before_denoise.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 inspect

import numpy as np
import torch

from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
from diffusers.utils.torch_utils import randn_tensor
from diffusers.modular_pipelines.modular_pipeline import ModularPipelineBlocks, PipelineState
from diffusers.modular_pipelines.modular_pipeline_utils import ComponentSpec, InputParam, OutputParam
from .modular_pipeline import Krea2ModularPipeline, Krea2Pachifier


# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift
def calculate_shift(
    image_seq_len,
    base_seq_len: int = 256,
    max_seq_len: int = 4096,
    base_shift: float = 0.5,
    max_shift: float = 1.15,
):
    m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
    b = base_shift - m * base_seq_len
    mu = image_seq_len * m + b
    return mu


# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
def retrieve_timesteps(
    scheduler,
    num_inference_steps: int | None = None,
    device: str | torch.device | None = None,
    timesteps: list[int] | None = None,
    sigmas: list[float] | None = None,
    **kwargs,
):
    r"""
    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.

    Args:
        scheduler (`SchedulerMixin`):
            The scheduler to get timesteps from.
        num_inference_steps (`int`):
            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
            must be `None`.
        device (`str` or `torch.device`, *optional*):
            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
        timesteps (`list[int]`, *optional*):
            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
            `num_inference_steps` and `sigmas` must be `None`.
        sigmas (`list[float]`, *optional*):
            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
            `num_inference_steps` and `timesteps` must be `None`.

    Returns:
        `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
        second element is the number of inference steps.
    """
    if timesteps is not None and sigmas is not None:
        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
    if timesteps is not None:
        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
        if not accepts_timesteps:
            raise ValueError(
                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
                f" timestep schedules. Please check whether you are using the correct scheduler."
            )
        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
        timesteps = scheduler.timesteps
        num_inference_steps = len(timesteps)
    elif sigmas is not None:
        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
        if not accept_sigmas:
            raise ValueError(
                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
                f" sigmas schedules. Please check whether you are using the correct scheduler."
            )
        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
        timesteps = scheduler.timesteps
        num_inference_steps = len(timesteps)
    else:
        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
        timesteps = scheduler.timesteps
    return timesteps, num_inference_steps


# Copied from diffusers.modular_pipelines.qwenimage.before_denoise.get_timesteps
def get_timesteps(scheduler, num_inference_steps, strength):
    # get the original timestep using init_timestep
    init_timestep = min(num_inference_steps * strength, num_inference_steps)

    t_start = int(max(num_inference_steps - init_timestep, 0))
    timesteps = scheduler.timesteps[t_start * scheduler.order :]
    if hasattr(scheduler, "set_begin_index"):
        scheduler.set_begin_index(t_start * scheduler.order)

    return timesteps, num_inference_steps - t_start


# ====================
# 1. PREPARE LATENTS
# ====================


class Krea2PrepareLatentsStep(ModularPipelineBlocks):
    model_name = "krea2"

    @property
    def description(self) -> str:
        return "Prepare initial random noise for the generation process"

    @property
    def expected_components(self) -> list[ComponentSpec]:
        return [
            ComponentSpec("pachifier", Krea2Pachifier, default_creation_method="from_config"),
        ]

    @property
    def inputs(self) -> list[InputParam]:
        return [
            InputParam.template("latents"),
            InputParam.template("height"),
            InputParam.template("width"),
            InputParam.template("num_images_per_prompt"),
            InputParam.template("generator"),
            InputParam.template("batch_size"),
            InputParam.template("dtype"),
        ]

    @property
    def intermediate_outputs(self) -> list[OutputParam]:
        return [
            OutputParam(name="height", type_hint=int, description="if not set, updated to default value"),
            OutputParam(name="width", type_hint=int, description="if not set, updated to default value"),
            OutputParam(
                name="latents",
                type_hint=torch.Tensor,
                description="The initial latents to use for the denoising process",
            ),
        ]

    @staticmethod
    def check_inputs(height, width, vae_scale_factor):
        if height is not None and height % (vae_scale_factor * 2) != 0:
            raise ValueError(f"Height must be divisible by {vae_scale_factor * 2} but is {height}")

        if width is not None and width % (vae_scale_factor * 2) != 0:
            raise ValueError(f"Width must be divisible by {vae_scale_factor * 2} but is {width}")

    @torch.no_grad()
    def __call__(self, components: Krea2ModularPipeline, state: PipelineState) -> PipelineState:
        block_state = self.get_block_state(state)

        self.check_inputs(
            height=block_state.height,
            width=block_state.width,
            vae_scale_factor=components.vae_scale_factor,
        )

        device = components._execution_device
        batch_size = block_state.batch_size * block_state.num_images_per_prompt

        # we can update the height and width here since it's used to generate the initial
        block_state.height = block_state.height or components.default_height
        block_state.width = block_state.width or components.default_width

        # VAE applies 8x compression on images but we must also account for packing which requires
        # latent height and width to be divisible by 2.
        latent_height = 2 * (int(block_state.height) // (components.vae_scale_factor * 2))
        latent_width = 2 * (int(block_state.width) // (components.vae_scale_factor * 2))

        shape = (batch_size, components.num_channels_latents, 1, latent_height, latent_width)
        if isinstance(block_state.generator, list) and len(block_state.generator) != batch_size:
            raise ValueError(
                f"You have passed a list of generators of length {len(block_state.generator)}, but requested an effective batch"
                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
            )
        if block_state.latents is None:
            block_state.latents = randn_tensor(
                shape, generator=block_state.generator, device=device, dtype=block_state.dtype
            )
            block_state.latents = components.pachifier.pack_latents(block_state.latents)

        self.set_block_state(state, block_state)
        return components, state


class Krea2PrepareLatentsWithStrengthStep(ModularPipelineBlocks):
    model_name = "krea2"

    @property
    def description(self) -> str:
        return "Step that adds noise to image latents for image-to-image/inpainting. Should be run after set_timesteps, prepare_latents. Both noise and image latents should already be patchified."

    @property
    def expected_components(self) -> list[ComponentSpec]:
        return [
            ComponentSpec("scheduler", FlowMatchEulerDiscreteScheduler),
        ]

    @property
    def inputs(self) -> list[InputParam]:
        return [
            InputParam(
                name="latents",
                required=True,
                type_hint=torch.Tensor,
                description="The initial random noised, can be generated in prepare latent step.",
            ),
            InputParam.template("image_latents", note="Can be generated from vae encoder and updated in input step."),
            InputParam(
                name="timesteps",
                required=True,
                type_hint=torch.Tensor,
                description="The timesteps to use for the denoising process. Can be generated in set_timesteps step.",
            ),
        ]

    @property
    def intermediate_outputs(self) -> list[OutputParam]:
        return [
            OutputParam(
                name="initial_noise",
                type_hint=torch.Tensor,
                description="The initial random noised used for inpainting denoising.",
            ),
            OutputParam(
                name="latents",
                type_hint=torch.Tensor,
                description="The scaled noisy latents to use for inpainting/image-to-image denoising.",
            ),
        ]

    @staticmethod
    def check_inputs(image_latents, latents):
        if image_latents.shape[0] != latents.shape[0]:
            raise ValueError(
                f"`image_latents` must have have same batch size as `latents`, but got {image_latents.shape[0]} and {latents.shape[0]}"
            )

        if image_latents.ndim != 3:
            raise ValueError(f"`image_latents` must have 3 dimensions (patchified), but got {image_latents.ndim}")

    @torch.no_grad()
    def __call__(self, components: Krea2ModularPipeline, state: PipelineState) -> PipelineState:
        block_state = self.get_block_state(state)

        self.check_inputs(
            image_latents=block_state.image_latents,
            latents=block_state.latents,
        )

        # prepare latent timestep
        latent_timestep = block_state.timesteps[:1].repeat(block_state.latents.shape[0])

        # make copy of initial_noise
        block_state.initial_noise = block_state.latents

        # scale noise
        block_state.latents = components.scheduler.scale_noise(
            block_state.image_latents, latent_timestep, block_state.latents
        )

        self.set_block_state(state, block_state)

        return components, state


class Krea2CreateMaskLatentsStep(ModularPipelineBlocks):
    model_name = "krea2"

    @property
    def description(self) -> str:
        return "Step that creates mask latents from preprocessed mask_image by interpolating to latent space."

    @property
    def expected_components(self) -> list[ComponentSpec]:
        return [
            ComponentSpec("pachifier", Krea2Pachifier, default_creation_method="from_config"),
        ]

    @property
    def inputs(self) -> list[InputParam]:
        return [
            InputParam(
                name="processed_mask_image",
                required=True,
                type_hint=torch.Tensor,
                description="The processed mask to use for the inpainting process.",
            ),
            InputParam.template("height", required=True),
            InputParam.template("width", required=True),
            InputParam.template("dtype"),
        ]

    @property
    def intermediate_outputs(self) -> list[OutputParam]:
        return [
            OutputParam(
                name="mask", type_hint=torch.Tensor, description="The mask to use for the inpainting process."
            ),
        ]

    @torch.no_grad()
    def __call__(self, components: Krea2ModularPipeline, state: PipelineState) -> PipelineState:
        block_state = self.get_block_state(state)

        device = components._execution_device

        # VAE applies 8x compression on images but we must also account for packing which requires
        # latent height and width to be divisible by 2.

        height_latents = 2 * (int(block_state.height) // (components.vae_scale_factor * 2))
        width_latents = 2 * (int(block_state.width) // (components.vae_scale_factor * 2))

        block_state.mask = torch.nn.functional.interpolate(
            block_state.processed_mask_image,
            size=(height_latents, width_latents),
        )

        block_state.mask = block_state.mask.unsqueeze(2)
        block_state.mask = block_state.mask.repeat(1, components.num_channels_latents, 1, 1, 1)
        block_state.mask = block_state.mask.to(device=device, dtype=block_state.dtype)

        block_state.mask = components.pachifier.pack_latents(block_state.mask)

        self.set_block_state(state, block_state)

        return components, state


# ====================
# 2. SET TIMESTEPS
# ====================


class Krea2SetTimestepsStep(ModularPipelineBlocks):
    model_name = "krea2"

    @property
    def description(self) -> str:
        return "Step that sets the scheduler's timesteps for text-to-image generation. Should be run after prepare latents step."

    @property
    def expected_components(self) -> list[ComponentSpec]:
        return [
            ComponentSpec("scheduler", FlowMatchEulerDiscreteScheduler),
        ]

    @property
    def inputs(self) -> list[InputParam]:
        return [
            InputParam.template("num_inference_steps", default=28),
            InputParam.template("sigmas"),
            InputParam(
                name="mu",
                type_hint=float,
                description="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).",
            ),
            InputParam(
                name="latents",
                required=True,
                type_hint=torch.Tensor,
                description="The initial random noised latents for the denoising process. Can be generated in prepare latents step.",
            ),
        ]

    @property
    def intermediate_outputs(self) -> list[OutputParam]:
        return [
            OutputParam(
                name="timesteps", type_hint=torch.Tensor, description="The timesteps to use for the denoising process"
            ),
        ]

    def __call__(self, components: Krea2ModularPipeline, state: PipelineState) -> PipelineState:
        block_state = self.get_block_state(state)

        device = components._execution_device
        sigmas = (
            np.linspace(1.0, 1 / block_state.num_inference_steps, block_state.num_inference_steps)
            if block_state.sigmas is None
            else block_state.sigmas
        )

        mu = block_state.mu
        if mu is None:
            mu = calculate_shift(
                image_seq_len=block_state.latents.shape[1],
                base_seq_len=components.scheduler.config.get("base_image_seq_len", 256),
                max_seq_len=components.scheduler.config.get("max_image_seq_len", 6400),
                base_shift=components.scheduler.config.get("base_shift", 0.5),
                max_shift=components.scheduler.config.get("max_shift", 1.15),
            )
        block_state.timesteps, block_state.num_inference_steps = retrieve_timesteps(
            scheduler=components.scheduler,
            num_inference_steps=block_state.num_inference_steps,
            device=device,
            sigmas=sigmas,
            mu=mu,
        )

        components.scheduler.set_begin_index(0)

        self.set_block_state(state, block_state)

        return components, state


class Krea2SetTimestepsWithStrengthStep(ModularPipelineBlocks):
    model_name = "krea2"

    @property
    def description(self) -> str:
        return "Step that sets the scheduler's timesteps for image-to-image generation, and inpainting. Should be run after prepare latents step."

    @property
    def expected_components(self) -> list[ComponentSpec]:
        return [
            ComponentSpec("scheduler", FlowMatchEulerDiscreteScheduler),
        ]

    @property
    def inputs(self) -> list[InputParam]:
        return [
            InputParam.template("num_inference_steps", default=28),
            InputParam.template("sigmas"),
            InputParam(
                name="mu",
                type_hint=float,
                description="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).",
            ),
            InputParam(
                "latents",
                required=True,
                type_hint=torch.Tensor,
                description="The latents to use for the denoising process. Can be generated in prepare latents step.",
            ),
            InputParam.template("strength"),
        ]

    @property
    def intermediate_outputs(self) -> list[OutputParam]:
        return [
            OutputParam(
                name="timesteps",
                type_hint=torch.Tensor,
                description="The timesteps to use for the denoising process.",
            ),
            OutputParam(
                name="num_inference_steps",
                type_hint=int,
                description="The number of denoising steps to perform at inference time. Updated based on strength.",
            ),
        ]

    def __call__(self, components: Krea2ModularPipeline, state: PipelineState) -> PipelineState:
        block_state = self.get_block_state(state)

        device = components._execution_device
        sigmas = (
            np.linspace(1.0, 1 / block_state.num_inference_steps, block_state.num_inference_steps)
            if block_state.sigmas is None
            else block_state.sigmas
        )

        mu = block_state.mu
        if mu is None:
            mu = calculate_shift(
                image_seq_len=block_state.latents.shape[1],
                base_seq_len=components.scheduler.config.get("base_image_seq_len", 256),
                max_seq_len=components.scheduler.config.get("max_image_seq_len", 6400),
                base_shift=components.scheduler.config.get("base_shift", 0.5),
                max_shift=components.scheduler.config.get("max_shift", 1.15),
            )
        block_state.timesteps, block_state.num_inference_steps = retrieve_timesteps(
            scheduler=components.scheduler,
            num_inference_steps=block_state.num_inference_steps,
            device=device,
            sigmas=sigmas,
            mu=mu,
        )

        block_state.timesteps, block_state.num_inference_steps = get_timesteps(
            scheduler=components.scheduler,
            num_inference_steps=block_state.num_inference_steps,
            strength=block_state.strength,
        )

        self.set_block_state(state, block_state)

        return components, state


# ====================
# 3. OTHER INPUTS FOR DENOISER
# ====================

## RoPE inputs for denoiser


class Krea2RoPEInputsStep(ModularPipelineBlocks):
    model_name = "krea2"

    @property
    def description(self) -> str:
        return (
            "Step that prepares the rotary position ids for the denoising process. Text tokens sit at the origin, "
            "image tokens carry their `(0, h, w)` latent-grid coordinates. Should be placed after prepare_latents step."
        )

    @property
    def expected_components(self) -> list[ComponentSpec]:
        return [
            ComponentSpec("pachifier", Krea2Pachifier, default_creation_method="from_config"),
        ]

    @property
    def inputs(self) -> list[InputParam]:
        return [
            InputParam.template("height", required=True),
            InputParam.template("width", required=True),
            InputParam.template("prompt_embeds_mask"),
        ]

    @property
    def intermediate_outputs(self) -> list[OutputParam]:
        return [
            OutputParam(
                name="position_ids",
                kwargs_type="denoiser_input_fields",
                type_hint=torch.Tensor,
                description="The rotary coordinates of shape (text_seq_len + grid_height * grid_width, 3) for the combined text-image sequence.",
            ),
        ]

    def __call__(self, components: Krea2ModularPipeline, state: PipelineState) -> PipelineState:
        block_state = self.get_block_state(state)

        device = components._execution_device
        patch_size = components.pachifier.config.patch_size

        text_seq_len = block_state.prompt_embeds_mask.shape[1]
        grid_height = block_state.height // (components.vae_scale_factor * patch_size)
        grid_width = block_state.width // (components.vae_scale_factor * patch_size)

        text_ids = torch.zeros(text_seq_len, 3, device=device)
        image_ids = torch.zeros(grid_height, grid_width, 3, device=device)
        image_ids[..., 1] = torch.arange(grid_height, device=device)[:, None]
        image_ids[..., 2] = torch.arange(grid_width, device=device)[None, :]
        image_ids = image_ids.reshape(grid_height * grid_width, 3)
        block_state.position_ids = torch.cat([text_ids, image_ids], dim=0)

        self.set_block_state(state, block_state)

        return components, state


class Krea2EditRoPEInputsStep(ModularPipelineBlocks):
    model_name = "krea2"

    @property
    def description(self) -> str:
        return (
            "Step that prepares the rotary position ids for the edit task: the `[text, image]` coordinates followed "
            "by the reference tokens' coordinates (each reference on its own frame axis). Should be placed after "
            "prepare_latents and the reference latents step."
        )

    @property
    def expected_components(self) -> list[ComponentSpec]:
        return [
            ComponentSpec("pachifier", Krea2Pachifier, default_creation_method="from_config"),
        ]

    @property
    def inputs(self) -> list[InputParam]:
        return [
            InputParam.template("height", required=True),
            InputParam.template("width", required=True),
            InputParam.template("prompt_embeds_mask"),
            InputParam(
                "reference_position_ids",
                required=True,
                type_hint=torch.Tensor,
                description="Rotary coordinates for the reference tokens. Can be generated in the reference latents step.",
            ),
        ]

    @property
    def intermediate_outputs(self) -> list[OutputParam]:
        return [
            OutputParam(
                name="position_ids",
                kwargs_type="denoiser_input_fields",
                type_hint=torch.Tensor,
                description="The rotary coordinates of shape (text_seq_len + grid_height * grid_width + ref_seq_len, 3) "
                "for the combined text-image-reference sequence.",
            ),
        ]

    def __call__(self, components: Krea2ModularPipeline, state: PipelineState) -> PipelineState:
        block_state = self.get_block_state(state)

        device = components._execution_device
        patch_size = components.pachifier.config.patch_size

        text_seq_len = block_state.prompt_embeds_mask.shape[1]
        grid_height = block_state.height // (components.vae_scale_factor * patch_size)
        grid_width = block_state.width // (components.vae_scale_factor * patch_size)

        text_ids = torch.zeros(text_seq_len, 3, device=device)
        image_ids = torch.zeros(grid_height, grid_width, 3, device=device)
        image_ids[..., 1] = torch.arange(grid_height, device=device)[:, None]
        image_ids[..., 2] = torch.arange(grid_width, device=device)[None, :]
        image_ids = image_ids.reshape(grid_height * grid_width, 3)
        block_state.position_ids = torch.cat([text_ids, image_ids, block_state.reference_position_ids], dim=0)

        self.set_block_state(state, block_state)

        return components, state