File size: 7,095 Bytes
afe9f02 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 | # 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 torch
from diffusers.configuration_utils import FrozenDict
from diffusers.guiders import ClassifierFreeGuidance
from diffusers.modular_pipelines import (
BlockState,
LoopSequentialPipelineBlocks,
ModularPipeline,
ModularPipelineBlocks,
PipelineState,
)
from diffusers.modular_pipelines.modular_pipeline_utils import (
ComponentSpec,
InputParam,
OutputParam,
)
from .components import scheduler_spec, transformer_spec
def pack_latents(latents: torch.Tensor, camera_latents: torch.Tensor) -> torch.Tensor:
"""Object latent [B, N, C] and camera latent [B, 1, K] -> one [B, N * C + K] tensor for the guider and scheduler."""
return torch.cat([latents.flatten(1), camera_latents.flatten(1)], dim=1)
def unpack_latents(packed: torch.Tensor, latents_shape: torch.Size, camera_shape: torch.Size):
num_latent_values = latents_shape[1] * latents_shape[2]
latents = packed[:, :num_latent_values].reshape(latents_shape)
camera_latents = packed[:, num_latent_values:].reshape(camera_shape)
return latents, camera_latents
class IdolLoopDenoiser(ModularPipelineBlocks):
model_name = "idol"
# transformer argument -> (conditional, unconditional) block_state fields
_guider_input_fields = {
"encoder_hidden_states": ("image_embeds", "negative_image_embeds"),
"encoder_hidden_states_vae": ("image_vae_embeds", "negative_image_vae_embeds"),
}
@property
def description(self) -> str:
return "Step within the denoising loop that predicts the guided velocity of the object and camera latents."
@property
def expected_components(self) -> list[ComponentSpec]:
return [
ComponentSpec(
"guider",
ClassifierFreeGuidance,
config=FrozenDict({"guidance_scale": 4.0}),
default_creation_method="from_config",
),
transformer_spec(),
]
@property
def inputs(self) -> list[InputParam]:
return [
InputParam("latents", required=True, type_hint=torch.Tensor, description="The object latent."),
InputParam("camera_latents", required=True, type_hint=torch.Tensor, description="The camera latent."),
InputParam("num_inference_steps", required=True, type_hint=int),
InputParam("image_embeds", required=True, type_hint=torch.Tensor),
InputParam("image_vae_embeds", required=True, type_hint=torch.Tensor),
InputParam("negative_image_embeds", type_hint=torch.Tensor),
InputParam("negative_image_vae_embeds", type_hint=torch.Tensor),
]
@torch.no_grad()
def __call__(self, components: ModularPipeline, block_state: BlockState, i: int, t: torch.Tensor):
components.guider.set_state(step=i, num_inference_steps=block_state.num_inference_steps, timestep=t)
guider_state = components.guider.prepare_inputs_from_block_state(block_state, self._guider_input_fields)
timestep = t.expand(block_state.latents.shape[0]).to(torch.float32)
for guider_state_batch in guider_state:
components.guider.prepare_models(components.transformer)
cond_kwargs = {name: getattr(guider_state_batch, name) for name in self._guider_input_fields}
sample, camera = components.transformer(
hidden_states=block_state.latents,
camera=block_state.camera_latents,
timestep=timestep,
return_dict=False,
**cond_kwargs,
)
guider_state_batch.noise_pred = pack_latents(sample, camera)
components.guider.cleanup_models(components.transformer)
block_state.noise_pred = components.guider(guider_state)[0]
return components, block_state
class IdolLoopAfterDenoiser(ModularPipelineBlocks):
model_name = "idol"
@property
def description(self) -> str:
return "Step within the denoising loop that takes one Euler step on the object and camera latents."
@property
def expected_components(self) -> list[ComponentSpec]:
return [scheduler_spec()]
@property
def intermediate_outputs(self) -> list[OutputParam]:
return [
OutputParam("latents", type_hint=torch.Tensor, description="The denoised object latent."),
OutputParam("camera_latents", type_hint=torch.Tensor, description="The denoised camera latent."),
]
@torch.no_grad()
def __call__(self, components: ModularPipeline, block_state: BlockState, i: int, t: torch.Tensor):
packed = pack_latents(block_state.latents, block_state.camera_latents)
packed = components.scheduler.step(block_state.noise_pred, t, packed, return_dict=False)[0]
block_state.latents, block_state.camera_latents = unpack_latents(
packed, block_state.latents.shape, block_state.camera_latents.shape
)
return components, block_state
class IdolDenoiseLoopWrapper(LoopSequentialPipelineBlocks):
model_name = "idol"
@property
def description(self) -> str:
return "Iteratively denoises the object and camera latents over `timesteps`."
@property
def loop_expected_components(self) -> list[ComponentSpec]:
return [scheduler_spec()]
@property
def loop_inputs(self) -> list[InputParam]:
return [
InputParam("timesteps", required=True, type_hint=torch.Tensor),
InputParam("num_inference_steps", required=True, type_hint=int),
]
@torch.no_grad()
def __call__(self, components: ModularPipeline, state: PipelineState) -> PipelineState:
block_state = self.get_block_state(state)
with self.progress_bar(total=block_state.num_inference_steps) as progress_bar:
for i, t in enumerate(block_state.timesteps):
components, block_state = self.loop_step(components, block_state, i=i, t=t)
progress_bar.update()
self.set_block_state(state, block_state)
return components, state
class IdolDenoiseStep(IdolDenoiseLoopWrapper):
block_classes = [IdolLoopDenoiser, IdolLoopAfterDenoiser]
block_names = ["denoiser", "after_denoiser"]
@property
def description(self) -> str:
return (
"Denoise step that iteratively denoises the object and camera latents. At each step it runs "
"`IdolLoopDenoiser` then `IdolLoopAfterDenoiser`."
)
|