeds (`Tensor`, *optional*):
              negative text embeddings used to guide the image generation. Can be generated from text_encoder step.
          negative_prompt_embeds_mask (`Tensor`, *optional*):
              mask for the negative text embeddings. Can be generated from text_encoder step.
          height (`int`, *optional*):
              The height in pixels of the generated image.
          width (`int`, *optional*):
              The width in pixels of the generated image.
          image_latents (`Tensor`):
              image latents used to guide the image generation. Can be generated from vae_encoder step.

      Outputs:
          batch_size (`int`):
              The batch size of the prompt embeddings
          dtype (`dtype`):
              The data type of the prompt embeddings
          prompt_embeds (`Tensor`):
              The prompt embeddings. (batch-expanded)
          prompt_embeds_mask (`Tensor`):
              The encoder attention mask. (batch-expanded)
          negative_prompt_embeds (`Tensor`):
              The negative prompt embeddings. (batch-expanded)
          negative_prompt_embeds_mask (`Tensor`):
              The negative prompt embeddings mask. (batch-expanded)
          image_height (`int`):
              The image height calculated from the image latents dimension
          image_width (`int`):
              The image width calculated from the image latents dimension
          height (`int`):
              if not provided, updated to image height
          width (`int`):
              if not provided, updated to image width
          image_latents (`Tensor`):
              image latents used to guide the image generation. Can be generated from vae_encoder step. (patchified and
              batch-expanded)
    """

    model_name = "krea2"
    block_classes = [Krea2TextInputsStep(), Krea2AdditionalInputsStep()]
    block_names = ["text_inputs", "additional_inputs"]

    @property
    def description(self):
        return (
            "Input step that prepares the inputs for the img2img denoising step. It:\n"
            " - make sure the text embeddings have consistent batch size as well as the additional inputs (`image_latents`).\n"
            " - update height/width based `image_latents`, patchify `image_latents`."
        )


# auto_docstring
class Krea2InpaintInputStep(SequentialPipelineBlocks):
    """
    Input step that prepares the inputs for the inpainting denoising step. It:
       - make sure the text embeddings have consistent batch size as well as the additional inputs (`image_latents` and
         `processed_mask_image`).
       - update height/width based `image_latents`, patchify `image_latents`.

      Components:
          pachifier (`Krea2Pachifier`)

      Inputs:
          num_images_per_prompt (`int`, *optional*, defaults to 1):
              The number of images to generate per prompt.
          prompt_embeds (`Tensor`):
              text embeddings used to guide the image generation. Can be generated from text_encoder step.
          prompt_embeds_mask (`Tensor`):
              mask for the text embeddings. Can be generated from text_encoder step.
          negative_prompt_embeds (`Tensor`, *optional*):
              negative text embeddings used to guide the image generation. Can be generated from text_encoder step.
          negative_prompt_embeds_mask (`Tensor`, *optional*):
              mask for the negative text embeddings. Can be generated from text_encoder step.
          height (`int`, *optional*):
              The height in pixels of the generated image.
          width (`int`, *optional*):
              The width in pixels of the generated image.
          image_latents (`Tensor`, *optional*):
              image latents used to guide the image generation. Can be generated from vae_encoder step.
          processed_mask_image (`Tensor`, *optional*):
              The processed mask image

      Outputs:
          batch_size (`int`):
              The batch size of the prompt embeddings
          dtype (`dtype`):
              The data type of the prompt embeddings
          prompt_embeds (`Tensor`):
              The prompt embeddings. (batch-expanded)
          prompt_embeds_mask (`Tensor`):
              The encoder attention mask. (batch-expanded)
          negative_prompt_embeds (`Tensor`):
              The negative prompt embeddings. (batch-expanded)
          negative_prompt_embeds_mask (`Tensor`):
              The negative prompt embeddings mask. (batch-expanded)
          image_height (`int`):
              The image height calculated from the image latents dimension
          image_width (`int`):
              The image width calculated from the image latents dimension
          height (`int`):
              if not provided, updated to image height
          width (`int`):
              if not provided, updated to image width
          image_latents (`Tensor`):
              image latents used to guide the image generation. Can be generated from vae_encoder step. (patchified and
              batch-expanded)
          processed_mask_image (`Tensor`):
              The processed mask image (batch-expanded)
    """

    model_name = "krea2"
    block_classes = [
        Krea2TextInputsStep(),
        Krea2AdditionalInputsStep(
            additional_batch_inputs=[
                InputParam(name="processed_mask_image", type_hint=torch.Tensor, description="The processed mask image")
            ]
        ),
    ]
    block_names = ["text_inputs", "additional_inputs"]

    @property
    def description(self):
        return (
            "Input step that prepares the inputs for the inpainting denoising step. It:\n"
            " - make sure the text embeddings have consistent batch size as well as the additional inputs (`image_latents` and `processed_mask_image`).\n"
            " - update height/width based `image_latents`, patchify `image_latents`."
        )


# assemble prepare latents steps
# auto_docstring
class Krea2InpaintPrepareLatentsStep(SequentialPipelineBlocks):
    """
    This step prepares the latents/image_latents and mask inputs for the inpainting denoising step. It:
       - Add noise to the image latents to create the latents input for the denoiser.
       - Create the patchified latents `mask` based on the processed mask image.

      Components:
          scheduler (`FlowMatchEulerDiscreteScheduler`) pachifier (`Krea2Pachifier`)

      Inputs:
          latents (`Tensor`):
              The initial random noised, can be generated in prepare latent step.
          image_latents (`Tensor`):
              image latents used to guide the image generation. Can be generated from vae_encoder step. (Can be
              generated from vae encoder and updated in input step.)
          timesteps (`Tensor`):
              The timesteps to use for the denoising process. Can be generated in set_timesteps step.
          processed_mask_image (`Tensor`):
              The processed mask to use for the inpainting process.
          height (`int`):
              The height in pixels of the generated image.
          width (`int`):
              The width in pixels of the generated image.
          dtype (`dtype`, *optional*, defaults to torch.float32):
              The dtype of the model inputs, can be generated in input step.

      Outputs:
          initial_noise (`Tensor`):
              The initial random noised used for inpainting denoising.
          latents (`Tensor`):
              The scaled noisy latents to use for inpainting/image-to-image denoising.
          mask (`Tensor`):
              The mask to use for the inpainting process.
    """

    model_name = "krea2"
    block_classes = [Krea2PrepareLatentsWithStrengthStep(), Krea2CreateMaskLatentsStep()]
    block_names = ["add_noise_to_latents", "create_mask_latents"]

    @property
    def description(self) -> str:
        return (
            "This step prepares the latents/image_latents and mask inputs for the inpainting denoising step. It:\n"
            " - Add noise to the image latents to create the latents input for the denoiser.\n"
            " - Create the patchified latents `mask` based on the processed mask image.\n"
        )


# assemble denoising steps


# Krea 2 (text2image)
# auto_docstring
class Krea2CoreDenoiseStep(SequentialPipelineBlocks):
    """
    step that denoise noise into image for text2image task. It includes the denoise loop, as well as prepare the inputs
    (timesteps, latents, rope inputs etc.).

      Components:
          pachifier (`Krea2Pachifier`) scheduler (`FlowMatchEulerDiscreteScheduler`) guider (`ClassifierFreeGuidance`)
          transformer (`Krea2Transformer2DModel`)

      Inputs:
          num_images_per_prompt (`int`, *optional*, defaults to 1):
              The number of images to generate per prompt.
          prompt_embeds (`Tensor`):
              text embeddings used to guide the image generation. Can be generated from text_encoder step.
          prompt_embeds_mask (`Tensor`):
              mask for the text embeddings. Can be generated from text_encoder step.
          negative_prompt_embeds (`Tensor`, *optional*):
              negative text embeddings used to guide the image generation. Can be generated from text_encoder step.
          negative_prompt_embeds_mask (`Tensor`, *optional*):
              mask for the negative text embeddings. Can be generated from text_encoder step.
          latents (`Tensor`, *optional*):
              Pre-generated noisy latents for image generation.
          height (`int`, *optional*):
              The height in pixels of the generated image.
          width (`int`, *optional*):
              The width in pixels of the generated image.
          generator (`Generator`, *optional*):
              Torch generator for deterministic generation.
          num_inference_steps (`int`, *optional*, defaults to 28):
              The number of denoising steps.
          sigmas (`list`, *optional*):
              Custom sigmas for the denoising process.
          mu (`float`, *optional*):
              Fixed timestep shift for the scheduler. Pass `1.15` for the few-step distilled (TDM/turbo) checkpoint; if
              not provided, computed from the image sequence length (base checkpoint behavior).
          **denoiser_input_fields (`None`, *optional*):
              conditional model inputs for the denoiser: e.g. prompt_embeds, negative_prompt_embeds, etc.

      Outputs:
          latents (`Tensor`):
              Denoised latents.
    """

    model_name = "krea2"
    block_classes = [
        Krea2TextInputsStep(),
        Krea2PrepareLatentsStep(),
        Krea2SetTimestepsStep(),
        Krea2RoPEInputsStep(),
        Krea2DenoiseStep(),
        Krea2AfterDenoiseStep(),
    ]
    block_names = [
        "input",
        "prepare_latents",
        "set_timesteps",
        "prepare_rope_inputs",
        "denoise",
        "after_denoise",
    ]

    @property
    def description(self):
        return "step that denoise noise into image for text2image task. It includes the denoise loop, as well as prepare the inputs (timesteps, latents, rope inputs etc.)."

    @property
    def outputs(self):
        return [
            OutputParam.template("latents"),
        ]


# Krea 2 (inpainting)
# auto_docstring
class Krea2InpaintCoreDenoiseStep(SequentialPipelineBlocks):
    """
    step that denoise noise into image for inpaint task. It includes the denoise loop, as well as prepare the inputs
    (timesteps, latents, rope inputs etc.).

      Components:
          pachifier (`Krea2Pachifier`) scheduler (`FlowMatchEulerDiscreteScheduler`) guider (`ClassifierFreeGuidance`)
          transformer (`Krea2Transformer2DModel`)

      Inputs:
          num_images_per_prompt (`int`, *optional*, defaults to 1):
              The number of images to generate per prompt.
          prompt_embeds (`Tensor`):
              text embeddings used to guide the image generation. Can be generated from text_encoder step.
          prompt_embeds_mask (`Tensor`):
              mask for the text embeddings. Can be generated from text_encoder step.
          negative_prompt_embeds (`Tensor`, *optional*):
              negative text embeddings used to guide the image generation. Can be generated from text_encoder step.
          negative_prompt_embeds_mask (`Tensor`, *optional*):
              mask for the negative text embeddings. Can be generated from text_encoder step.
          height (`int`, *optional*):
              The height in pixels of the generated image.
          width (`int`, *optional*):
              The width in pixels of the generated image.
          image_latents (`Tensor`, *optional*):
              image latents used to guide the image generation. Can be generated from vae_encoder step.
          processed_mask_image (`Tensor`, *optional*):
              The processed mask image
          latents (`Tensor`, *optional*):
              Pre-generated noisy latents for image generation.
          generator (`Generator`, *optional*):
              Torch generator for deterministic generation.
          num_inference_steps (`int`, *optional*, defaults to 28):
              The number of denoising steps.
          sigmas (`list`, *optional*):
              Custom sigmas for the denoising process.
          mu (`float`, *optional*):
              Fixed timestep shift for the scheduler. Pass `1.15` for the few-step distilled (TDM/turbo) checkpoint; if
              not provided, computed from the image sequence length (base checkpoint behavior).
          strength (`float`, *optional*, defaults to 0.9):
              Strength for img2img/inpainting.
          **denoiser_input_fields (`None`, *optional*):
              conditional model inputs for the denoiser: e.g. prompt_embeds, negative_prompt_embeds, etc.

      Outputs:
          latents (`Tensor`):
              Denoised latents.
    """

    model_name = "krea2"
    block_classes = [
        Krea2InpaintInputStep(),
        Krea2PrepareLatentsStep(),
        Krea2SetTimestepsWithStrengthStep(),
        Krea2InpaintPrepareLatentsStep(),
        Krea2RoPEInputsStep(),
        Krea2InpaintDenoiseStep(),
        Krea2AfterDenoiseStep(),
    ]
    block_names = [
        "input",
        "prepare_latents",
        "set_timesteps",
        "prepare_inpaint_latents",
        "prepare_rope_inputs",
        "denoise",
        "after_denoise",
    ]

    @property
    def description(self):
        return "step that denoise noise into image for inpaint task. It includes the denoise loop, as well as prepare the inputs (timesteps, latents, rope inputs etc.)."

    @property
    def outputs(self):
        return [
            OutputParam.template("latents"),
        ]


# Krea 2 (image2image)
# auto_docstring
class Krea2Img2ImgCoreDenoiseStep(SequentialPipelineBlocks):
    """
    step that denoise noise into image for img2img task. It includes the denoise loop, as well as prepare the inputs
    (timesteps, latents, rope inputs etc.).

      Components:
          pachifier (`Krea2Pachifier`) scheduler (`FlowMatchEulerDiscreteScheduler`) guider (`ClassifierFreeGuidance`)
          transformer (`Krea2Transformer2DModel`)

      Inputs:
          num_images_per_prompt (`int`, *optional*, defaults to 1):
              The number of images to generate per prompt.
          prompt_embeds (`Tensor`):
              text embeddings used to guide the image generation. Can be generated from text_encoder step.
          prompt_embeds_mask (`Tensor`):
              mask for the text embeddings. Can be generated from text_encoder step.
          negative_prompt_embeds (`Tensor`, *optional*):
              negative text embeddings used to guide the image generation. Can be generated from text_encoder step.
          negative_prompt_embeds_mask (`Tensor`, *optional*):
              mask for the negative text embeddings. Can be generated from text_encoder step.
          height (`int`, *optional*):
              The height in pixels of the generated image.
          width (`int`, *optional*):
              The width in pixels of the generated image.
          image_latents (`Tensor`):
              image latents used to guide the image generation. Can be generated from vae_encoder step.
          latents (`Tensor`, *optional*):
              Pre-generated noisy latents for image generation.
          generator (`Generator`, *optional*):
              Torch generator for deterministic generation.
          num_inference_steps (`int`, *optional*, defaults to 28):
              The number of denoising steps.
          sigmas (`list`, *optional*):
              Custom sigmas for the denoising process.
          mu (`float`, *optional*):
              Fixed timestep shift for the scheduler. Pass `1.15` for the few-step distilled (TDM/turbo) checkpoint; if
              not provided, computed from the image sequence length (base checkpoint behavior).
          strength (`float`, *optional*, defaults to 0.9):
              Strength for img2img/inpainting.
          **denoiser_input_fields (`None`, *optional*):
              conditional model inputs for the denoiser: e.g. prompt_embeds, negative_prompt_embeds, etc.

      Outputs:
          latents (`Tensor`):
              Denoised latents.
    """

    model_name = "krea2"
    block_classes = [
        Krea2Img2ImgInputStep(),
        Krea2PrepareLatentsStep(),
        Krea2SetTimestepsWithStrengthStep(),
        Krea2PrepareLatentsWithStrengthStep(),
        Krea2RoPEInputsStep(),
        Krea2DenoiseStep(),
        Krea2AfterDenoiseStep(),
    ]
    block_names = [
        "input",
        "prepare_latents",
        "set_timesteps",
        "prepare_img2img_latents",
        "prepare_rope_inputs",
        "denoise",
        "after_denoise",
    ]

    @property
    def description(self):
        return "step that denoise noise into image for img2img task. It includes the denoise loop, as well as prepare the inputs (timesteps, latents, rope inputs etc.)."

    @property
    def outputs(self):
        return [
            OutputParam.template("latents"),
        ]


# Auto denoise step for Krea 2
class Krea2AutoCoreDenoiseStep(AutoPipelineBlocks):
    model_name = "krea2"
    block_classes = [
        Krea2InpaintCoreDenoiseStep,
        Krea2Img2ImgCoreDenoiseStep,
        Krea2CoreDenoiseStep,
    ]
    block_names = [
        "inpaint",
        "img2img",
        "text2image",
    ]
    block_trigger_inputs = ["processed_mask_image", "image_latents", None]

    @property
    def description(self):
        return (
            "Core step that performs the denoising process. \n"
            + " - `Krea2InpaintCoreDenoiseStep` (inpaint) is used when `processed_mask_image` is provided.\n"
            + " - `Krea2Img2ImgCoreDenoiseStep` (img2img) is used when `image_latents` is provided.\n"
            + " - `Krea2CoreDenoiseStep` (text2image) is used otherwise.\n"
            + "This step support text-to-image, image-to-image, and inpainting tasks for Krea 2:\n"
            + " - for image-to-image generation, you need to provide `image_latents`\n"
            + " - for inpainting, you need to provide `processed_mask_image` and `image_latents`\n"
            + " - for text-to-image generation, all you need to provide is prompt embeddings"
        )

    @property
    def outputs(self):
        return [
            OutputParam.template("latents"),
        ]


# ====================
# 4. DECODE
# ====================


# standard decode step works for most tasks except for inpaint
# auto_docstring
class Krea2DecodeStep(SequentialPipelineBlocks):
    """
    Decode step that decodes the latents to images and postprocess the generated image.

      Components:
          vae (`AutoencoderKLQwenImage`) image_processor (`VaeImageProcessor`)

      Inputs:
          latents (`Tensor`):
              The denoised latents to decode, can be generated in the denoise step and unpacked in the after denoise
              step.
          output_type (`str`, *optional*, defaults to pil):
              Output format: 'pil', 'np', 'pt'.

      Outputs:
          images (`list`):
              Generated images. (tensor output of the vae decoder.)
    """

    model_name = "krea2"
    block_classes = [Krea2DecoderStep(), Krea2ProcessImagesOutputStep()]
    block_names = ["decode", "postprocess"]

    @property
    def description(self):
        return "Decode step that decodes the latents to images and postprocess the generated image."


# Inpaint decode step
# auto_docstring
class Krea2InpaintDecodeStep(SequentialPipelineBlocks):
    """
    Decode step that decodes the latents to images and postprocess the generated image, optionally apply the mask
    overlay to the original image.

      Components:
          vae (`AutoencoderKLQwenImage`) image_mask_processor (`InpaintProcessor`)

      Inputs:
          latents (`Tensor`):
              The denoised latents to decode, can be generated in the denoise step and unpacked in the after denoise
              step.
          output_type (`str`, *optional*, defaults to pil):
              Output format: 'pil', 'np', 'pt'.
          mask_overlay_kwargs (`dict`, *optional*):
              The kwargs for the postprocess step to apply the mask overlay. generated in
              Krea2InpaintProcessImagesInputStep.

      Outputs:
          images (`list`):
              Generated images. (tensor output of the vae decoder.)
    """

    model_name = "krea2"
    block_classes = [Krea2DecoderStep(), Krea2InpaintProcessImagesOutputStep()]
    block_names = ["decode", "postprocess"]

    @property
    def description(self):
        return "Decode step that decodes the latents to images and postprocess the generated image, optionally apply the mask overlay to the original image."


# Auto decode step for Krea 2
class Krea2AutoDecodeStep(AutoPipelineBlocks):
    model_name = "krea2"
    block_classes = [Krea2InpaintDecodeStep, Krea2DecodeStep]
    block_names = ["inpaint_decode", "decode"]
    block_trigger_inputs = ["mask", None]

    @property
    def description(self):
        return (
            "Decode step that decode the latents into images. \n"
            " This is an auto pipeline block that works for inpaint/text2image/img2img tasks.\n"
            + " - `Krea2InpaintDecodeStep` (inpaint_decode) is used when `mask` is provided.\n"
            + " - `Krea2DecodeStep` (decode) is used when `mask` is not provided.\n"
        )


# ====================
# 5. AUTO BLOCKS & PRESETS
# ====================
AUTO_BLOCKS = InsertableDict(
    [
        ("text_encoder", Krea2AutoTextEncoderStep()),
        ("vae_encoder", Krea2AutoVaeEncoderStep()),
        ("denoise", Krea2AutoCoreDenoiseStep()),
        ("decode", Krea2AutoDecodeStep()),
    ]
)


# auto_docstring
class Krea2AutoBlocks(SequentialPipelineBlocks):
    """
    Auto Modular pipeline for text-to-image, image-to-image, and inpainting tasks using Krea 2.

      Supported workflows:
        - `text2image`: requires `prompt`
        - `image2image`: requires `prompt`, `image`
        - `inpainting`: requires `prompt`, `mask_image`, `image`

      Components:
          text_encoder (`Qwen3VLModel`): The text encoder to use tokenizer (`Qwen2Tokenizer`): The tokenizer to use
          guider (`ClassifierFreeGuidance`) image_mask_processor (`InpaintProcessor`) vae (`AutoencoderKLQwenImage`)
          image_processor (`VaeImageProcessor`) pachifier (`Krea2Pachifier`) scheduler
          (`FlowMatchEulerDiscreteScheduler`) transformer (`Krea2Transformer2DModel`)

      Inputs:
          prompt (`str`, *optional*):
              The prompt or prompts to guide image generation.
          negative_prompt (`str`, *optional*):
              The prompt or prompts not to guide the image generation.
          max_sequence_length (`int`, *optional*, defaults to 512):
              Maximum sequence length for prompt encoding.
          mask_image (`Image`, *optional*):
              Mask image for inpainting.
          image (`Image | list`, *optional*):
              Reference image(s) for denoising. Can be a single image or list of images.
          height (`int`, *optional*):
              The height in pixels of the generated image.
          width (`int`, *optional*):
              The width in pixels of the generated image.
          padding_mask_crop (`int`, *optional*):
              Padding for mask cropping in inpainting.
          generator (`Generator`, *optional*):
              Torch generator for deterministic generation.
          num_images_per_prompt (`int`, *optional*, defaults to 1):
              The number of images to generate per prompt.
          prompt_embeds (`Tensor`):
              text embeddings used to guide the image generation. Can be generated from text_encoder step.
          prompt_embeds_mask (`Tensor`):
              mask for the text embeddings. Can be generated from text_encoder step.
          negative_prompt_embeds (`Tensor`, *optional*):
              negative text embeddings used to guide the image generation. Can be generated from text_encoder step.
          negative_prompt_embeds_mask (`Tensor`, *optional*):
              mask for the negative text embeddings. Can be generated from text_encoder step.
          image_latents (`Tensor`, *optional*):
              image latents used to guide the image generation. Can be generated from vae_encoder step.
          processed_mask_image (`Tensor`, *optional*):
              The processed mask image
          latents (`Tensor`):
              Pre-generated noisy latents for image generation.
          num_inference_steps (`int`):
              The number of denoising steps.
          sigmas (`list`, *optional*):
              Custom sigmas for the denoising process.
          mu (`float`, *optional*):
              Fixed timestep shift for the scheduler. Pass `1.15` for the few-step distilled (TDM/turbo) checkpoint; if
              not provided, computed from the image sequence length (base checkpoint behavior).
          strength (`float`, *optional*, defaults to 0.9):
              Strength for img2img/inpainting.
          **denoiser_input_fields (`None`, *optional*):
              conditional model inputs for the denoiser: e.g. prompt_embeds, negative_prompt_embeds, etc.
          output_type (`str`, *optional*, defaults to pil):
              Output format: 'pil', 'np', 'pt'.
          mask_overlay_kwargs (`dict`, *optional*):
              The kwargs for the postprocess step to apply the mask overlay. generated in
              Krea2InpaintProcessImagesInputStep.

      Outputs:
          images (`list`):
              Generated images.
    """

    model_name = "krea2"

    block_classes = AUTO_BLOCKS.values()
    block_names = AUTO_BLOCKS.keys()

    # Workflow map defines the trigger conditions for each workflow.
    # How to define:
    #   - Only include required inputs and trigger inputs (inputs that determine which blocks run)
    #   - currently, only supports `True` means the workflow triggers when the input is not None

    _workflow_map = {
        "text2image": {"prompt": True},
        "image2image": {"prompt": True, "image": True},
        "inpainting": {"prompt": True, "mask_image": True, "image": True},
    }

    @property
    def description(self):
        return "Auto Modular pipeline for text-to-image, image-to-image, and inpainting tasks using Krea 2."

    @property
    def outputs(self):
        return [OutputParam.template("images")]