feat: inital commit

latent_space_explorer/__init__.py

@@ -0,0 +1,434 @@
+from typing import List
+
+from dataclasses import dataclass
+from .fast_sd import fast_diffusion_pipeline
+
+import torch
+import pygame
+import numpy as np
+import time
+from PIL import Image
+
+from .game_objects import Point, TextPrompt
+from .sampling import (
+    DistanceSampling,
+    CircleSampling
+)
+
+@dataclass
+class GameConfig:
+    point_thickness : float = 10 # Thickness for each point
+    zoom_speed : float = 0.75 # How fast we zoom in or out
+    move_speed : float = 0.75 # How fast we move around canvas
+    point_font_size : int = 25 # Size of fonts for points on screen
+
+    prompt_font_size : int = 30 # Size of font for prompt on screen
+
+    # screen size
+    width : int = 1920
+    height : int = 1080
+
+    # size of sample in top left
+    sample_width : int = 512
+    sample_height : int = 512
+
+    compile : bool = False # compile the sd model with torch.compile?
+    sampler : str = "distance" # "distance" or "circle"
+    seed : int = 0 # Seed for initial latent noise
+    call_every : int = 90 # Only calls draw function every *this many* ms. This is to prevent lag. Set this to be around the latency of the model
+
+class LatentSpaceExplorer:
+    def __init__(self, config : GameConfig = GameConfig()):
+        self.config = config
+
+        self.pipe = fast_diffusion_pipeline(compile = self.config.compile)
+        self.points : List[Point] = []
+        self.player_pos = None # [2,] np array in R2 space
+
+        self.dragging_point_idx = None
+        self.selected_point_idx = None
+
+        self.zoom_level = 300.0
+        self.translation = np.array([-self.config.width/2, -self.config.height/2])
+
+        self.point_kwargs = {}
+        if self.config.sampler == "distance":
+            self.sampler = DistanceSampling
+        elif self.config.sampler == "circle":
+            self.sampler = CircleSampling
+            self.point_kwargs['on_edge'] = True
+        else:
+            raise ValueError(f"Invalid sampler choice: {self.config.sampler}")
+
+        pygame.init()
+        self.screen = pygame.display.set_mode((self.config.width, self.config.height))
+        self.clock = pygame.time.Clock()
+        self.ms_elapsed = 0
+
+        # (n_samples, running average)
+        self.avg_latency = (0, 0) # Track average latency of generation for debug
+        
+        self.sample_image = None
+        self.sample_font = pygame.font.Font(None, self.config.point_font_size)
+        
+        # User input
+        self.input_font = pygame.font.Font(None, self.config.prompt_font_size)
+        self.inputting_text = False
+        self.inputting_text_for = None # oneof ["modify", "add"]
+        self.text_prompt : TextPrompt = None
+    
+    def tick(self):
+        self.clock.tick()
+        self.ms_elapsed += self.clock.get_time()
+
+    def update_latency(self, new_observation):
+        n = self.avg_latency[0]
+        old_avg = self.avg_latency[1]
+        self.avg_latency = (n + 1, (old_avg * n + new_observation) / (n + 1))
+
+    def create_text_prompt(self, prompt_text):
+        self.text_prompt = TextPrompt(prompt_text, self.input_font, self.screen)
+
+    def switch_sampler(self):
+        if self.config.sampler == "distance":
+            self.config.sampler = "circle"
+            self.sampler = CircleSampling
+            self.point_kwargs = {'on_edge' : True}
+        elif self.config.sampler == "circle":
+            self.config.sampler = "distance"
+            self.sampler = DistanceSampling
+            self.point_kwargs = {}
+        self.set_prompts(self.prompts, reset = True)
+    
+    @property
+    def encodes(self):
+        """
+        Get encodings directly from points as a tuple with batched encodings
+        """
+        if not self.points:
+            return None
+        encode_list = [p.encoding for p in self.points] # list of N-tuples
+        n = len(encode_list[0])
+        res = []
+        for i in range(n):
+            res.append(torch.cat([e[i] for e in encode_list], dim = 0) if encode_list[0][i] is not None else None)
+
+        return tuple(res)
+    
+    @property
+    def prompts(self):
+        """
+        Get a list of current prompts
+        """
+        return [p.text for p in self.points]
+    
+    @property
+    def r2_points(self):
+        """
+        Get all points in terms of R2 space
+        """
+        points = [np.array(p.xy_pos) for p in self.points]
+        points = np.stack(points, axis = 0) # [n, 2]
+        return points
+
+    @property
+    def screen_space_points(self):
+        """
+        Get all points in terms of screen space
+        """
+        screen_space = (self.r2_points * self.zoom_level) - self.translation[None,:]
+        return screen_space # list of points in scren space
+    
+    @property
+    def mouse_pos(self):
+        return np.array(pygame.mouse.get_pos())
+
+    def invert_screen_space(self, point):
+        """
+        taking position as [2,] np array in screen space, return R2 pos
+        """
+        return (point + self.translation) / self.zoom_level
+
+    def screen_space(self, point):
+        """
+        R2 -> screenspace as [2,] array
+        """
+        return (point * self.zoom_level) - self.translation
+    
+    def fixed_seed(self):
+        """
+        Controls random number generator for initial latent noise
+        """
+        return torch.Generator('cuda').manual_seed(self.config.seed)
+    
+    def get_encodes(self, text):
+        """
+        Get text encodings for some prompt then split them so we can associate points with thier encodings
+        """
+        encodes = self.pipe.get_encodes(text, generator = self.fixed_seed())
+        # (n-tuple of lists) into (list of n-tuples)
+        if not isinstance(encodes, tuple) and not isinstance(encodes, list):
+            return encodes # Already a tensor, no problem
+        
+        res_list = []
+        for i in range(len(encodes[0])):
+            res_list_i = [encodes_j[i].unsqueeze(0) if encodes_j is not None else None for encodes_j in encodes]
+            res_list.append(tuple(res_list_i))
+
+        return res_list
+    
+    def draw_sample(self):
+        """
+        Draw sample with current points and player position
+        """
+        if self.player_pos is not None and self.encodes is not None:
+            if self.ms_elapsed >= self.config.call_every:
+                time_start = time.time()
+                encoding = self.sampler(self.encodes)(self.player_pos, self.r2_points)
+                self.sample_image = self.pipe.generate_from_encodes(encoding, generator = self.fixed_seed()).images[0]
+                time_total = float(time.time() - time_start) * 1000 # s -> ms
+
+                self.update_latency(time_total)
+
+                self.ms_elapsed = 0
+
+    def get_player_pos_r2(self):
+        """
+        Get player position in R2 from the 
+        """
+        self.player_pos = self.invert_screen_space(self.mouse_pos)
+
+    def get_player_pos_screenspace(self):
+        """
+        Get player pos in screen space
+        """
+        if self.player_pos is not None: return self.screen_space(self.player_pos)
+    
+    def detect_mouse_on_point(self):
+        """
+        Detect if mouse is currently in a point. If so, returns index of point, otherwise returns none.
+        """
+        if not self.points:
+            return None
+        
+        mouse_pos = self.mouse_pos
+        points = self.screen_space_points
+
+        distances = np.linalg.norm(points - mouse_pos[None,:], axis = 1)
+        close_idx = np.argmin(distances)
+
+        if distances[close_idx] <= self.config.point_thickness:
+            return close_idx
+        return None
+
+    # === POINT/NODE CONTROL ===
+
+    def modify_node(self, new_prompt):
+        idx = self.selected_point_idx
+        new_prompts = self.prompts
+        new_prompts[idx] = new_prompt
+        self.set_prompts(new_prompts, reset = False)
+    
+    def add_node(self, new_prompt):
+        self.set_prompts(self.prompts + [new_prompt], reset = False)
+
+    def del_node(self):
+        idx = self.selected_point_idx
+        new_prompts = list(self.prompts)
+        del new_prompts[idx]
+        self.set_prompts(new_prompts, reset = False)
+        self.selected_point_idx = None
+
+    def prepare_to_prompt(self, mode):
+        """
+        Get ready to show the textbox. Call when we want the text prompt to come
+        """
+        self.inputting_text = True
+        self.inputting_text_for = mode
+
+        if mode == "modify":
+            self.create_text_prompt("Enter New Prompt To Replace Node:")
+        elif mode == "add":
+            self.create_text_prompt("Enter New Prompt To Create Node:")
+    
+    def handle_prompt(self):
+        """
+        After enter pressed with textbox, this is called to go back to normal game
+        """
+        done_prompting = self.text_prompt.update()
+
+        if done_prompting:
+            new_prompt = self.text_prompt.user_input.strip()
+            if self.inputting_text_for == "modify":
+                self.modify_node(new_prompt)
+            elif self.inputting_text_for == "add":
+                self.add_node(new_prompt)
+            self.text_prompt = None
+            self.inputting_text = False
+
+    def set_prompts(self, prompts : List[str], reset : bool = False):
+        """
+        :param prompts: New prompts to update to
+        :param reset: Reset xy positions of points?
+        """
+
+        if len(prompts) > 0:
+            encodes = self.get_encodes(prompts)
+
+        # First call
+        if not self.points or reset:
+            self.points = [Point(prompt, encoding, xy_init_kwargs = self.point_kwargs) for (prompt, encoding) in zip(prompts, encodes)]
+            return
+    
+        # Modifications
+        old_len = len(self.points)
+        new_len = len(prompts)
+    
+        pos = [tuple(pos_i) for pos_i in self.r2_points] # positions for each point
+
+        if old_len <= new_len: # Additions or modification
+            pos += [None] * (new_len - old_len) # randomly init this many new positions
+            self.points = [Point(prompt, encoding, pos_i, xy_init_kwargs = self.point_kwargs) for (prompt, encoding, pos_i) in zip(prompts, encodes, pos)]
+            return
+        elif old_len > new_len: # Deletions
+            idx_to_keep = []
+            for idx, prompt in enumerate(self.prompts):
+                if prompt in prompts:
+                    idx_to_keep.append(idx)
+            self.points = [self.points[idx] for idx in idx_to_keep]
+            return
+
+    # === CONTROLS ===
+
+    def handle_event_controls(self):
+        """
+        Handles discrete (i.e. keydown, mousedown) controls through events
+        """
+        for event in pygame.event.get():
+            if event.type == pygame.QUIT:
+                pygame.quit()
+                quit()
+            elif event.type == pygame.MOUSEBUTTONDOWN:
+                # Click
+                if event.button == 1: # Left click
+                    self.selected_point_idx = self.detect_mouse_on_point()
+                    if self.selected_point_idx is not None: self.dragging_point_idx = None
+                    else: # If no point was selected, we move player cursor
+                        self.get_player_pos_r2()
+                        self.draw_sample()
+                elif event.button == 3: # Right click
+                    self.dragging_point_idx = self.detect_mouse_on_point()
+                    if self.dragging_point_idx is not None: self.selected_point_idx = None
+            elif event.type == pygame.MOUSEBUTTONUP:
+                if event.button == 3: # Right Up
+                    self.dragging_point_idx = None # Disable drag
+            elif event.type == pygame.MOUSEMOTION:
+                if self.dragging_point_idx is not None:
+                    # Drag point
+                    self.points[self.dragging_point_idx].move(self.invert_screen_space(self.mouse_pos))
+                elif pygame.mouse.get_pressed()[0]:
+                    self.get_player_pos_r2()
+                    self.draw_sample()
+            elif event.type == pygame.KEYDOWN:
+                keys = pygame.key.get_pressed()
+                if keys[pygame.K_r]:
+                    self.set_prompts(self.prompts, reset = True)
+                elif keys[pygame.K_t] and self.selected_point_idx is not None: # Modify existing node
+                    self.prepare_to_prompt("modify")
+                    return
+                elif keys[pygame.K_p]: # Adding a node
+                    self.prepare_to_prompt("add")
+                    return
+                elif keys[pygame.K_o] and self.selected_point_idx is not None:
+                    # Remove node
+                    self.del_node()
+                elif keys[pygame.K_g]:
+                    if self.sample_image is not None:
+                        self.sample_image.save("sample.png")
+                elif keys[pygame.K_m]:
+                    # Change sampler mode
+                    self.switch_sampler()
+
+
+    def handle_continuous_controls(self):
+        """
+        Continuous controls for movement (i.e. zoom, movement)
+        """
+        keys = pygame.key.get_pressed()
+        if keys[pygame.K_q]:
+            self.zoom_level = max(0.01, self.zoom_level - self.config.zoom_speed)
+        elif keys[pygame.K_e]:
+            self.zoom_level = self.zoom_level + self.config.zoom_speed
+
+        idx, sign = None, None
+
+        # up, down, left, right
+        if keys[pygame.K_w]:
+            idx, sign = 1, -1
+        elif keys[pygame.K_s]:
+            idx, sign = 1, 1
+        elif keys[pygame.K_a]:
+            idx, sign = 0, -1
+        elif keys[pygame.K_d]:
+            idx, sign = 0, 1
+
+        if idx is not None and sign is not None:
+            self.translation[idx] += sign * self.config.move_speed
+
+    # === DRAWING THINGS ===
+
+    def draw_main_screen(self):
+        """
+        Draw main screen. Sample image, points, etc.
+        """
+        def get_point_color(idx):
+            color = (255, 255, 255) # default to white
+            if idx == self.selected_point_idx:
+                color = (0, 127.5, 0)
+            if idx == self.dragging_point_idx:
+                color = (255, 0, 0)
+            return color
+        
+        if self.config.sampler == "circle":
+            # Draw unit circle on screen
+            center = np.array([0,0])
+            border = np.array([1,0])
+
+            center = self.screen_space(center)
+            border = self.screen_space(border)
+            radius = abs(border[0] - center[0])
+
+            pygame.draw.circle(self.screen, (255, 255, 255), center, int(radius), 1)
+        
+        if len(self.points) > 0:
+            for idx, point in enumerate(self.screen_space_points):
+                pygame.draw.circle(self.screen, get_point_color(idx), point, self.config.point_thickness)
+                text = self.sample_font.render(self.points[idx].text, True, get_point_color(idx))
+                self.screen.blit(text, point)
+        
+        player_pos = self.get_player_pos_screenspace()
+        if player_pos is not None:
+            pygame.draw.circle(self.screen, (0, 255, 0), player_pos, self.config.point_thickness/2)
+        
+        if self.sample_image is not None:
+            pygame_image = pygame.image.fromstring(self.sample_image.tobytes(), self.sample_image.size, self.sample_image.mode)
+            pygame_image = pygame.transform.scale(pygame_image, (self.config.sample_width, self.config.sample_height))
+            self.screen.blit(pygame_image, (0, 0))
+
+    def update(self):
+        """
+        Main pygame loop
+        """
+
+        if not self.inputting_text:
+            self.handle_event_controls()
+            self.handle_continuous_controls()
+        self.tick()
+        
+        self.screen.fill((0,0,0))
+        self.draw_main_screen()
+
+        # Handle prompt after so it can be drawn over the main screen
+        if self.inputting_text:
+            self.handle_prompt()
+        pygame.display.flip()

latent_space_explorer/fast_sd.py

@@ -0,0 +1,34 @@
+from diffusers import AutoencoderTiny, StableDiffusionXLPipeline
+from .hacked_sdxl_pipeline import HackedSDXLPipeline
+import torch
+
+def fast_diffusion_pipeline(model_id = "stabilityai/sdxl-turbo", vae_id = "madebyollin/taesdxl", compile = False):
+    """
+    :param compile: If true, does a bunch of stuff to make calls fast, but the first call will be very slow as a consequence
+        - If you use this, don't vary the batch size (probably)
+    """
+
+    pipe = HackedSDXLPipeline.from_pretrained(model_id, torch_dtype = torch.float16)
+    pipe.set_progress_bar_config(disable=True)
+    pipe.cached_encode = None
+    pipe.vae = AutoencoderTiny.from_pretrained(vae_id, torch_dtype=torch.float16)
+
+    pipe.to('cuda')
+
+    if compile:
+        pipe.unet = torch.compile(pipe.unet)
+        pipe.vae.decode = torch.compile(pipe.vae.decode)
+        """
+        from sfast.compilers.stable_diffusion_pipeline_compiler import (compile, CompilationConfig)
+
+        config = CompilationConfig()
+        config.enable_jit = True
+        config.enable_jit_freeze = True
+        config.trace_scheduler = True
+        config.enable_cnn_optimization = True
+        config.preserve_parameters = False
+        config.prefer_lowp_gemm = True
+
+        pipe = compile(pipe, config)
+        """
+    return pipe

latent_space_explorer/game_objects.py

@@ -0,0 +1,97 @@
+from .utils import random_circle_init
+
+import math
+import pygame
+
+class Point:
+    # Point representing a prompt/generation
+    """
+    :param text: Text associated with the point
+    :param encoding: Encoding associated with the point
+    :param xy_pos: Tuple (x, y) for the position of point in R2. If not given, will initialize randomly
+    :param xy_init_kwargs: kwargs to random_circle_init for randomly init'ing xy_pos
+    """
+    def __init__(self, text, encoding, xy_pos = None, xy_init_kwargs = {}):
+        self.text = text
+
+        if xy_pos is not None:
+            self.xy_pos = xy_pos # Tuple of x and y in R2 space (not screen space)
+        else:
+            self.xy_pos = random_circle_init(**xy_init_kwargs)
+        self.encoding = encoding
+
+        if "on_edge" in xy_init_kwargs:
+            self.on_edge = xy_init_kwargs['on_edge']
+        else:
+            self.on_edge = False
+
+    def move(self, new_xy_pos):
+        if self.on_edge:
+            x, y = new_xy_pos
+            length = math.sqrt(x**2 + y**2)
+            self.xy_pos = (x/length, y/length)
+        else:
+            self.xy_pos = new_xy_pos
+
+class TextPrompt:
+    def __init__(self, prompt_text, font, screen):
+        self.prompt_text = prompt_text
+        self.font = font
+        self.screen = screen
+
+        self.user_input = ""
+
+    def draw_main_blocks(self):
+        """
+        Draw main text block and text inside it
+        """
+        screen_width, screen_height = pygame.display.get_surface().get_size()
+
+        # Margins
+        rect_height_fraction = 0.3
+        rect_width_fraction = 0.7
+        text_prompt_fraction = 0.3
+        border_thickness = 10
+
+        rect_height = int(screen_height * rect_height_fraction)
+        rect_width = int(screen_width * rect_width_fraction)
+
+        rect_x = (screen_width - rect_width) // 2
+        rect_y = (screen_height - rect_height) // 2
+
+        rect = pygame.Rect(rect_x, rect_y, rect_width, rect_height)
+
+        pygame.draw.rect(self.screen, (0, 0, 0), rect)  # Fill rectangle with black
+        pygame.draw.rect(self.screen, (255, 255, 255), rect, border_thickness)  # Draw white border
+
+        text_surface = self.font.render(self.prompt_text, True, (255, 255, 255))  # Render text
+        text_rect = text_surface.get_rect()  # Get text rectangle
+        text_rect.centerx = rect.centerx  # Center text horizontally
+        text_rect.y = rect.y + int(rect.height * text_prompt_fraction)  # Position text vertically based on fraction
+        self.screen.blit(text_surface, text_rect)  # Draw text
+
+        user_text_surface = self.font.render(self.user_input, True, (255, 255,255))
+        user_text_rect = user_text_surface.get_rect()
+        user_text_rect.centerx = rect.centerx
+        user_text_rect.y = rect.y + int(rect.height * (1 - text_prompt_fraction))
+        self.screen.blit(user_text_surface, user_text_rect)
+
+    def get_user_input(self) -> bool:
+        """
+        Get user input, update self.user_input, and return True if user pressed enter
+        """
+        for event in pygame.event.get():
+            if event.type == pygame.KEYDOWN:
+                if event.key == pygame.K_BACKSPACE:
+                    self.user_input = self.user_input[:-1]
+                elif event.key == pygame.K_RETURN:
+                    return True
+                else:
+                    self.user_input += event.unicode
+        return False
+
+    def update(self):
+        is_done = self.get_user_input()
+        self.draw_main_blocks()
+
+        return is_done

latent_space_explorer/hacked_sdxl_pipeline.py

@@ -0,0 +1,502 @@
+"""
+SDXL pipeline hacked to cache embeddings for later use
+
+Summary of changes:
+- __call__ takes "mode" that can be "cache" or "call"
+    - If "cache", just computes embeddings and returns nothing
+    - If "call", uses pre-computed embeddings ()
+    - Otherwise just has normal behaviour
+
+- There's a few custom methods after the init that you should look at if using
+"""
+
+from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl import *
+
+class HackedSDXLPipeline(StableDiffusionXLPipeline):
+    def get_encodes(self, *args, **kwargs):
+        """
+        Get encodings/latents for given prompt. Inputs are identical to if you were calling the pipeline.
+        """
+
+        self.__call__(*args, mode="cache", guidance_scale = 0.0, num_inference_steps = 1, **kwargs)
+        return self.cached_encodes
+    
+    def generate_from_encodes(self, encodes, *args, **kwargs):
+        """
+        Assuming you have some encodings/latents, pass here to generate from them.
+        """
+
+        if len(encodes[0].shape) == 2:
+            encodes[0] = encodes[0].unsqueeze(0)
+        if len(encodes[2].shape) == 1:
+            encodes[2] = encodes[2].unsqueeze(0)
+
+        self.cached_encodes = encodes
+        if 'prompt' in kwargs:
+            del kwargs['prompt']
+
+        return self.__call__(*args, prompt = [""] * len(self.cached_encodes[0]), guidance_scale = 0.0, num_inference_steps = 1, mode = "call", **kwargs)
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        prompt_2: Optional[Union[str, List[str]]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        timesteps: List[int] = None,
+        denoising_end: Optional[float] = None,
+        guidance_scale: float = 0.0,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt_2: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
+        ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        guidance_rescale: float = 0.0,
+        original_size: Optional[Tuple[int, int]] = None,
+        crops_coords_top_left: Tuple[int, int] = (0, 0),
+        target_size: Optional[Tuple[int, int]] = None,
+        negative_original_size: Optional[Tuple[int, int]] = None,
+        negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
+        negative_target_size: Optional[Tuple[int, int]] = None,
+        clip_skip: Optional[int] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        mode = "call", # cache or call
+        **kwargs,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                used in both text-encoders
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image. This is set to 1024 by default for the best results.
+                Anything below 512 pixels won't work well for
+                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+                and checkpoints that are not specifically fine-tuned on low resolutions.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image. This is set to 1024 by default for the best results.
+                Anything below 512 pixels won't work well for
+                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+                and checkpoints that are not specifically fine-tuned on low resolutions.
+            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.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            denoising_end (`float`, *optional*):
+                When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
+                completed before it is intentionally prematurely terminated. As a result, the returned sample will
+                still retain a substantial amount of noise as determined by the discrete timesteps selected by the
+                scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a
+                "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
+                Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output)
+            guidance_scale (`float`, *optional*, defaults to 5.0):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+                `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+            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 (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noireturn Nonesy 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
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+            ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+                Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+                Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+                if `do_classifier_free_guidance` is set to `True`.
+                If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
+                of a plain tuple.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            guidance_rescale (`float`, *optional*, defaults to 0.0):
+                Guidancettioning return Noneas explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                To negatively condition the generation process based on a target image resolution. It should be as same
+                as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
+            `tuple`. When returning a tuple, the first element is a list with the generated images.
+        """
+
+        callback = kwargs.pop("callback", None)
+        callback_steps = kwargs.pop("callback_steps", None)
+
+        if callback is not None:
+            deprecate(
+                "callback",
+                "1.0.0",
+                "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+            )
+        if callback_steps is not None:
+            deprecate(
+                "callback_steps",
+                "1.0.0",
+                "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+            )
+
+        # 0. Default height and width to unet
+        height = height or self.default_sample_size * self.vae_scale_factor
+        width = width or self.default_sample_size * self.vae_scale_factor
+
+        original_size = original_size or (height, width)
+        target_size = target_size or (height, width)
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            prompt_2,
+            height,
+            width,
+            callback_steps,
+            negative_prompt,
+            negative_prompt_2,
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+            ip_adapter_image,
+            ip_adapter_image_embeds,
+            callback_on_step_end_tensor_inputs,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._guidance_rescale = guidance_rescale
+        self._clip_skip = clip_skip
+        self._cross_attention_kwargs = cross_attention_kwargs
+        self._denoising_end = denoising_end
+        self._interrupt = False
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        # 3. Encode input prompt
+        lora_scale = (
+            self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+        )
+
+
+        if mode == "cache":
+            self.cached_encodes = self.encode_prompt(
+                prompt=prompt,
+                prompt_2=prompt_2,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                do_classifier_free_guidance=self.do_classifier_free_guidance,
+                negative_prompt=negative_prompt,
+                negative_prompt_2=negative_prompt_2,
+                prompt_embeds=prompt_embeds,
+                negative_prompt_embeds=negative_prompt_embeds,
+                pooled_prompt_embeds=pooled_prompt_embeds,
+                negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+                lora_scale=lora_scale,
+                clip_skip=self.clip_skip,
+            )
+            return None
+        elif mode == "call":
+            (
+                prompt_embeds,
+                negative_prompt_embeds,
+                pooled_prompt_embeds,
+                negative_pooled_prompt_embeds,
+            ) = self.cached_encodes
+        else: # Normal behaviour
+            (
+                prompt_embeds,
+                negative_prompt_embeds,
+                pooled_prompt_embeds,
+                negative_pooled_prompt_embeds,
+            ) = self.encode_prompt(
+                prompt=prompt,
+                prompt_2=prompt_2,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                do_classifier_free_guidance=self.do_classifier_free_guidance,
+                negative_prompt=negative_prompt,
+                negative_prompt_2=negative_prompt_2,
+                prompt_embeds=prompt_embeds,
+                negative_prompt_embeds=negative_prompt_embeds,
+                pooled_prompt_embeds=pooled_prompt_embeds,
+                negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+                lora_scale=lora_scale,
+                clip_skip=self.clip_skip,
+            )          
+
+        # 4. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.unet.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. 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)
+
+        # 7. Prepare added time ids & embeddings
+        add_text_embeds = pooled_prompt_embeds
+        if self.text_encoder_2 is None:
+            text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
+        else:
+            text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
+
+        add_time_ids = self._get_add_time_ids(
+            original_size,
+            crops_coords_top_left,
+            target_size,
+            dtype=prompt_embeds.dtype,
+            text_encoder_projection_dim=text_encoder_projection_dim,
+        )
+        if negative_original_size is not None and negative_target_size is not None:
+            negative_add_time_ids = self._get_add_time_ids(
+                negative_original_size,
+                negative_crops_coords_top_left,
+                negative_target_size,
+                dtype=prompt_embeds.dtype,
+                text_encoder_projection_dim=text_encoder_projection_dim,
+            )
+        else:
+            negative_add_time_ids = add_time_ids
+
+        if False:#self.do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
+            add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
+
+        prompt_embeds = prompt_embeds.to(device)
+        add_text_embeds = add_text_embeds.to(device)
+        add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
+
+        if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+            image_embeds = self.prepare_ip_adapter_image_embeds(
+                ip_adapter_image,
+                ip_adapter_image_embeds,
+                device,
+                batch_size * num_images_per_prompt,
+                self.do_classifier_free_guidance,
+            )
+
+        # 8. Denoising loop
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+        # 8.1 Apply denoising_end
+        if (
+            self.denoising_end is not None
+            and isinstance(self.denoising_end, float)
+            and self.denoising_end > 0
+            and self.denoising_end < 1
+        ):
+            discrete_timestep_cutoff = int(
+                round(
+                    self.scheduler.config.num_train_timesteps
+                    - (self.denoising_end * self.scheduler.config.num_train_timesteps)
+                )
+            )
+            num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
+            timesteps = timesteps[:num_inference_steps]
+
+        # 9. Optionally get Guidance Scale Embedding
+        timestep_cond = None
+        if self.unet.config.time_cond_proj_dim is not None:
+            guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+            timestep_cond = self.get_guidance_scale_embedding(
+                guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+            ).to(device=device, dtype=latents.dtype)
+
+        self._num_timesteps = len(timesteps)
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                # predict the noise residual
+                added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
+                if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+                    added_cond_kwargs["image_embeds"] = image_embeds
+                noise_pred = self.unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep_cond=timestep_cond,
+                    cross_attention_kwargs=self.cross_attention_kwargs,
+                    added_cond_kwargs=added_cond_kwargs,
+                    return_dict=False,
+                )[0]
+
+                # perform guidance
+                if self.do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
+                    # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+                    noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+                    add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds)
+                    negative_pooled_prompt_embeds = callback_outputs.pop(
+                        "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
+                    )
+                    add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids)
+                    negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+                    if callback is not None and i % callback_steps == 0:
+                        step_idx = i // getattr(self.scheduler, "order", 1)
+                        callback(step_idx, t, latents)
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        if not output_type == "latent":
+            # make sure the VAE is in float32 mode, as it overflows in float16
+            needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+            if needs_upcasting:
+                self.upcast_vae()
+                latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
+            # unscale/denormalize the latents
+            # denormalize with the mean and std if available and not None
+            has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None
+            has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None
+            if has_latents_mean and has_latents_std:
+                latents_mean = (
+                    torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+                )
+                latents_std = (
+                    torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+                )
+                latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean
+            else:
+                latents = latents / self.vae.config.scaling_factor
+
+            image = self.vae.decode(latents, return_dict=False)[0]
+
+            # cast back to fp16 if needed
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float16)
+        else:
+            image = latents
+
+        if not output_type == "latent":
+            # apply watermark if available
+            if self.watermark is not None:
+                image = self.watermark.apply_watermark(image)
+
+            image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image,)
+
+        return StableDiffusionXLPipelineOutput(images=image)