OzzyGT HF Staff

add bocks and code

afe9f02 1 day ago

4.2 kB

# Copyright 2026 Sperid Labs 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 numpy as np
import torch
from diffusers.modular_pipelines import (
    ModularPipeline,
    ModularPipelineBlocks,
    PipelineState,
)
from diffusers.modular_pipelines.modular_pipeline_utils import (
    ComponentSpec,
    InputParam,
    OutputParam,
)
from diffusers.utils.torch_utils import randn_tensor

from .components import scheduler_spec, transformer_spec


class IdolSetTimestepsStep(ModularPipelineBlocks):
    model_name = "idol"

    @property
    def description(self) -> str:
        return "Sets the shifted flow-matching timesteps (time shift from the scheduler's `shift`, 3.0 by default)."

    @property
    def expected_components(self) -> list[ComponentSpec]:
        return [scheduler_spec()]

    @property
    def inputs(self) -> list[InputParam]:
        return [InputParam.template("num_inference_steps")]

    @property
    def intermediate_outputs(self) -> list[OutputParam]:
        return [
            OutputParam("timesteps", type_hint=torch.Tensor, description="The timesteps of the denoising loop."),
            OutputParam("num_inference_steps", type_hint=int, description="The number of denoising steps."),
        ]

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

        sigmas = np.linspace(1.0, 0.0, block_state.num_inference_steps + 1)[:-1]
        components.scheduler.set_timesteps(sigmas=sigmas, device=device)
        block_state.timesteps = components.scheduler.timesteps

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


class IdolPrepareLatentsStep(ModularPipelineBlocks):
    model_name = "idol"

    @property
    def description(self) -> str:
        return "Draws the initial noise for the object latent and the camera latent."

    @property
    def expected_components(self) -> list[ComponentSpec]:
        return [transformer_spec()]

    @property
    def inputs(self) -> list[InputParam]:
        return [
            InputParam(
                "image_embeds",
                required=True,
                type_hint=torch.Tensor,
                description="DINOv3 image tokens; sets the batch size.",
            ),
            InputParam.template("generator"),
        ]

    @property
    def intermediate_outputs(self) -> list[OutputParam]:
        return [
            OutputParam("latents", type_hint=torch.Tensor, description="Initial noise of the object latent."),
            OutputParam("camera_latents", type_hint=torch.Tensor, description="Initial noise of the camera latent."),
        ]

    @torch.no_grad()
    def __call__(self, components: ModularPipeline, state: PipelineState) -> PipelineState:
        block_state = self.get_block_state(state)
        device = components._execution_device
        config = components.transformer.config
        batch_size = block_state.image_embeds.shape[0]

        # the sampler integrates in float32, like the reference
        block_state.latents = randn_tensor(
            (batch_size, config.num_latent_tokens, config.in_channels),
            generator=block_state.generator,
            device=device,
            dtype=torch.float32,
        )
        block_state.camera_latents = randn_tensor(
            (batch_size, 1, config.camera_channels),
            generator=block_state.generator,
            device=device,
            dtype=torch.float32,
        )

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

	# Copyright 2026 Sperid Labs 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 numpy as np
	import torch
	from diffusers.modular_pipelines import (
	ModularPipeline,
	ModularPipelineBlocks,
	PipelineState,
	)
	from diffusers.modular_pipelines.modular_pipeline_utils import (
	ComponentSpec,
	InputParam,
	OutputParam,
	)
	from diffusers.utils.torch_utils import randn_tensor

	from .components import scheduler_spec, transformer_spec


	class IdolSetTimestepsStep(ModularPipelineBlocks):
	model_name = "idol"

	@property
	def description(self) -> str:
	return "Sets the shifted flow-matching timesteps (time shift from the scheduler's `shift`, 3.0 by default)."

	@property
	def expected_components(self) -> list[ComponentSpec]:
	return [scheduler_spec()]

	@property
	def inputs(self) -> list[InputParam]:
	return [InputParam.template("num_inference_steps")]

	@property
	def intermediate_outputs(self) -> list[OutputParam]:
	return [
	OutputParam("timesteps", type_hint=torch.Tensor, description="The timesteps of the denoising loop."),
	OutputParam("num_inference_steps", type_hint=int, description="The number of denoising steps."),
	]

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

	sigmas = np.linspace(1.0, 0.0, block_state.num_inference_steps + 1)[:-1]
	components.scheduler.set_timesteps(sigmas=sigmas, device=device)
	block_state.timesteps = components.scheduler.timesteps

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


	class IdolPrepareLatentsStep(ModularPipelineBlocks):
	model_name = "idol"

	@property
	def description(self) -> str:
	return "Draws the initial noise for the object latent and the camera latent."

	@property
	def expected_components(self) -> list[ComponentSpec]:
	return [transformer_spec()]

	@property
	def inputs(self) -> list[InputParam]:
	return [
	InputParam(
	"image_embeds",
	required=True,
	type_hint=torch.Tensor,
	description="DINOv3 image tokens; sets the batch size.",
	),
	InputParam.template("generator"),
	]

	@property
	def intermediate_outputs(self) -> list[OutputParam]:
	return [
	OutputParam("latents", type_hint=torch.Tensor, description="Initial noise of the object latent."),
	OutputParam("camera_latents", type_hint=torch.Tensor, description="Initial noise of the camera latent."),
	]

	@torch.no_grad()
	def __call__(self, components: ModularPipeline, state: PipelineState) -> PipelineState:
	block_state = self.get_block_state(state)
	device = components._execution_device
	config = components.transformer.config
	batch_size = block_state.image_embeds.shape[0]

	# the sampler integrates in float32, like the reference
	block_state.latents = randn_tensor(
	(batch_size, config.num_latent_tokens, config.in_channels),
	generator=block_state.generator,
	device=device,
	dtype=torch.float32,
	)
	block_state.camera_latents = randn_tensor(
	(batch_size, 1, config.camera_channels),
	generator=block_state.generator,
	device=device,
	dtype=torch.float32,
	)

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