Move ConditioningData and its field classes to their own file. This will allow new conditioning types to be added more cleanly without introducing circular dependencies.

invokeai/backend/stable_diffusion/__init__.py

@@ -2,14 +2,8 @@
 Initialization file for the invokeai.backend.stable_diffusion package
 """
 from .diffusers_pipeline import (  # noqa: F401
-    ConditioningData,
     PipelineIntermediateState,
     StableDiffusionGeneratorPipeline,
 )
 from .diffusion import InvokeAIDiffuserComponent  # noqa: F401
 from .diffusion.cross_attention_map_saving import AttentionMapSaver  # noqa: F401
-from .diffusion.shared_invokeai_diffusion import (  # noqa: F401
-    PostprocessingSettings,
-    BasicConditioningInfo,
-    SDXLConditioningInfo,
-)

invokeai/backend/stable_diffusion/diffusers_pipeline.py

@@ -28,14 +28,12 @@ from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
 
 from invokeai.app.services.config import InvokeAIAppConfig
 from invokeai.backend.ip_adapter.ip_adapter import IPAdapter, IPAdapterPlus, IPAdapterXL
+from invokeai.backend.stable_diffusion.diffusion.conditioning_data import (
+    ConditioningData,
+)
 
 from ..util import auto_detect_slice_size, normalize_device
-from .diffusion import (
-    AttentionMapSaver,
-    BasicConditioningInfo,
-    InvokeAIDiffuserComponent,
-    PostprocessingSettings,
-)
+from .diffusion import AttentionMapSaver, InvokeAIDiffuserComponent
 
 
 @dataclass
@@ -181,42 +179,6 @@ class IPAdapterData:
     weight: float = Field(default=1.0)
 
 
-@dataclass
-class ConditioningData:
-    unconditioned_embeddings: BasicConditioningInfo
-    text_embeddings: BasicConditioningInfo
-    guidance_scale: Union[float, List[float]]
-    """
-    Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
-    `guidance_scale` is defined as `w` of equation 2. of [Imagen Paper](https://arxiv.org/pdf/2205.11487.pdf).
-    Guidance scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate
-    images that are closely linked to the text `prompt`, usually at the expense of lower image quality.
-    """
-    extra: Optional[InvokeAIDiffuserComponent.ExtraConditioningInfo] = None
-    scheduler_args: dict[str, Any] = field(default_factory=dict)
-    """
-    Additional arguments to pass to invokeai_diffuser.do_latent_postprocessing().
-    """
-    postprocessing_settings: Optional[PostprocessingSettings] = None
-
-    @property
-    def dtype(self):
-        return self.text_embeddings.dtype
-
-    def add_scheduler_args_if_applicable(self, scheduler, **kwargs):
-        scheduler_args = dict(self.scheduler_args)
-        step_method = inspect.signature(scheduler.step)
-        for name, value in kwargs.items():
-            try:
-                step_method.bind_partial(**{name: value})
-            except TypeError:
-                # FIXME: don't silently discard arguments
-                pass  # debug("%s does not accept argument named %r", scheduler, name)
-            else:
-                scheduler_args[name] = value
-        return dataclasses.replace(self, scheduler_args=scheduler_args)
-
-
 @dataclass
 class InvokeAIStableDiffusionPipelineOutput(StableDiffusionPipelineOutput):
     r"""

invokeai/backend/stable_diffusion/diffusion/__init__.py

@@ -3,9 +3,4 @@ Initialization file for invokeai.models.diffusion
 """
 from .cross_attention_control import InvokeAICrossAttentionMixin  # noqa: F401
 from .cross_attention_map_saving import AttentionMapSaver  # noqa: F401
-from .shared_invokeai_diffusion import (  # noqa: F401
-    InvokeAIDiffuserComponent,
-    PostprocessingSettings,
-    BasicConditioningInfo,
-    SDXLConditioningInfo,
-)
+from .shared_invokeai_diffusion import InvokeAIDiffuserComponent  # noqa: F401

invokeai/backend/stable_diffusion/diffusion/conditioning_data.py

@@ -0,0 +1,87 @@
+import dataclasses
+import inspect
+from dataclasses import dataclass, field
+from typing import Any, List, Optional, Union
+
+import torch
+
+from .cross_attention_control import Arguments
+
+
+@dataclass
+class ExtraConditioningInfo:
+    tokens_count_including_eos_bos: int
+    cross_attention_control_args: Optional[Arguments] = None
+
+    @property
+    def wants_cross_attention_control(self):
+        return self.cross_attention_control_args is not None
+
+
+@dataclass
+class BasicConditioningInfo:
+    embeds: torch.Tensor
+    # TODO(ryand): Right now we awkwardly copy the extra conditioning info from here up to `ConditioningData`. This
+    # should only be stored in one place.
+    extra_conditioning: Optional[ExtraConditioningInfo]
+    # weight: float
+    # mode: ConditioningAlgo
+
+    def to(self, device, dtype=None):
+        self.embeds = self.embeds.to(device=device, dtype=dtype)
+        return self
+
+
+@dataclass
+class SDXLConditioningInfo(BasicConditioningInfo):
+    pooled_embeds: torch.Tensor
+    add_time_ids: torch.Tensor
+
+    def to(self, device, dtype=None):
+        self.pooled_embeds = self.pooled_embeds.to(device=device, dtype=dtype)
+        self.add_time_ids = self.add_time_ids.to(device=device, dtype=dtype)
+        return super().to(device=device, dtype=dtype)
+
+
+@dataclass(frozen=True)
+class PostprocessingSettings:
+    threshold: float
+    warmup: float
+    h_symmetry_time_pct: Optional[float]
+    v_symmetry_time_pct: Optional[float]
+
+
+@dataclass
+class ConditioningData:
+    unconditioned_embeddings: BasicConditioningInfo
+    text_embeddings: BasicConditioningInfo
+    guidance_scale: Union[float, List[float]]
+    """
+    Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+    `guidance_scale` is defined as `w` of equation 2. of [Imagen Paper](https://arxiv.org/pdf/2205.11487.pdf).
+    Guidance scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate
+    images that are closely linked to the text `prompt`, usually at the expense of lower image quality.
+    """
+    extra: Optional[ExtraConditioningInfo] = None
+    scheduler_args: dict[str, Any] = field(default_factory=dict)
+    """
+    Additional arguments to pass to invokeai_diffuser.do_latent_postprocessing().
+    """
+    postprocessing_settings: Optional[PostprocessingSettings] = None
+
+    @property
+    def dtype(self):
+        return self.text_embeddings.dtype
+
+    def add_scheduler_args_if_applicable(self, scheduler, **kwargs):
+        scheduler_args = dict(self.scheduler_args)
+        step_method = inspect.signature(scheduler.step)
+        for name, value in kwargs.items():
+            try:
+                step_method.bind_partial(**{name: value})
+            except TypeError:
+                # FIXME: don't silently discard arguments
+                pass  # debug("%s does not accept argument named %r", scheduler, name)
+            else:
+                scheduler_args[name] = value
+        return dataclasses.replace(self, scheduler_args=scheduler_args)

invokeai/backend/stable_diffusion/diffusion/shared_invokeai_diffusion.py

@@ -1,8 +1,7 @@
 from __future__ import annotations
 
-from contextlib import contextmanager
-from dataclasses import dataclass
 import math
+from contextlib import contextmanager
 from typing import Any, Callable, Optional, Union
 
 import torch
@@ -10,9 +9,13 @@ from diffusers import UNet2DConditionModel
 from typing_extensions import TypeAlias
 
 from invokeai.app.services.config import InvokeAIAppConfig
+from invokeai.backend.stable_diffusion.diffusion.conditioning_data import (
+    ExtraConditioningInfo,
+    PostprocessingSettings,
+    SDXLConditioningInfo,
+)
 
 from .cross_attention_control import (
-    Arguments,
     Context,
     CrossAttentionType,
     SwapCrossAttnContext,
@@ -31,37 +34,6 @@ ModelForwardCallback: TypeAlias = Union[
 ]
 
 
-@dataclass
-class BasicConditioningInfo:
-    embeds: torch.Tensor
-    extra_conditioning: Optional[InvokeAIDiffuserComponent.ExtraConditioningInfo]
-    # weight: float
-    # mode: ConditioningAlgo
-
-    def to(self, device, dtype=None):
-        self.embeds = self.embeds.to(device=device, dtype=dtype)
-        return self
-
-
-@dataclass
-class SDXLConditioningInfo(BasicConditioningInfo):
-    pooled_embeds: torch.Tensor
-    add_time_ids: torch.Tensor
-
-    def to(self, device, dtype=None):
-        self.pooled_embeds = self.pooled_embeds.to(device=device, dtype=dtype)
-        self.add_time_ids = self.add_time_ids.to(device=device, dtype=dtype)
-        return super().to(device=device, dtype=dtype)
-
-
-@dataclass(frozen=True)
-class PostprocessingSettings:
-    threshold: float
-    warmup: float
-    h_symmetry_time_pct: Optional[float]
-    v_symmetry_time_pct: Optional[float]
-
-
 class InvokeAIDiffuserComponent:
     """
     The aim of this component is to provide a single place for code that can be applied identically to
@@ -75,15 +47,6 @@ class InvokeAIDiffuserComponent:
     debug_thresholding = False
     sequential_guidance = False
 
-    @dataclass
-    class ExtraConditioningInfo:
-        tokens_count_including_eos_bos: int
-        cross_attention_control_args: Optional[Arguments] = None
-
-        @property
-        def wants_cross_attention_control(self):
-            return self.cross_attention_control_args is not None
-
     def __init__(
         self,
         model,