Fix handling handling of 0-step denoising process (#6544)

## Summary

#6522 introduced a change in
behavior in cases where start/end were set such that there are 0
timesteps. This PR reverts that change.

cc @StAlKeR7779 

## QA Instructions

Run with euler, 5 steps, start: 0.0, end: 0.05. I ran this test before
#6522, after #6522, and on this branch. This branch restores the
behavior to pre-#6522 i.e. noise is injected even if no denoising steps
are applied.


## Checklist

- [x] _The PR has a short but descriptive title, suitable for a
changelog_
- [x] _Tests added / updated (if applicable)_
- [x] _Documentation added / updated (if applicable)_

Author: RyanJDick

invokeai/app/invocations/denoise_latents.py

@@ -625,6 +625,7 @@ def init_scheduler(
         t_start_idx *= scheduler.order
         t_end_idx *= scheduler.order
 
+        init_timestep = timesteps[t_start_idx : t_start_idx + 1]
         timesteps = timesteps[t_start_idx : t_start_idx + t_end_idx]
 
         scheduler_step_kwargs: Dict[str, Any] = {}
@@ -647,7 +648,7 @@ def init_scheduler(
         if isinstance(scheduler, TCDScheduler):
             scheduler_step_kwargs.update({"eta": 1.0})
 
-        return timesteps, scheduler_step_kwargs
+        return timesteps, init_timestep, scheduler_step_kwargs
 
     def prep_inpaint_mask(
         self, context: InvocationContext, latents: torch.Tensor
@@ -813,7 +814,7 @@ def _lora_loader() -> Iterator[Tuple[LoRAModelRaw, float]]:
                 dtype=unet.dtype,
             )
 
-            timesteps, scheduler_step_kwargs = self.init_scheduler(
+            timesteps, init_timestep, scheduler_step_kwargs = self.init_scheduler(
                 scheduler,
                 device=unet.device,
                 steps=self.steps,
@@ -825,6 +826,7 @@ def _lora_loader() -> Iterator[Tuple[LoRAModelRaw, float]]:
             result_latents = pipeline.latents_from_embeddings(
                 latents=latents,
                 timesteps=timesteps,
+                init_timestep=init_timestep,
                 noise=noise,
                 seed=seed,
                 mask=mask,

invokeai/app/invocations/tiled_multi_diffusion_denoise_latents.py

@@ -252,7 +252,7 @@ def _lora_loader() -> Iterator[Tuple[LoRAModelRaw, float]]:
                     )
                 )
 
-            timesteps, scheduler_step_kwargs = DenoiseLatentsInvocation.init_scheduler(
+            timesteps, init_timestep, scheduler_step_kwargs = DenoiseLatentsInvocation.init_scheduler(
                 scheduler,
                 device=unet.device,
                 steps=self.steps,
@@ -269,6 +269,7 @@ def _lora_loader() -> Iterator[Tuple[LoRAModelRaw, float]]:
                 scheduler_step_kwargs=scheduler_step_kwargs,
                 noise=noise,
                 timesteps=timesteps,
+                init_timestep=init_timestep,
                 callback=step_callback,
             )
 

invokeai/backend/stable_diffusion/diffusers_pipeline.py

@@ -273,6 +273,7 @@ def latents_from_embeddings(
         noise: Optional[torch.Tensor],
         seed: int,
         timesteps: torch.Tensor,
+        init_timestep: torch.Tensor,
         callback: Callable[[PipelineIntermediateState], None],
         control_data: list[ControlNetData] | None = None,
         ip_adapter_data: Optional[list[IPAdapterData]] = None,
@@ -298,6 +299,8 @@ def latents_from_embeddings(
                 HACK(ryand): seed is only used in a particular case when `noise` is None, but we need to re-generate the
                 same noise used earlier in the pipeline. This should really be handled in a clearer way.
             timesteps: The timestep schedule for the denoising process.
+            init_timestep: The first timestep in the schedule. This is used to determine the initial noise level, so
+                should be populated if you want noise applied *even* if timesteps is empty.
             callback: A callback function that is called to report progress during the denoising process.
             control_data: ControlNet data.
             ip_adapter_data: IP-Adapter data.
@@ -312,17 +315,16 @@ def latents_from_embeddings(
                 SD UNet model.
             is_gradient_mask: A flag indicating whether `mask` is a gradient mask or not.
         """
-        if timesteps.shape[0] == 0:
+        if init_timestep.shape[0] == 0:
             return latents
 
         orig_latents = latents.clone()
+
         batch_size = latents.shape[0]
+        batched_init_timestep = init_timestep.expand(batch_size)
 
         # noise can be None if the latents have already been noised (e.g. when running the SDXL refiner).
         if noise is not None:
-            # batched_init_timestep should have shape (batch_size, 1).
-            batched_init_timestep = timesteps[0:1].expand(batch_size)
-
             # TODO(ryand): I'm pretty sure we should be applying init_noise_sigma in cases where we are starting with
             # full noise. Investigate the history of why this got commented out.
             # latents = noise * self.scheduler.init_noise_sigma # it's like in t2l according to diffusers

invokeai/backend/stable_diffusion/multi_diffusion_pipeline.py

@@ -44,20 +44,19 @@ def multi_diffusion_denoise(
         scheduler_step_kwargs: dict[str, Any],
         noise: Optional[torch.Tensor],
         timesteps: torch.Tensor,
+        init_timestep: torch.Tensor,
         callback: Callable[[PipelineIntermediateState], None],
     ) -> torch.Tensor:
         self._check_regional_prompting(multi_diffusion_conditioning)
 
-        if timesteps.shape[0] == 0:
+        if init_timestep.shape[0] == 0:
             return latents
 
         batch_size, _, latent_height, latent_width = latents.shape
+        batched_init_timestep = init_timestep.expand(batch_size)
 
         # noise can be None if the latents have already been noised (e.g. when running the SDXL refiner).
         if noise is not None:
-            # batched_init_timestep should have shape (batch_size, 1).
-            batched_init_timestep = timesteps[0:1].expand(batch_size)
-
             # TODO(ryand): I'm pretty sure we should be applying init_noise_sigma in cases where we are starting with
             # full noise. Investigate the history of why this got commented out.
             # latents = noise * self.scheduler.init_noise_sigma # it's like in t2l according to diffusers