Fix bug #2931

invokeai/backend/generator/base.py

@@ -347,7 +347,6 @@ class Generator:
             h_symmetry_time_pct=h_symmetry_time_pct,
             v_symmetry_time_pct=v_symmetry_time_pct,
             attention_maps_callback=attention_maps_callback,
-            seed=seed,
             **kwargs,
         )
         results = []
@@ -375,7 +374,8 @@ class Generator:
                         print("** An error occurred while getting initial noise **")
                         print(traceback.format_exc())
 
-                image = make_image(x_T)
+                # Pass on the seed in case a layer beneath us needs to generate noise on its own.
+                image = make_image(x_T, seed)
 
                 if self.safety_checker is not None:
                     image = self.safety_checker.check(image)

invokeai/backend/generator/img2img.py

@@ -37,7 +37,6 @@ class Img2Img(Generator):
         h_symmetry_time_pct=None,
         v_symmetry_time_pct=None,
         attention_maps_callback=None,
-        seed=None,
         **kwargs,
     ):
         """
@@ -64,7 +63,7 @@ class Img2Img(Generator):
             ),
         ).add_scheduler_args_if_applicable(pipeline.scheduler, eta=ddim_eta)
 
-        def make_image(x_T):
+        def make_image(x_T: torch.Tensor, seed: int):
             # FIXME: use x_T for initial seeded noise
             # We're not at the moment because the pipeline automatically resizes init_image if
             # necessary, which the x_T input might not match.
@@ -77,7 +76,7 @@ class Img2Img(Generator):
                 conditioning_data,
                 noise_func=self.get_noise_like,
                 callback=step_callback,
-                seed=seed
+                seed=seed,
             )
             if (
                 pipeline_output.attention_map_saver is not None
@@ -88,9 +87,7 @@ class Img2Img(Generator):
 
         return make_image
 
-    def get_noise_like(self, like: torch.Tensor, seed: Optional[int]):
-        if seed is not None:
-            set_seed(seed)
+    def get_noise_like(self, like: torch.Tensor):
         device = like.device
         if device.type == "mps":
             x = torch.randn_like(like, device="cpu").to(device)

invokeai/backend/generator/inpaint.py

@@ -311,7 +311,7 @@ class Inpaint(Img2Img):
             uc, c, cfg_scale
         ).add_scheduler_args_if_applicable(pipeline.scheduler, eta=ddim_eta)
 
-        def make_image(x_T):
+        def make_image(x_T: torch.Tensor, seed: int):
             pipeline_output = pipeline.inpaint_from_embeddings(
                 init_image=init_image,
                 mask=1 - mask,  # expects white means "paint here."
@@ -320,7 +320,7 @@ class Inpaint(Img2Img):
                 conditioning_data=conditioning_data,
                 noise_func=self.get_noise_like,
                 callback=step_callback,
-                seed=seed
+                seed=seed,
             )
 
             if (

invokeai/backend/generator/txt2img.py

@@ -61,7 +61,7 @@ class Txt2Img(Generator):
             ),
         ).add_scheduler_args_if_applicable(pipeline.scheduler, eta=ddim_eta)
 
-        def make_image(x_T) -> PIL.Image.Image:
+        def make_image(x_T: torch.Tensor, _: int) -> PIL.Image.Image:
             pipeline_output = pipeline.image_from_embeddings(
                 latents=torch.zeros_like(x_T, dtype=self.torch_dtype()),
                 noise=x_T,

invokeai/backend/generator/txt2img2img.py

@@ -64,7 +64,7 @@ class Txt2Img2Img(Generator):
             ),
         ).add_scheduler_args_if_applicable(pipeline.scheduler, eta=ddim_eta)
 
-        def make_image(x_T):
+        def make_image(x_T: torch.Tensor, _: int):
             first_pass_latent_output, _ = pipeline.latents_from_embeddings(
                 latents=torch.zeros_like(x_T),
                 num_inference_steps=steps,

invokeai/backend/stable_diffusion/diffusers_pipeline.py

@@ -9,6 +9,7 @@ from typing import Any, Callable, Generic, List, Optional, Type, TypeVar, Union
 
 import einops
 import PIL.Image
+from accelerate.utils import set_seed
 import psutil
 import torch
 import torchvision.transforms as T
@@ -694,7 +695,9 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
             device=self._model_group.device_for(self.unet),
             dtype=self.unet.dtype,
         )
-        noise = noise_func(initial_latents, seed)
+        if seed is not None:
+            set_seed(seed)
+        noise = noise_func(initial_latents)
 
         return self.img2img_from_latents_and_embeddings(
             initial_latents,
@@ -796,7 +799,9 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         init_image_latents = self.non_noised_latents_from_image(
             init_image, device=device, dtype=latents_dtype
         )
-        noise = noise_func(init_image_latents, seed)
+        if seed is not None:
+            set_seed(seed)
+        noise = noise_func(init_image_latents)
 
         if mask.dim() == 3:
             mask = mask.unsqueeze(0)