Upscaling fixed (#1402)

* Upscaling fixed

* up

* more fixes

* fix

* more fixes

* finish again

* up

src/diffusers/pipeline_utils.py

@@ -544,7 +544,14 @@ class DiffusionPipeline(ConfigMixin):
         init_kwargs = {**init_kwargs, **passed_pipe_kwargs}
 
         # remove `null` components
-        init_dict = {k: v for k, v in init_dict.items() if v[0] is not None}
+        def load_module(name, value):
+            if value[0] is None:
+                return False
+            if name in passed_class_obj and passed_class_obj[name] is None:
+                return False
+            return True
+
+        init_dict = {k: v for k, v in init_dict.items() if load_module(k, v)}
 
         if len(unused_kwargs) > 0:
             logger.warning(f"Keyword arguments {unused_kwargs} not recognized.")
@@ -560,12 +567,11 @@ class DiffusionPipeline(ConfigMixin):
 
             is_pipeline_module = hasattr(pipelines, library_name)
             loaded_sub_model = None
-            sub_model_should_be_defined = True
 
             # if the model is in a pipeline module, then we load it from the pipeline
             if name in passed_class_obj:
                 # 1. check that passed_class_obj has correct parent class
-                if not is_pipeline_module and passed_class_obj[name] is not None:
+                if not is_pipeline_module:
                     library = importlib.import_module(library_name)
                     class_obj = getattr(library, class_name)
                     importable_classes = LOADABLE_CLASSES[library_name]
@@ -581,12 +587,6 @@ class DiffusionPipeline(ConfigMixin):
                             f"{passed_class_obj[name]} is of type: {type(passed_class_obj[name])}, but should be"
                             f" {expected_class_obj}"
                         )
-                elif passed_class_obj[name] is None and name not in pipeline_class._optional_components:
-                    logger.warning(
-                        f"You have passed `None` for {name} to disable its functionality in {pipeline_class}. Note"
-                        f" that this might lead to problems when using {pipeline_class} and is not recommended."
-                    )
-                    sub_model_should_be_defined = False
                 else:
                     logger.warning(
                         f"You have passed a non-standard module {passed_class_obj[name]}. We cannot verify whether it"
@@ -608,7 +608,7 @@ class DiffusionPipeline(ConfigMixin):
                 importable_classes = LOADABLE_CLASSES[library_name]
                 class_candidates = {c: getattr(library, c, None) for c in importable_classes.keys()}
 
-            if loaded_sub_model is None and sub_model_should_be_defined:
+            if loaded_sub_model is None:
                 load_method_name = None
                 for class_name, class_candidate in class_candidates.items():
                     if class_candidate is not None and issubclass(class_obj, class_candidate):

src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion.py

@@ -141,6 +141,7 @@ class AltDiffusionPipeline(DiffusionPipeline):
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
         )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
     def enable_xformers_memory_efficient_attention(self):
@@ -379,7 +380,7 @@ class AltDiffusionPipeline(DiffusionPipeline):
             )
 
     def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
-        shape = (batch_size, num_channels_latents, height // 8, width // 8)
+        shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
         if latents is None:
             if device.type == "mps":
                 # randn does not work reproducibly on mps
@@ -420,9 +421,9 @@ class AltDiffusionPipeline(DiffusionPipeline):
         Args:
             prompt (`str` or `List[str]`):
                 The prompt or prompts to guide the image generation.
-            height (`int`, *optional*, defaults to self.unet.config.sample_size * 8):
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                 The height in pixels of the generated image.
-            width (`int`, *optional*, defaults to self.unet.config.sample_size * 8):
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                 The width in pixels of the generated image.
             num_inference_steps (`int`, *optional*, defaults to 50):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
@@ -469,8 +470,8 @@ class AltDiffusionPipeline(DiffusionPipeline):
             (nsfw) content, according to the `safety_checker`.
         """
         # 0. Default height and width to unet
-        height = height or self.unet.config.sample_size * 8
-        width = width or self.unet.config.sample_size * 8
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
 
         # 1. Check inputs. Raise error if not correct
         self.check_inputs(prompt, height, width, callback_steps)

src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion_img2img.py

@@ -154,6 +154,7 @@ class AltDiffusionImg2ImgPipeline(DiffusionPipeline):
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
         )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
     def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):

src/diffusers/pipelines/latent_diffusion/pipeline_latent_diffusion.py

@@ -60,6 +60,7 @@ class LDMTextToImagePipeline(DiffusionPipeline):
     ):
         super().__init__()
         self.register_modules(vqvae=vqvae, bert=bert, tokenizer=tokenizer, unet=unet, scheduler=scheduler)
+        self.vae_scale_factor = 2 ** (len(self.vqvae.config.block_out_channels) - 1)
 
     @torch.no_grad()
     def __call__(
@@ -79,9 +80,9 @@ class LDMTextToImagePipeline(DiffusionPipeline):
         Args:
             prompt (`str` or `List[str]`):
                 The prompt or prompts to guide the image generation.
-            height (`int`, *optional*, defaults to self.unet.config.sample_size * 8):
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                 The height in pixels of the generated image.
-            width (`int`, *optional*, defaults to self.unet.config.sample_size * 8):
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                 The width in pixels of the generated image.
             num_inference_steps (`int`, *optional*, defaults to 50):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
@@ -107,8 +108,8 @@ class LDMTextToImagePipeline(DiffusionPipeline):
             generated images.
         """
         # 0. Default height and width to unet
-        height = height or self.unet.config.sample_size * 8
-        width = width or self.unet.config.sample_size * 8
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
 
         if isinstance(prompt, str):
             batch_size = 1

src/diffusers/pipelines/stable_diffusion/pipeline_flax_stable_diffusion.py

@@ -106,6 +106,7 @@ class FlaxStableDiffusionPipeline(FlaxDiffusionPipeline):
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
         )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
 
     def prepare_inputs(self, prompt: Union[str, List[str]]):
         if not isinstance(prompt, (str, list)):
@@ -168,8 +169,8 @@ class FlaxStableDiffusionPipeline(FlaxDiffusionPipeline):
         neg_prompt_ids: jnp.array = None,
     ):
         # 0. Default height and width to unet
-        height = height or self.unet.config.sample_size * 8
-        width = width or self.unet.config.sample_size * 8
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
 
         if height % 8 != 0 or width % 8 != 0:
             raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
@@ -192,7 +193,12 @@ class FlaxStableDiffusionPipeline(FlaxDiffusionPipeline):
         uncond_embeddings = self.text_encoder(uncond_input, params=params["text_encoder"])[0]
         context = jnp.concatenate([uncond_embeddings, text_embeddings])
 
-        latents_shape = (batch_size, self.unet.in_channels, height // 8, width // 8)
+        latents_shape = (
+            batch_size,
+            self.unet.in_channels,
+            height // self.vae_scale_factor,
+            width // self.vae_scale_factor,
+        )
         if latents is None:
             latents = jax.random.normal(prng_seed, shape=latents_shape, dtype=jnp.float32)
         else:
@@ -269,9 +275,9 @@ class FlaxStableDiffusionPipeline(FlaxDiffusionPipeline):
         Args:
             prompt (`str` or `List[str]`):
                 The prompt or prompts to guide the image generation.
-            height (`int`, *optional*, defaults to self.unet.config.sample_size * 8):
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                 The height in pixels of the generated image.
-            width (`int`, *optional*, defaults to self.unet.config.sample_size * 8):
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                 The width in pixels of the generated image.
             num_inference_steps (`int`, *optional*, defaults to 50):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
@@ -307,8 +313,8 @@ class FlaxStableDiffusionPipeline(FlaxDiffusionPipeline):
             "not-safe-for-work" (nsfw) content, according to the `safety_checker`.
         """
         # 0. Default height and width to unet
-        height = height or self.unet.config.sample_size * 8
-        width = width or self.unet.config.sample_size * 8
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
 
         if jit:
             images = _p_generate(

src/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion.py

@@ -108,6 +108,7 @@ class OnnxStableDiffusionPipeline(DiffusionPipeline):
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
         )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
     def _encode_prompt(self, prompt, num_images_per_prompt, do_classifier_free_guidance, negative_prompt):
@@ -206,8 +207,8 @@ class OnnxStableDiffusionPipeline(DiffusionPipeline):
         **kwargs,
     ):
         # 0. Default height and width to unet
-        height = height or self.unet.config.sample_size * 8
-        width = width or self.unet.config.sample_size * 8
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
 
         if isinstance(prompt, str):
             batch_size = 1
@@ -241,7 +242,12 @@ class OnnxStableDiffusionPipeline(DiffusionPipeline):
 
         # get the initial random noise unless the user supplied it
         latents_dtype = text_embeddings.dtype
-        latents_shape = (batch_size * num_images_per_prompt, 4, height // 8, width // 8)
+        latents_shape = (
+            batch_size * num_images_per_prompt,
+            4,
+            height // self.vae_scale_factor,
+            width // self.vae_scale_factor,
+        )
         if latents is None:
             latents = generator.randn(*latents_shape).astype(latents_dtype)
         elif latents.shape != latents_shape:

src/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_inpaint.py

@@ -158,6 +158,7 @@ class OnnxStableDiffusionInpaintPipeline(DiffusionPipeline):
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
         )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_onnx_stable_diffusion.OnnxStableDiffusionPipeline._encode_prompt
@@ -273,9 +274,9 @@ class OnnxStableDiffusionInpaintPipeline(DiffusionPipeline):
                 repainted, while black pixels will be preserved. If `mask_image` is a PIL image, it will be converted
                 to a single channel (luminance) before use. If it's a tensor, it should contain one color channel (L)
                 instead of 3, so the expected shape would be `(B, H, W, 1)`.
-            height (`int`, *optional*, defaults to self.unet.config.sample_size * 8):
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                 The height in pixels of the generated image.
-            width (`int`, *optional*, defaults to self.unet.config.sample_size * 8):
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                 The width in pixels of the generated image.
             num_inference_steps (`int`, *optional*, defaults to 50):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
@@ -321,8 +322,8 @@ class OnnxStableDiffusionInpaintPipeline(DiffusionPipeline):
             (nsfw) content, according to the `safety_checker`.
         """
         # 0. Default height and width to unet
-        height = height or self.unet.config.sample_size * 8
-        width = width or self.unet.config.sample_size * 8
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
 
         if isinstance(prompt, str):
             batch_size = 1
@@ -358,7 +359,12 @@ class OnnxStableDiffusionInpaintPipeline(DiffusionPipeline):
         )
 
         num_channels_latents = NUM_LATENT_CHANNELS
-        latents_shape = (batch_size * num_images_per_prompt, num_channels_latents, height // 8, width // 8)
+        latents_shape = (
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height // self.vae_scale_factor,
+            width // self.vae_scale_factor,
+        )
         latents_dtype = text_embeddings.dtype
         if latents is None:
             latents = generator.randn(*latents_shape).astype(latents_dtype)

src/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_inpaint_legacy.py

@@ -27,11 +27,11 @@ def preprocess(image):
     return 2.0 * image - 1.0
 
 
-def preprocess_mask(mask):
+def preprocess_mask(mask, scale_factor=8):
     mask = mask.convert("L")
     w, h = mask.size
     w, h = map(lambda x: x - x % 32, (w, h))  # resize to integer multiple of 32
-    mask = mask.resize((w // 8, h // 8), resample=PIL.Image.NEAREST)
+    mask = mask.resize((w // scale_factor, h // scale_factor), resample=PIL.Image.NEAREST)
     mask = np.array(mask).astype(np.float32) / 255.0
     mask = np.tile(mask, (4, 1, 1))
     mask = mask[None].transpose(0, 1, 2, 3)  # what does this step do?
@@ -143,6 +143,7 @@ class OnnxStableDiffusionInpaintPipelineLegacy(DiffusionPipeline):
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
         )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_onnx_stable_diffusion.OnnxStableDiffusionPipeline._encode_prompt
@@ -349,7 +350,7 @@ class OnnxStableDiffusionInpaintPipelineLegacy(DiffusionPipeline):
 
         # preprocess mask
         if not isinstance(mask_image, np.ndarray):
-            mask_image = preprocess_mask(mask_image)
+            mask_image = preprocess_mask(mask_image, self.vae_scale_factor)
         mask_image = mask_image.astype(latents_dtype)
         mask = np.concatenate([mask_image] * num_images_per_prompt, axis=0)
 

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py

@@ -140,6 +140,7 @@ class StableDiffusionPipeline(DiffusionPipeline):
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
         )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
     def enable_xformers_memory_efficient_attention(self):
@@ -378,7 +379,7 @@ class StableDiffusionPipeline(DiffusionPipeline):
             )
 
     def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
-        shape = (batch_size, num_channels_latents, height // 8, width // 8)
+        shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
         if latents is None:
             if device.type == "mps":
                 # randn does not work reproducibly on mps
@@ -419,9 +420,9 @@ class StableDiffusionPipeline(DiffusionPipeline):
         Args:
             prompt (`str` or `List[str]`):
                 The prompt or prompts to guide the image generation.
-            height (`int`, *optional*, defaults to self.unet.config.sample_size * 8):
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                 The height in pixels of the generated image.
-            width (`int`, *optional*, defaults to self.unet.config.sample_size * 8):
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                 The width in pixels of the generated image.
             num_inference_steps (`int`, *optional*, defaults to 50):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
@@ -468,8 +469,8 @@ class StableDiffusionPipeline(DiffusionPipeline):
             (nsfw) content, according to the `safety_checker`.
         """
         # 0. Default height and width to unet
-        height = height or self.unet.config.sample_size * 8
-        width = width or self.unet.config.sample_size * 8
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
 
         # 1. Check inputs. Raise error if not correct
         self.check_inputs(prompt, height, width, callback_steps)

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_image_variation.py

@@ -108,6 +108,7 @@ class StableDiffusionImageVariationPipeline(DiffusionPipeline):
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
         )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_xformers_memory_efficient_attention
@@ -281,7 +282,7 @@ class StableDiffusionImageVariationPipeline(DiffusionPipeline):
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
     def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
-        shape = (batch_size, num_channels_latents, height // 8, width // 8)
+        shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
         if latents is None:
             if device.type == "mps":
                 # randn does not work reproducibly on mps
@@ -324,9 +325,9 @@ class StableDiffusionImageVariationPipeline(DiffusionPipeline):
                 configuration of
                 [this](https://huggingface.co/lambdalabs/sd-image-variations-diffusers/blob/main/feature_extractor/preprocessor_config.json)
                 `CLIPFeatureExtractor`
-            height (`int`, *optional*, defaults to self.unet.config.sample_size * 8):
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                 The height in pixels of the generated image.
-            width (`int`, *optional*, defaults to self.unet.config.sample_size * 8):
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                 The width in pixels of the generated image.
             num_inference_steps (`int`, *optional*, defaults to 50):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
@@ -370,8 +371,8 @@ class StableDiffusionImageVariationPipeline(DiffusionPipeline):
             (nsfw) content, according to the `safety_checker`.
         """
         # 0. Default height and width to unet
-        height = height or self.unet.config.sample_size * 8
-        width = width or self.unet.config.sample_size * 8
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
 
         # 1. Check inputs. Raise error if not correct
         self.check_inputs(image, height, width, callback_steps)

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py

@@ -153,6 +153,7 @@ class StableDiffusionImg2ImgPipeline(DiffusionPipeline):
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
         )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_attention_slicing

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py

@@ -218,6 +218,7 @@ class StableDiffusionInpaintPipeline(DiffusionPipeline):
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
         )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_attention_slicing
@@ -468,7 +469,7 @@ class StableDiffusionInpaintPipeline(DiffusionPipeline):
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
     def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
-        shape = (batch_size, num_channels_latents, height // 8, width // 8)
+        shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
         if latents is None:
             if device.type == "mps":
                 # randn does not work reproducibly on mps
@@ -490,7 +491,9 @@ class StableDiffusionInpaintPipeline(DiffusionPipeline):
         # resize the mask to latents shape as we concatenate the mask to the latents
         # we do that before converting to dtype to avoid breaking in case we're using cpu_offload
         # and half precision
-        mask = torch.nn.functional.interpolate(mask, size=(height // 8, width // 8))
+        mask = torch.nn.functional.interpolate(
+            mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor)
+        )
         mask = mask.to(device=device, dtype=dtype)
 
         masked_image = masked_image.to(device=device, dtype=dtype)
@@ -547,9 +550,9 @@ class StableDiffusionInpaintPipeline(DiffusionPipeline):
                 repainted, while black pixels will be preserved. If `mask_image` is a PIL image, it will be converted
                 to a single channel (luminance) before use. If it's a tensor, it should contain one color channel (L)
                 instead of 3, so the expected shape would be `(B, H, W, 1)`.
-            height (`int`, *optional*, defaults to self.unet.config.sample_size * 8):
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                 The height in pixels of the generated image.
-            width (`int`, *optional*, defaults to self.unet.config.sample_size * 8):
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                 The width in pixels of the generated image.
             num_inference_steps (`int`, *optional*, defaults to 50):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
@@ -596,8 +599,8 @@ class StableDiffusionInpaintPipeline(DiffusionPipeline):
             (nsfw) content, according to the `safety_checker`.
         """
         # 0. Default height and width to unet
-        height = height or self.unet.config.sample_size * 8
-        width = width or self.unet.config.sample_size * 8
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
 
         # 1. Check inputs
         self.check_inputs(prompt, height, width, callback_steps)

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint_legacy.py

@@ -51,11 +51,11 @@ def preprocess_image(image):
     return 2.0 * image - 1.0
 
 
-def preprocess_mask(mask):
+def preprocess_mask(mask, scale_factor=8):
     mask = mask.convert("L")
     w, h = mask.size
     w, h = map(lambda x: x - x % 32, (w, h))  # resize to integer multiple of 32
-    mask = mask.resize((w // 8, h // 8), resample=PIL_INTERPOLATION["nearest"])
+    mask = mask.resize((w // scale_factor, h // scale_factor), resample=PIL_INTERPOLATION["nearest"])
     mask = np.array(mask).astype(np.float32) / 255.0
     mask = np.tile(mask, (4, 1, 1))
     mask = mask[None].transpose(0, 1, 2, 3)  # what does this step do?
@@ -166,6 +166,7 @@ class StableDiffusionInpaintPipelineLegacy(DiffusionPipeline):
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
         )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_attention_slicing
@@ -541,7 +542,7 @@ class StableDiffusionInpaintPipelineLegacy(DiffusionPipeline):
             init_image = preprocess_image(init_image)
 
         if not isinstance(mask_image, torch.FloatTensor):
-            mask_image = preprocess_mask(mask_image)
+            mask_image = preprocess_mask(mask_image, self.vae_scale_factor)
 
         # 5. set timesteps
         self.scheduler.set_timesteps(num_inference_steps, device=device)

src/diffusers/pipelines/stable_diffusion_safe/pipeline_stable_diffusion_safe.py

@@ -136,6 +136,7 @@ class StableDiffusionPipelineSafe(DiffusionPipeline):
             feature_extractor=feature_extractor,
         )
         self._safety_text_concept = safety_concept
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
     @property
@@ -443,7 +444,7 @@ class StableDiffusionPipelineSafe(DiffusionPipeline):
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
     def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
-        shape = (batch_size, num_channels_latents, height // 8, width // 8)
+        shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
         if latents is None:
             if device.type == "mps":
                 # randn does not work reproducibly on mps
@@ -531,9 +532,9 @@ class StableDiffusionPipelineSafe(DiffusionPipeline):
         Args:
             prompt (`str` or `List[str]`):
                 The prompt or prompts to guide the image generation.
-            height (`int`, *optional*, defaults to self.unet.config.sample_size * 8):
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                 The height in pixels of the generated image.
-            width (`int`, *optional*, defaults to self.unet.config.sample_size * 8):
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                 The width in pixels of the generated image.
             num_inference_steps (`int`, *optional*, defaults to 50):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
@@ -600,8 +601,8 @@ class StableDiffusionPipelineSafe(DiffusionPipeline):
             (nsfw) content, according to the `safety_checker`.
         """
         # 0. Default height and width to unet
-        height = height or self.unet.config.sample_size * 8
-        width = width or self.unet.config.sample_size * 8
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
 
         # 1. Check inputs. Raise error if not correct
         self.check_inputs(prompt, height, width, callback_steps)

src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion.py

@@ -78,6 +78,7 @@ class VersatileDiffusionPipeline(DiffusionPipeline):
             vae=vae,
             scheduler=scheduler,
         )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
 
     def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
         r"""
@@ -131,9 +132,9 @@ class VersatileDiffusionPipeline(DiffusionPipeline):
         Args:
             image (`PIL.Image.Image`, `List[PIL.Image.Image]` or `torch.Tensor`):
                 The image prompt or prompts to guide the image generation.
-            height (`int`, *optional*, defaults to self.image_unet.config.sample_size * 8):
+            height (`int`, *optional*, defaults to self.image_unet.config.sample_size * self.vae_scale_factor):
                 The height in pixels of the generated image.
-            width (`int`, *optional*, defaults to self.image_unet.config.sample_size * 8):
+            width (`int`, *optional*, defaults to self.image_unet.config.sample_size * self.vae_scale_factor):
                 The width in pixels of the generated image.
             num_inference_steps (`int`, *optional*, defaults to 50):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
@@ -247,9 +248,9 @@ class VersatileDiffusionPipeline(DiffusionPipeline):
         Args:
             prompt (`str` or `List[str]`):
                 The prompt or prompts to guide the image generation.
-            height (`int`, *optional*, defaults to self.image_unet.config.sample_size * 8):
+            height (`int`, *optional*, defaults to self.image_unet.config.sample_size * self.vae_scale_factor):
                 The height in pixels of the generated image.
-            width (`int`, *optional*, defaults to self.image_unet.config.sample_size * 8):
+            width (`int`, *optional*, defaults to self.image_unet.config.sample_size * self.vae_scale_factor):
                 The width in pixels of the generated image.
             num_inference_steps (`int`, *optional*, defaults to 50):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
@@ -360,9 +361,9 @@ class VersatileDiffusionPipeline(DiffusionPipeline):
         Args:
             prompt (`str` or `List[str]`):
                 The prompt or prompts to guide the image generation.
-            height (`int`, *optional*, defaults to self.image_unet.config.sample_size * 8):
+            height (`int`, *optional*, defaults to self.image_unet.config.sample_size * self.vae_scale_factor):
                 The height in pixels of the generated image.
-            width (`int`, *optional*, defaults to self.image_unet.config.sample_size * 8):
+            width (`int`, *optional*, defaults to self.image_unet.config.sample_size * self.vae_scale_factor):
                 The width in pixels of the generated image.
             num_inference_steps (`int`, *optional*, defaults to 50):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the

src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion_dual_guided.py

@@ -87,6 +87,7 @@ class VersatileDiffusionDualGuidedPipeline(DiffusionPipeline):
             vae=vae,
             scheduler=scheduler,
         )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
 
         if self.text_unet is not None and (
             "dual_cross_attention" not in self.image_unet.config or not self.image_unet.config.dual_cross_attention
@@ -419,7 +420,7 @@ class VersatileDiffusionDualGuidedPipeline(DiffusionPipeline):
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
     def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
-        shape = (batch_size, num_channels_latents, height // 8, width // 8)
+        shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
         if latents is None:
             if device.type == "mps":
                 # randn does not work reproducibly on mps
@@ -474,9 +475,9 @@ class VersatileDiffusionDualGuidedPipeline(DiffusionPipeline):
         Args:
             prompt (`str` or `List[str]`):
                 The prompt or prompts to guide the image generation.
-            height (`int`, *optional*, defaults to self.image_unet.config.sample_size * 8):
+            height (`int`, *optional*, defaults to self.image_unet.config.sample_size * self.vae_scale_factor):
                 The height in pixels of the generated image.
-            width (`int`, *optional*, defaults to self.image_unet.config.sample_size * 8):
+            width (`int`, *optional*, defaults to self.image_unet.config.sample_size * self.vae_scale_factor):
                 The width in pixels of the generated image.
             num_inference_steps (`int`, *optional*, defaults to 50):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
@@ -552,8 +553,8 @@ class VersatileDiffusionDualGuidedPipeline(DiffusionPipeline):
             returning a tuple, the first element is a list with the generated images.
         """
         # 0. Default height and width to unet
-        height = height or self.image_unet.config.sample_size * 8
-        width = width or self.image_unet.config.sample_size * 8
+        height = height or self.image_unet.config.sample_size * self.vae_scale_factor
+        width = width or self.image_unet.config.sample_size * self.vae_scale_factor
 
         # 1. Check inputs. Raise error if not correct
         self.check_inputs(prompt, image, height, width, callback_steps)

src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion_image_variation.py

@@ -71,6 +71,7 @@ class VersatileDiffusionImageVariationPipeline(DiffusionPipeline):
             vae=vae,
             scheduler=scheduler,
         )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_xformers_memory_efficient_attention with unet->image_unet
     def enable_xformers_memory_efficient_attention(self):
@@ -277,7 +278,7 @@ class VersatileDiffusionImageVariationPipeline(DiffusionPipeline):
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
     def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
-        shape = (batch_size, num_channels_latents, height // 8, width // 8)
+        shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
         if latents is None:
             if device.type == "mps":
                 # randn does not work reproducibly on mps
@@ -318,9 +319,9 @@ class VersatileDiffusionImageVariationPipeline(DiffusionPipeline):
         Args:
             image (`PIL.Image.Image`, `List[PIL.Image.Image]` or `torch.Tensor`):
                 The image prompt or prompts to guide the image generation.
-            height (`int`, *optional*, defaults to self.image_unet.config.sample_size * 8):
+            height (`int`, *optional*, defaults to self.image_unet.config.sample_size * self.vae_scale_factor):
                 The height in pixels of the generated image.
-            width (`int`, *optional*, defaults to self.image_unet.config.sample_size * 8):
+            width (`int`, *optional*, defaults to self.image_unet.config.sample_size * self.vae_scale_factor):
                 The width in pixels of the generated image.
             num_inference_steps (`int`, *optional*, defaults to 50):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
@@ -392,8 +393,8 @@ class VersatileDiffusionImageVariationPipeline(DiffusionPipeline):
             (nsfw) content, according to the `safety_checker`.
         """
         # 0. Default height and width to unet
-        height = height or self.image_unet.config.sample_size * 8
-        width = width or self.image_unet.config.sample_size * 8
+        height = height or self.image_unet.config.sample_size * self.vae_scale_factor
+        width = width or self.image_unet.config.sample_size * self.vae_scale_factor
 
         # 1. Check inputs. Raise error if not correct
         self.check_inputs(image, height, width, callback_steps)

src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion_text_to_image.py

@@ -75,6 +75,7 @@ class VersatileDiffusionTextToImagePipeline(DiffusionPipeline):
             vae=vae,
             scheduler=scheduler,
         )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
 
         if self.text_unet is not None:
             self._swap_unet_attention_blocks()
@@ -337,7 +338,7 @@ class VersatileDiffusionTextToImagePipeline(DiffusionPipeline):
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
     def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
-        shape = (batch_size, num_channels_latents, height // 8, width // 8)
+        shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
         if latents is None:
             if device.type == "mps":
                 # randn does not work reproducibly on mps
@@ -378,9 +379,9 @@ class VersatileDiffusionTextToImagePipeline(DiffusionPipeline):
         Args:
             prompt (`str` or `List[str]`):
                 The prompt or prompts to guide the image generation.
-            height (`int`, *optional*, defaults to self.image_unet.config.sample_size * 8):
+            height (`int`, *optional*, defaults to self.image_unet.config.sample_size * self.vae_scale_factor):
                 The height in pixels of the generated image.
-            width (`int`, *optional*, defaults to self.image_unet.config.sample_size * 8):
+            width (`int`, *optional*, defaults to self.image_unet.config.sample_size * self.vae_scale_factor):
                 The width in pixels of the generated image.
             num_inference_steps (`int`, *optional*, defaults to 50):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
@@ -444,8 +445,8 @@ class VersatileDiffusionTextToImagePipeline(DiffusionPipeline):
             (nsfw) content, according to the `safety_checker`.
         """
         # 0. Default height and width to unet
-        height = height or self.image_unet.config.sample_size * 8
-        width = width or self.image_unet.config.sample_size * 8
+        height = height or self.image_unet.config.sample_size * self.vae_scale_factor
+        width = width or self.image_unet.config.sample_size * self.vae_scale_factor
 
         # 1. Check inputs. Raise error if not correct
         self.check_inputs(prompt, height, width, callback_steps)

tests/pipelines/altdiffusion/test_alt_diffusion.py

@@ -172,7 +172,9 @@ class AltDiffusionPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         image_from_tuple_slice = image_from_tuple[0, -3:, -3:, -1]
 
         assert image.shape == (1, 64, 64, 3)
-        expected_slice = np.array([0.5748162, 0.60447145, 0.48821217, 0.50100636, 0.5431185, 0.45763683, 0.49657696, 0.48132733, 0.47573093])
+        expected_slice = np.array(
+            [0.5748162, 0.60447145, 0.48821217, 0.50100636, 0.5431185, 0.45763683, 0.49657696, 0.48132733, 0.47573093]
+        )
 
         assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
         assert np.abs(image_from_tuple_slice.flatten() - expected_slice).max() < 1e-2
@@ -219,7 +221,9 @@ class AltDiffusionPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         image_from_tuple_slice = image_from_tuple[0, -3:, -3:, -1]
 
         assert image.shape == (1, 64, 64, 3)
-        expected_slice = np.array([0.51605093, 0.5707241, 0.47365507, 0.50578886, 0.5633877, 0.4642503, 0.5182081, 0.48763484, 0.49084237])
+        expected_slice = np.array(
+            [0.51605093, 0.5707241, 0.47365507, 0.50578886, 0.5633877, 0.4642503, 0.5182081, 0.48763484, 0.49084237]
+        )
         assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
         assert np.abs(image_from_tuple_slice.flatten() - expected_slice).max() < 1e-2
 

tests/pipelines/latent_diffusion/test_latent_diffusion.py

@@ -111,8 +111,8 @@ class LDMTextToImagePipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         image_slice = image[0, -3:, -3:, -1]
         image_from_tuple_slice = image_from_tuple[0, -3:, -3:, -1]
 
-        assert image.shape == (1, 64, 64, 3)
-        expected_slice = np.array([0.5074, 0.5026, 0.4998, 0.4056, 0.3523, 0.4649, 0.5289, 0.5299, 0.4897])
+        assert image.shape == (1, 16, 16, 3)
+        expected_slice = np.array([0.6806, 0.5454, 0.5638, 0.4893, 0.4656, 0.4257, 0.6248, 0.5217, 0.5498])
         assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
         assert np.abs(image_from_tuple_slice.flatten() - expected_slice).max() < 1e-2