Remove scheduler_args from ConditioningData structure.

invokeai/app/invocations/latent.py

@@ -1,5 +1,6 @@
 # Copyright (c) 2023 Kyle Schouviller (https://github.com/kyle0654)
 
+import inspect
 import math
 from contextlib import ExitStack
 from functools import singledispatchmethod
@@ -375,14 +376,6 @@ class DenoiseLatentsInvocation(BaseInvocation):
             guidance_rescale_multiplier=self.cfg_rescale_multiplier,
         )
 
-        conditioning_data = conditioning_data.add_scheduler_args_if_applicable(
-            scheduler,
-            # for ddim scheduler
-            eta=0.0,  # ddim_eta
-            # for ancestral and sde schedulers
-            # flip all bits to have noise different from initial
-            generator=torch.Generator(device=unet.device).manual_seed(seed ^ 0xFFFFFFFF),
-        )
         return conditioning_data
 
     def create_pipeline(
@@ -642,7 +635,7 @@ class DenoiseLatentsInvocation(BaseInvocation):
 
     # original idea by https://github.com/AmericanPresidentJimmyCarter
     # TODO: research more for second order schedulers timesteps
-    def init_scheduler(self, scheduler, device, steps, denoising_start, denoising_end):
+    def init_scheduler(self, scheduler, device, steps, denoising_start, denoising_end, seed: int):
         if scheduler.config.get("cpu_only", False):
             scheduler.set_timesteps(steps, device="cpu")
             timesteps = scheduler.timesteps.to(device=device)
@@ -669,7 +662,15 @@ class DenoiseLatentsInvocation(BaseInvocation):
         timesteps = timesteps[t_start_idx : t_start_idx + t_end_idx]
         num_inference_steps = len(timesteps) // scheduler.order
 
-        return num_inference_steps, timesteps, init_timestep
+        scheduler_step_kwargs = {}
+        scheduler_step_signature = inspect.signature(scheduler.step)
+        if "generator" in scheduler_step_signature.parameters:
+            # At some point, someone decided that schedulers that accept a generator should use the original seed with
+            # all bits flipped. I don't know the original rationale for this, but now we must keep it like this for
+            # reproducibility.
+            scheduler_step_kwargs = {"generator": torch.Generator(device=device).manual_seed(seed ^ 0xFFFFFFFF)}
+
+        return num_inference_steps, timesteps, init_timestep, scheduler_step_kwargs
 
     def prep_inpaint_mask(self, context, latents):
         if self.denoise_mask is None:
@@ -783,12 +784,13 @@ class DenoiseLatentsInvocation(BaseInvocation):
                     exit_stack=exit_stack,
                 )
 
-                num_inference_steps, timesteps, init_timestep = self.init_scheduler(
+                num_inference_steps, timesteps, init_timestep, scheduler_step_kwargs = self.init_scheduler(
                     scheduler,
                     device=unet.device,
                     steps=self.steps,
                     denoising_start=self.denoising_start,
                     denoising_end=self.denoising_end,
+                    seed=seed,
                 )
 
                 result_latents = pipeline.latents_from_embeddings(
@@ -800,6 +802,7 @@ class DenoiseLatentsInvocation(BaseInvocation):
                     mask=mask,
                     masked_latents=masked_latents,
                     num_inference_steps=num_inference_steps,
+                    scheduler_step_kwargs=scheduler_step_kwargs,
                     conditioning_data=conditioning_data,
                     control_data=controlnet_data,
                     ip_adapter_data=ip_adapter_data,

invokeai/backend/stable_diffusion/diffusers_pipeline.py

@@ -309,6 +309,7 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         self,
         latents: torch.Tensor,
         num_inference_steps: int,
+        scheduler_step_kwargs: dict[str, Any],
         conditioning_data: ConditioningData,
         *,
         noise: Optional[torch.Tensor],
@@ -368,6 +369,7 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
                 latents,
                 timesteps,
                 conditioning_data,
+                scheduler_step_kwargs=scheduler_step_kwargs,
                 additional_guidance=additional_guidance,
                 control_data=control_data,
                 ip_adapter_data=ip_adapter_data,
@@ -388,6 +390,7 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         latents: torch.Tensor,
         timesteps,
         conditioning_data: ConditioningData,
+        scheduler_step_kwargs: dict[str, Any],
         *,
         additional_guidance: List[Callable] = None,
         control_data: List[ControlNetData] = None,
@@ -454,6 +457,7 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
                     conditioning_data,
                     step_index=i,
                     total_step_count=len(timesteps),
+                    scheduler_step_kwargs=scheduler_step_kwargs,
                     additional_guidance=additional_guidance,
                     control_data=control_data,
                     ip_adapter_data=ip_adapter_data,
@@ -485,6 +489,7 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         conditioning_data: ConditioningData,
         step_index: int,
         total_step_count: int,
+        scheduler_step_kwargs: dict[str, Any],
         additional_guidance: List[Callable] = None,
         control_data: List[ControlNetData] = None,
         ip_adapter_data: Optional[list[IPAdapterData]] = None,
@@ -584,7 +589,7 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
             )
 
         # compute the previous noisy sample x_t -> x_t-1
-        step_output = self.scheduler.step(noise_pred, timestep, latents, **conditioning_data.scheduler_args)
+        step_output = self.scheduler.step(noise_pred, timestep, latents, **scheduler_step_kwargs)
 
         # TODO: issue to diffusers?
         # undo internal counter increment done by scheduler.step, so timestep can be resolved as before call

invokeai/backend/stable_diffusion/diffusion/conditioning_data.py

@@ -1,7 +1,5 @@
-import dataclasses
-import inspect
-from dataclasses import dataclass, field
-from typing import Any, List, Optional, Union
+from dataclasses import dataclass
+from typing import List, Optional, Union
 
 import torch
 
@@ -73,19 +71,5 @@ class ConditioningData:
      ref [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf)
     """
     guidance_rescale_multiplier: float = 0
-    scheduler_args: dict[str, Any] = field(default_factory=dict)
 
     ip_adapter_conditioning: Optional[list[IPAdapterConditioningInfo]] = None
-
-    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)