Reproducible images by supplying latents to pipeline (huggingface#247)

* Accept latents as input for StableDiffusionPipeline.

* Notebook to demonstrate reusable seeds (latents).

* More accurate type annotation

Co-authored-by: Suraj Patil <[email protected]>

* Review comments: move to device, raise instead of assert.

* Actually commit the test notebook.

I had mistakenly pushed an empty file instead.

* Adapt notebook to Colab.

* Update examples readme.

* Move notebook to personal repo.

Co-authored-by: Suraj Patil <[email protected]>

pipelines/stable_diffusion/pipeline_stable_diffusion.py

@@ -46,6 +46,7 @@ def __call__(
         guidance_scale: Optional[float] = 7.5,
         eta: Optional[float] = 0.0,
         generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.FloatTensor] = None,
         output_type: Optional[str] = "pil",
         **kwargs,
     ):
@@ -98,12 +99,18 @@ def __call__(
             # to avoid doing two forward passes
             text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
 
-        # get the intial random noise
-        latents = torch.randn(
-            (batch_size, self.unet.in_channels, height // 8, width // 8),
-            generator=generator,
-            device=self.device,
-        )
+        # get the initial random noise unless the user supplied it
+        latents_shape = (batch_size, self.unet.in_channels, height // 8, width // 8)
+        if latents is None:
+            latents = torch.randn(
+                latents_shape,
+                generator=generator,
+                device=self.device,
+            )
+        else:
+            if latents.shape != latents_shape:
+                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
+            latents = latents.to(self.device)
 
         # set timesteps
         accepts_offset = "offset" in set(inspect.signature(self.scheduler.set_timesteps).parameters.keys())