Update pipeline.py

pipeline.py

@@ -6,11 +6,10 @@ import numpy as np
 import torch
 
 import PIL
-from diffusers.onnx_utils import OnnxRuntimeModel
-from diffusers.pipeline_utils import DiffusionPipeline
+from diffusers import OnnxStableDiffusionPipeline, SchedulerMixin
+from diffusers.onnx_utils import ORT_TO_NP_TYPE, OnnxRuntimeModel
 from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
-from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
-from diffusers.utils import logging
+from diffusers.utils import PIL_INTERPOLATION, deprecate, logging
 from transformers import CLIPFeatureExtractor, CLIPTokenizer
 
 
@@ -38,7 +37,7 @@ re_attention = re.compile(
 
 def parse_prompt_attention(text):
     """
-    Parses a string with attention tokens and returns a list of pairs: text and its assoicated weight.
+    Parses a string with attention tokens and returns a list of pairs: text and its associated weight.
     Accepted tokens are:
       (abc) - increases attention to abc by a multiplier of 1.1
       (abc:3.12) - increases attention to abc by a multiplier of 3.12
@@ -236,12 +235,12 @@ def get_weighted_text_embeddings(
     r"""
     Prompts can be assigned with local weights using brackets. For example,
     prompt 'A (very beautiful) masterpiece' highlights the words 'very beautiful',
-    and the embedding tokens corresponding to the words get multipled by a constant, 1.1.
+    and the embedding tokens corresponding to the words get multiplied by a constant, 1.1.
 
-    Also, to regularize of the embedding, the weighted embedding would be scaled to preserve the origional mean.
+    Also, to regularize of the embedding, the weighted embedding would be scaled to preserve the original mean.
 
     Args:
-        pipe (`DiffusionPipeline`):
+        pipe (`OnnxStableDiffusionPipeline`):
             Pipe to provide access to the tokenizer and the text encoder.
         prompt (`str` or `List[str]`):
             The prompt or prompts to guide the image generation.
@@ -365,17 +364,17 @@ def get_weighted_text_embeddings(
 def preprocess_image(image):
     w, h = image.size
     w, h = map(lambda x: x - x % 32, (w, h))  # resize to integer multiple of 32
-    image = image.resize((w, h), resample=PIL.Image.LANCZOS)
+    image = image.resize((w, h), resample=PIL_INTERPOLATION["lanczos"])
     image = np.array(image).astype(np.float32) / 255.0
     image = image[None].transpose(0, 3, 1, 2)
     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?
@@ -383,7 +382,7 @@ def preprocess_mask(mask):
     return mask
 
 
-class OnnxStableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
+class OnnxStableDiffusionLongPromptWeightingPipeline(OnnxStableDiffusionPipeline):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion without tokens length limit, and support parsing
     weighting in prompt.
@@ -399,12 +398,12 @@ class OnnxStableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
         text_encoder: OnnxRuntimeModel,
         tokenizer: CLIPTokenizer,
         unet: OnnxRuntimeModel,
-        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+        scheduler: SchedulerMixin,
         safety_checker: OnnxRuntimeModel,
         feature_extractor: CLIPFeatureExtractor,
+        requires_safety_checker: bool = True,
     ):
-        super().__init__()
-        self.register_modules(
+        super().__init__(
             vae_encoder=vae_encoder,
             vae_decoder=vae_decoder,
             text_encoder=text_encoder,
@@ -413,14 +412,177 @@ class OnnxStableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
             scheduler=scheduler,
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
+            requires_safety_checker=requires_safety_checker,
         )
+        self.unet_in_channels = 4
+        self.vae_scale_factor = 8
+
+    def _encode_prompt(
+        self,
+        prompt,
+        num_images_per_prompt,
+        do_classifier_free_guidance,
+        negative_prompt,
+        max_embeddings_multiples,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `list(int)`):
+                prompt to be encoded
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_prompt (`str` or `List[str]`):
+                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+                if `guidance_scale` is less than `1`).
+            max_embeddings_multiples (`int`, *optional*, defaults to `3`):
+                The max multiple length of prompt embeddings compared to the max output length of text encoder.
+        """
+        batch_size = len(prompt) if isinstance(prompt, list) else 1
+
+        if negative_prompt is None:
+            negative_prompt = [""] * batch_size
+        elif isinstance(negative_prompt, str):
+            negative_prompt = [negative_prompt] * batch_size
+        if batch_size != len(negative_prompt):
+            raise ValueError(
+                f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                " the batch size of `prompt`."
+            )
+
+        text_embeddings, uncond_embeddings = get_weighted_text_embeddings(
+            pipe=self,
+            prompt=prompt,
+            uncond_prompt=negative_prompt if do_classifier_free_guidance else None,
+            max_embeddings_multiples=max_embeddings_multiples,
+        )
+
+        text_embeddings = text_embeddings.repeat(num_images_per_prompt, 0)
+        if do_classifier_free_guidance:
+            uncond_embeddings = uncond_embeddings.repeat(num_images_per_prompt, 0)
+            text_embeddings = np.concatenate([uncond_embeddings, text_embeddings])
+
+        return text_embeddings
+
+    def check_inputs(self, prompt, height, width, strength, callback_steps):
+        if not isinstance(prompt, str) and not isinstance(prompt, list):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if strength < 0 or strength > 1:
+            raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
+
+        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}.")
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+    def get_timesteps(self, num_inference_steps, strength, is_text2img):
+        if is_text2img:
+            return self.scheduler.timesteps, num_inference_steps
+        else:
+            # get the original timestep using init_timestep
+            offset = self.scheduler.config.get("steps_offset", 0)
+            init_timestep = int(num_inference_steps * strength) + offset
+            init_timestep = min(init_timestep, num_inference_steps)
+
+            t_start = max(num_inference_steps - init_timestep + offset, 0)
+            timesteps = self.scheduler.timesteps[t_start:]
+            return timesteps, num_inference_steps - t_start
+
+    def run_safety_checker(self, image):
+        if self.safety_checker is not None:
+            safety_checker_input = self.feature_extractor(
+                self.numpy_to_pil(image), return_tensors="np"
+            ).pixel_values.astype(image.dtype)
+            # There will throw an error if use safety_checker directly and batchsize>1
+            images, has_nsfw_concept = [], []
+            for i in range(image.shape[0]):
+                image_i, has_nsfw_concept_i = self.safety_checker(
+                    clip_input=safety_checker_input[i : i + 1], images=image[i : i + 1]
+                )
+                images.append(image_i)
+                has_nsfw_concept.append(has_nsfw_concept_i[0])
+            image = np.concatenate(images)
+        else:
+            has_nsfw_concept = None
+        return image, has_nsfw_concept
+
+    def decode_latents(self, latents):
+        latents = 1 / 0.18215 * latents
+        # image = self.vae_decoder(latent_sample=latents)[0]
+        # it seems likes there is a strange result for using half-precision vae decoder if batchsize>1
+        image = np.concatenate(
+            [self.vae_decoder(latent_sample=latents[i : i + 1])[0] for i in range(latents.shape[0])]
+        )
+        image = np.clip(image / 2 + 0.5, 0, 1)
+        image = image.transpose((0, 2, 3, 1))
+        return image
+
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def prepare_latents(self, image, timestep, batch_size, height, width, dtype, generator, latents=None):
+        if image is None:
+            shape = (
+                batch_size,
+                self.unet_in_channels,
+                height // self.vae_scale_factor,
+                width // self.vae_scale_factor,
+            )
+
+            if latents is None:
+                latents = torch.randn(shape, generator=generator, device="cpu").numpy().astype(dtype)
+            else:
+                if latents.shape != shape:
+                    raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+
+            # scale the initial noise by the standard deviation required by the scheduler
+            latents = (torch.from_numpy(latents) * self.scheduler.init_noise_sigma).numpy()
+            return latents, None, None
+        else:
+            init_latents = self.vae_encoder(sample=image)[0]
+            init_latents = 0.18215 * init_latents
+            init_latents = np.concatenate([init_latents] * batch_size, axis=0)
+            init_latents_orig = init_latents
+            shape = init_latents.shape
+
+            # add noise to latents using the timesteps
+            noise = torch.randn(shape, generator=generator, device="cpu").numpy().astype(dtype)
+            latents = self.scheduler.add_noise(
+                torch.from_numpy(init_latents), torch.from_numpy(noise), timestep
+            ).numpy()
+            return latents, init_latents_orig, noise
 
     @torch.no_grad()
     def __call__(
         self,
         prompt: Union[str, List[str]],
         negative_prompt: Optional[Union[str, List[str]]] = None,
-        init_image: Union[np.ndarray, PIL.Image.Image] = None,
+        image: Union[np.ndarray, PIL.Image.Image] = None,
         mask_image: Union[np.ndarray, PIL.Image.Image] = None,
         height: int = 512,
         width: int = 512,
@@ -429,12 +591,13 @@ class OnnxStableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
         strength: float = 0.8,
         num_images_per_prompt: Optional[int] = 1,
         eta: float = 0.0,
-        generator: Optional[np.random.RandomState] = None,
+        generator: Optional[torch.Generator] = None,
         latents: Optional[np.ndarray] = None,
         max_embeddings_multiples: Optional[int] = 3,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
         callback: Optional[Callable[[int, int, np.ndarray], None]] = None,
+        is_cancelled_callback: Optional[Callable[[], bool]] = None,
         callback_steps: Optional[int] = 1,
         **kwargs,
     ):
@@ -447,11 +610,11 @@ class OnnxStableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
             negative_prompt (`str` or `List[str]`, *optional*):
                 The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
                 if `guidance_scale` is less than `1`).
-            init_image (`np.ndarray` or `PIL.Image.Image`):
+            image (`np.ndarray` or `PIL.Image.Image`):
                 `Image`, or tensor representing an image batch, that will be used as the starting point for the
                 process.
             mask_image (`np.ndarray` or `PIL.Image.Image`):
-                `Image`, or tensor representing an image batch, to mask `init_image`. White pixels in the mask will be
+                `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be
                 replaced by noise and therefore 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)`.
@@ -469,18 +632,19 @@ class OnnxStableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
                 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
                 usually at the expense of lower image quality.
             strength (`float`, *optional*, defaults to 0.8):
-                Conceptually, indicates how much to transform the reference `init_image`. Must be between 0 and 1.
-                `init_image` will be used as a starting point, adding more noise to it the larger the `strength`. The
+                Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1.
+                `image` will be used as a starting point, adding more noise to it the larger the `strength`. The
                 number of denoising steps depends on the amount of noise initially added. When `strength` is 1, added
                 noise will be maximum and the denoising process will run for the full number of iterations specified in
-                `num_inference_steps`. A value of 1, therefore, essentially ignores `init_image`.
+                `num_inference_steps`. A value of 1, therefore, essentially ignores `image`.
             num_images_per_prompt (`int`, *optional*, defaults to 1):
                 The number of images to generate per prompt.
             eta (`float`, *optional*, defaults to 0.0):
                 Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
                 [`schedulers.DDIMScheduler`], will be ignored for others.
-            generator (`np.random.RandomState`, *optional*):
-                A np.random.RandomState to make generation deterministic.
+            generator (`torch.Generator`, *optional*):
+                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
+                deterministic.
             latents (`np.ndarray`, *optional*):
                 Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
                 generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
@@ -496,213 +660,142 @@ class OnnxStableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
             callback (`Callable`, *optional*):
                 A function that will be called every `callback_steps` steps during inference. The function will be
                 called with the following arguments: `callback(step: int, timestep: int, latents: np.ndarray)`.
+            is_cancelled_callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. If the function returns
+                `True`, the inference will be cancelled.
             callback_steps (`int`, *optional*, defaults to 1):
                 The frequency at which the `callback` function will be called. If not specified, the callback will be
                 called at every step.
 
         Returns:
+            `None` if cancelled by `is_cancelled_callback`,
             [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
             [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
             When returning a tuple, the first element is a list with the generated images, and the second element is a
             list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
             (nsfw) content, according to the `safety_checker`.
         """
+        message = "Please use `image` instead of `init_image`."
+        init_image = deprecate("init_image", "0.12.0", message, take_from=kwargs)
+        image = init_image or image
 
-        if isinstance(prompt, str):
-            batch_size = 1
-            prompt = [prompt]
-        elif isinstance(prompt, list):
-            batch_size = len(prompt)
-        else:
-            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
-
-        if strength < 0 or strength > 1:
-            raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
-
-        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}.")
-
-        if (callback_steps is None) or (
-            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
-        ):
-            raise ValueError(
-                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
-                f" {type(callback_steps)}."
-            )
+        # 0. Default height and width to unet
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
 
-        # get prompt text embeddings
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(prompt, height, width, strength, callback_steps)
 
+        # 2. Define call parameters
+        batch_size = 1 if isinstance(prompt, str) else len(prompt)
         # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
         # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
         # corresponds to doing no classifier free guidance.
         do_classifier_free_guidance = guidance_scale > 1.0
-        # get unconditional embeddings for classifier free guidance
-        if negative_prompt is None:
-            negative_prompt = [""] * batch_size
-        elif isinstance(negative_prompt, str):
-            negative_prompt = [negative_prompt] * batch_size
-        if batch_size != len(negative_prompt):
-            raise ValueError(
-                f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
-                f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
-                " the batch size of `prompt`."
-            )
 
-        if generator is None:
-            generator = np.random
-
-        text_embeddings, uncond_embeddings = get_weighted_text_embeddings(
-            pipe=self,
-            prompt=prompt,
-            uncond_prompt=negative_prompt if do_classifier_free_guidance else None,
-            max_embeddings_multiples=max_embeddings_multiples,
-            **kwargs,
+        # 3. Encode input prompt
+        text_embeddings = self._encode_prompt(
+            prompt,
+            num_images_per_prompt,
+            do_classifier_free_guidance,
+            negative_prompt,
+            max_embeddings_multiples,
         )
-
-        text_embeddings = text_embeddings.repeat(num_images_per_prompt, 0)
-        if do_classifier_free_guidance:
-            uncond_embeddings = uncond_embeddings.repeat(num_images_per_prompt, 0)
-            text_embeddings = np.concatenate([uncond_embeddings, text_embeddings])
-
-        # set timesteps
-        self.scheduler.set_timesteps(num_inference_steps)
-
-        latents_dtype = text_embeddings.dtype
-        init_latents_orig = None
-        mask = None
-        noise = None
-
-        if init_image is None:
-            latents_shape = (
-                batch_size * num_images_per_prompt,
-                4,
-                height // 8,
-                width // 8,
-            )
-
-            if latents is None:
-                latents = generator.randn(*latents_shape).astype(latents_dtype)
-            elif latents.shape != latents_shape:
-                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
-
-            timesteps = self.scheduler.timesteps.to(self.device)
-
-            # scale the initial noise by the standard deviation required by the scheduler
-            latents = latents * self.scheduler.init_noise_sigma
+        dtype = text_embeddings.dtype
+
+        # 4. Preprocess image and mask
+        if isinstance(image, PIL.Image.Image):
+            image = preprocess_image(image)
+        if image is not None:
+            image = image.astype(dtype)
+        if isinstance(mask_image, PIL.Image.Image):
+            mask_image = preprocess_mask(mask_image, self.vae_scale_factor)
+        if mask_image is not None:
+            mask = mask_image.astype(dtype)
+            mask = np.concatenate([mask] * batch_size * num_images_per_prompt)
         else:
-            if isinstance(init_image, PIL.Image.Image):
-                init_image = preprocess_image(init_image)
-            # encode the init image into latents and scale the latents
-            init_image = init_image.astype(latents_dtype)
-            init_latents = self.vae_encoder(sample=init_image)[0]
-            init_latents = 0.18215 * init_latents
-            init_latents = np.concatenate([init_latents] * batch_size * num_images_per_prompt)
-            init_latents_orig = init_latents
-
-            # preprocess mask
-            if mask_image is not None:
-                if isinstance(mask_image, PIL.Image.Image):
-                    mask_image = preprocess_mask(mask_image)
-                mask_image = mask_image.astype(latents_dtype)
-                mask = np.concatenate([mask_image] * batch_size * num_images_per_prompt)
+            mask = None
 
-                # check sizes
-                if not mask.shape == init_latents.shape:
-                    print(mask.shape, init_latents.shape)
-                    raise ValueError("The mask and init_image should be the same size!")
-
-            # get the original timestep using init_timestep
-            offset = self.scheduler.config.get("steps_offset", 0)
-            init_timestep = int(num_inference_steps * strength) + offset
-            init_timestep = min(init_timestep, num_inference_steps)
-
-            timesteps = self.scheduler.timesteps[-init_timestep]
-            timesteps = torch.tensor([timesteps] * batch_size * num_images_per_prompt)
-
-            # add noise to latents using the timesteps
-            noise = generator.randn(*init_latents.shape).astype(latents_dtype)
-            latents = self.scheduler.add_noise(
-                torch.from_numpy(init_latents), torch.from_numpy(noise), timesteps
-            ).numpy()
-
-            t_start = max(num_inference_steps - init_timestep + offset, 0)
-            timesteps = self.scheduler.timesteps[t_start:]
-
-        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
-        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
-        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
-        # and should be between [0, 1]
-        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
-        extra_step_kwargs = {}
-        if accepts_eta:
-            extra_step_kwargs["eta"] = eta
-
-        for i, t in enumerate(self.progress_bar(timesteps)):
-            # expand the latents if we are doing classifier free guidance
-            latent_model_input = np.concatenate([latents] * 2) if do_classifier_free_guidance else latents
-            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
-
-            # predict the noise residual
-            noise_pred = self.unet(
-                sample=latent_model_input,
-                timestep=np.array([t]),
-                encoder_hidden_states=text_embeddings,
-            )
-            noise_pred = noise_pred[0]
-
-            # perform guidance
-            if do_classifier_free_guidance:
-                noise_pred_uncond, noise_pred_text = np.split(noise_pred, 2)
-                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
-
-            # compute the previous noisy sample x_t -> x_t-1
-            latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample.numpy()
-
-            if mask is not None:
-                # masking
-                init_latents_proper = self.scheduler.add_noise(
-                    torch.from_numpy(init_latents_orig),
-                    torch.from_numpy(noise),
-                    torch.tensor([t]),
-                ).numpy()
-                latents = (init_latents_proper * mask) + (latents * (1 - mask))
-
-            # call the callback, if provided
-            if callback is not None and i % callback_steps == 0:
-                callback(i, t, latents)
+        # 5. set timesteps
+        self.scheduler.set_timesteps(num_inference_steps)
+        timestep_dtype = next(
+            (input.type for input in self.unet.model.get_inputs() if input.name == "timestep"), "tensor(float)"
+        )
+        timestep_dtype = ORT_TO_NP_TYPE[timestep_dtype]
+        timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, image is None)
+        latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+
+        # 6. Prepare latent variables
+        latents, init_latents_orig, noise = self.prepare_latents(
+            image,
+            latent_timestep,
+            batch_size * num_images_per_prompt,
+            height,
+            width,
+            dtype,
+            generator,
+            latents,
+        )
 
-        latents = 1 / 0.18215 * latents
-        # image = self.vae_decoder(latent_sample=latents)[0]
-        # it seems likes there is a problem for using half-precision vae decoder if batchsize>1
-        image = []
-        for i in range(latents.shape[0]):
-            image.append(self.vae_decoder(latent_sample=latents[i : i + 1])[0])
-        image = np.concatenate(image)
+        # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 8. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = np.concatenate([latents] * 2) if do_classifier_free_guidance else latents
+                latent_model_input = self.scheduler.scale_model_input(torch.from_numpy(latent_model_input), t)
+                latent_model_input = latent_model_input.numpy()
+
+                # predict the noise residual
+                noise_pred = self.unet(
+                    sample=latent_model_input,
+                    timestep=np.array([t], dtype=timestep_dtype),
+                    encoder_hidden_states=text_embeddings,
+                )
+                noise_pred = noise_pred[0]
 
-        image = np.clip(image / 2 + 0.5, 0, 1)
-        image = image.transpose((0, 2, 3, 1))
+                # perform guidance
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = np.split(noise_pred, 2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
 
-        if self.safety_checker is not None:
-            safety_checker_input = self.feature_extractor(
-                self.numpy_to_pil(image), return_tensors="np"
-            ).pixel_values.astype(image.dtype)
-            # There will throw an error if use safety_checker directly and batchsize>1
-            images, has_nsfw_concept = [], []
-            for i in range(image.shape[0]):
-                image_i, has_nsfw_concept_i = self.safety_checker(
-                    clip_input=safety_checker_input[i : i + 1], images=image[i : i + 1]
+                # compute the previous noisy sample x_t -> x_t-1
+                scheduler_output = self.scheduler.step(
+                    torch.from_numpy(noise_pred), t, torch.from_numpy(latents), **extra_step_kwargs
                 )
-                images.append(image_i)
-                has_nsfw_concept.append(has_nsfw_concept_i[0])
-            image = np.concatenate(images)
-        else:
-            has_nsfw_concept = None
-
+                latents = scheduler_output.prev_sample.numpy()
+
+                if mask is not None:
+                    # masking
+                    init_latents_proper = self.scheduler.add_noise(
+                        torch.from_numpy(init_latents_orig),
+                        torch.from_numpy(noise),
+                        t,
+                    ).numpy()
+                    latents = (init_latents_proper * mask) + (latents * (1 - mask))
+
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+                    if i % callback_steps == 0:
+                        if callback is not None:
+                            callback(i, t, latents)
+                        if is_cancelled_callback is not None and is_cancelled_callback():
+                            return None
+        # 9. Post-processing
+        image = self.decode_latents(latents)
+
+        # 10. Run safety checker
+        image, has_nsfw_concept = self.run_safety_checker(image)
+
+        # 11. Convert to PIL
         if output_type == "pil":
             image = self.numpy_to_pil(image)
 
         if not return_dict:
-            return (image, has_nsfw_concept)
+            return image, has_nsfw_concept
 
         return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
 
@@ -716,7 +809,7 @@ class OnnxStableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
         guidance_scale: float = 7.5,
         num_images_per_prompt: Optional[int] = 1,
         eta: float = 0.0,
-        generator: Optional[np.random.RandomState] = None,
+        generator: Optional[torch.Generator] = None,
         latents: Optional[np.ndarray] = None,
         max_embeddings_multiples: Optional[int] = 3,
         output_type: Optional[str] = "pil",
@@ -751,8 +844,9 @@ class OnnxStableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
             eta (`float`, *optional*, defaults to 0.0):
                 Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
                 [`schedulers.DDIMScheduler`], will be ignored for others.
-            generator (`np.random.RandomState`, *optional*):
-                A np.random.RandomState to make generation deterministic.
+            generator (`torch.Generator`, *optional*):
+                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
+                deterministic.
             latents (`np.ndarray`, *optional*):
                 Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
                 generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
@@ -799,7 +893,7 @@ class OnnxStableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
 
     def img2img(
         self,
-        init_image: Union[np.ndarray, PIL.Image.Image],
+        image: Union[np.ndarray, PIL.Image.Image],
         prompt: Union[str, List[str]],
         negative_prompt: Optional[Union[str, List[str]]] = None,
         strength: float = 0.8,
@@ -807,7 +901,7 @@ class OnnxStableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
         guidance_scale: Optional[float] = 7.5,
         num_images_per_prompt: Optional[int] = 1,
         eta: Optional[float] = 0.0,
-        generator: Optional[np.random.RandomState] = None,
+        generator: Optional[torch.Generator] = None,
         max_embeddings_multiples: Optional[int] = 3,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
@@ -818,7 +912,7 @@ class OnnxStableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
         r"""
         Function for image-to-image generation.
         Args:
-            init_image (`np.ndarray` or `PIL.Image.Image`):
+            image (`np.ndarray` or `PIL.Image.Image`):
                 `Image`, or ndarray representing an image batch, that will be used as the starting point for the
                 process.
             prompt (`str` or `List[str]`):
@@ -827,11 +921,11 @@ class OnnxStableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
                 The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
                 if `guidance_scale` is less than `1`).
             strength (`float`, *optional*, defaults to 0.8):
-                Conceptually, indicates how much to transform the reference `init_image`. Must be between 0 and 1.
-                `init_image` will be used as a starting point, adding more noise to it the larger the `strength`. The
+                Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1.
+                `image` will be used as a starting point, adding more noise to it the larger the `strength`. The
                 number of denoising steps depends on the amount of noise initially added. When `strength` is 1, added
                 noise will be maximum and the denoising process will run for the full number of iterations specified in
-                `num_inference_steps`. A value of 1, therefore, essentially ignores `init_image`.
+                `num_inference_steps`. A value of 1, therefore, essentially ignores `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
                 expense of slower inference. This parameter will be modulated by `strength`.
@@ -846,8 +940,9 @@ class OnnxStableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
             eta (`float`, *optional*, defaults to 0.0):
                 Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
                 [`schedulers.DDIMScheduler`], will be ignored for others.
-            generator (`np.random.RandomState`, *optional*):
-                A np.random.RandomState to make generation deterministic.
+            generator (`torch.Generator`, *optional*):
+                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
+                deterministic.
             max_embeddings_multiples (`int`, *optional*, defaults to `3`):
                 The max multiple length of prompt embeddings compared to the max output length of text encoder.
             output_type (`str`, *optional*, defaults to `"pil"`):
@@ -872,7 +967,7 @@ class OnnxStableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
         return self.__call__(
             prompt=prompt,
             negative_prompt=negative_prompt,
-            init_image=init_image,
+            image=image,
             num_inference_steps=num_inference_steps,
             guidance_scale=guidance_scale,
             strength=strength,
@@ -889,7 +984,7 @@ class OnnxStableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
 
     def inpaint(
         self,
-        init_image: Union[np.ndarray, PIL.Image.Image],
+        image: Union[np.ndarray, PIL.Image.Image],
         mask_image: Union[np.ndarray, PIL.Image.Image],
         prompt: Union[str, List[str]],
         negative_prompt: Optional[Union[str, List[str]]] = None,
@@ -898,7 +993,7 @@ class OnnxStableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
         guidance_scale: Optional[float] = 7.5,
         num_images_per_prompt: Optional[int] = 1,
         eta: Optional[float] = 0.0,
-        generator: Optional[np.random.RandomState] = None,
+        generator: Optional[torch.Generator] = None,
         max_embeddings_multiples: Optional[int] = 3,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
@@ -909,11 +1004,11 @@ class OnnxStableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
         r"""
         Function for inpaint.
         Args:
-            init_image (`np.ndarray` or `PIL.Image.Image`):
+            image (`np.ndarray` or `PIL.Image.Image`):
                 `Image`, or tensor representing an image batch, that will be used as the starting point for the
                 process. This is the image whose masked region will be inpainted.
             mask_image (`np.ndarray` or `PIL.Image.Image`):
-                `Image`, or tensor representing an image batch, to mask `init_image`. White pixels in the mask will be
+                `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be
                 replaced by noise and therefore 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)`.
@@ -925,7 +1020,7 @@ class OnnxStableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
             strength (`float`, *optional*, defaults to 0.8):
                 Conceptually, indicates how much to inpaint the masked area. Must be between 0 and 1. When `strength`
                 is 1, the denoising process will be run on the masked area for the full number of iterations specified
-                in `num_inference_steps`. `init_image` will be used as a reference for the masked area, adding more
+                in `num_inference_steps`. `image` will be used as a reference for the masked area, adding more
                 noise to that region the larger the `strength`. If `strength` is 0, no inpainting will occur.
             num_inference_steps (`int`, *optional*, defaults to 50):
                 The reference number of denoising steps. More denoising steps usually lead to a higher quality image at
@@ -941,8 +1036,9 @@ class OnnxStableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
             eta (`float`, *optional*, defaults to 0.0):
                 Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
                 [`schedulers.DDIMScheduler`], will be ignored for others.
-            generator (`np.random.RandomState`, *optional*):
-                A np.random.RandomState to make generation deterministic.
+            generator (`torch.Generator`, *optional*):
+                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
+                deterministic.
             max_embeddings_multiples (`int`, *optional*, defaults to `3`):
                 The max multiple length of prompt embeddings compared to the max output length of text encoder.
             output_type (`str`, *optional*, defaults to `"pil"`):
@@ -967,7 +1063,7 @@ class OnnxStableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
         return self.__call__(
             prompt=prompt,
             negative_prompt=negative_prompt,
-            init_image=init_image,
+            image=image,
             mask_image=mask_image,
             num_inference_steps=num_inference_steps,
             guidance_scale=guidance_scale,