Skip to content

GitLab

Unverified Commit 05a36d5c authored by Patrick von Platen's avatar Patrick von Platen Committed by GitHub
Browse files

Upscaling fixed (#1402) 

* Upscaling fixed

* up

* more fixes

* fix

* more fixes

* finish again

* up
parent cbfed0c2
Hide whitespace changes
Inline Side-by-side
Showing with 124 additions and 86 deletions
src/diffusers/pipeline_utils.py
View file @ 05a36d5c
......@@ -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
View file @ 05a36d5c
......@@ -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
View file @ 05a36d5c
......@@ -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
View file @ 05a36d5c
......@@ -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
View file @ 05a36d5c
......@@ -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
View file @ 05a36d5c
......@@ -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
View file @ 05a36d5c
......@@ -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
View file @ 05a36d5c
......@@ -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
View file @ 05a36d5c
......@@ -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
View file @ 05a36d5c
......@@ -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
View file @ 05a36d5c
......@@ -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
View file @ 05a36d5c
......@@ -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
View file @ 05a36d5c
......@@ -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
View file @ 05a36d5c
......@@ -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
View file @ 05a36d5c
......@@ -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
View file @ 05a36d5c
......@@ -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
View file @ 05a36d5c
......@@ -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
View file @ 05a36d5c
......@@ -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
View file @ 05a36d5c
......@@ -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
View file @ 05a36d5c
......@@ -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
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment