#!/usr/bin/env python import os import random import uuid import gradio as gr import numpy as np from PIL import Image import spaces import torch from diffusers import StableDiffusionXLPipeline, EulerAncestralDiscreteScheduler from typing import Tuple css = ''' .gradio-container{max-width: 570px !important} h1{text-align:center} footer { visibility: hidden } ''' DESCRIPTIONXX = """ ## REALVISXL V5 + LIGHTNING ⚡ """ examples = [ "Illustration of A starry night camp in the mountains, 4k, cinematic --ar 85:128 --v 6.0 --style raw", "A delicious ceviche cheesecake slice, 4k, octane render, ray tracing, Ultra-High-Definition" ] MODEL_OPTIONS = { "REALVISXL V5.0": "SG161222/RealVisXL_V5.0", # "LIGHTNING V5.0": "SG161222/RealVisXL_V5.0_Lightning", } MAX_IMAGE_SIZE = int(os.getenv("MAX_IMAGE_SIZE", "4096")) USE_TORCH_COMPILE = os.getenv("USE_TORCH_COMPILE", "0") == "1" ENABLE_CPU_OFFLOAD = os.getenv("ENABLE_CPU_OFFLOAD", "0") == "1" BATCH_SIZE = int(os.getenv("BATCH_SIZE", "1")) device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") style_list = [ { "name": "3840 x 2160", "prompt": "hyper-realistic 8K image of {prompt}. ultra-detailed, lifelike, high-resolution, sharp, vibrant colors, photorealistic", "negative_prompt": "cartoonish, low resolution, blurry, simplistic, abstract, deformed, ugly", }, { "name": "2560 x 1440", "prompt": "hyper-realistic 4K image of {prompt}. ultra-detailed, lifelike, high-resolution, sharp, vibrant colors, photorealistic", "negative_prompt": "cartoonish, low resolution, blurry, simplistic, abstract, deformed, ugly", }, { "name": "HD+", "prompt": "hyper-realistic 2K image of {prompt}. ultra-detailed, lifelike, high-resolution, sharp, vibrant colors, photorealistic", "negative_prompt": "cartoonish, low resolution, blurry, simplistic, abstract, deformed, ugly", }, { "name": "Style Zero", "prompt": "{prompt}", "negative_prompt": "", }, ] styles = {k["name"]: (k["prompt"], k["negative_prompt"]) for k in style_list} DEFAULT_STYLE_NAME = "3840 x 2160" STYLE_NAMES = list(styles.keys()) def apply_style(style_name: str, positive: str, negative: str = "") -> Tuple[str, str]: if style_name in styles: p, n = styles.get(style_name, styles[DEFAULT_STYLE_NAME]) else: p, n = styles[DEFAULT_STYLE_NAME] if not negative: negative = "" return p.replace("{prompt}", positive), n + negative def load_and_prepare_model(model_id): pipe = StableDiffusionXLPipeline.from_pretrained( model_id, torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32, use_safetensors=True, add_watermarker=False, ).to(device) pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config) if USE_TORCH_COMPILE: pipe.compile() if ENABLE_CPU_OFFLOAD: pipe.enable_model_cpu_offload() return pipe # Preload and compile both models models = {key: load_and_prepare_model(value) for key, value in MODEL_OPTIONS.items()} MAX_SEED = np.iinfo(np.int32).max def save_image(img): unique_name = str(uuid.uuid4()) + ".png" img.save(unique_name) return unique_name def randomize_seed_fn(seed: int, randomize_seed: bool) -> int: if randomize_seed: seed = random.randint(0, MAX_SEED) return seed @spaces.GPU(duration=60, enable_queue=True) def generate( model_choice: str, prompt: str, negative_prompt: str = "", use_negative_prompt: bool = False, style_selection: str = DEFAULT_STYLE_NAME, seed: int = 1, width: int = 1024, height: int = 1024, guidance_scale: float = 3, num_inference_steps: int = 25, randomize_seed: bool = False, use_resolution_binning: bool = True, num_images: int = 1, progress=gr.Progress(track_tqdm=True), ): global models pipe = models[model_choice] seed = int(randomize_seed_fn(seed, randomize_seed)) generator = torch.Generator(device=device).manual_seed(seed) prompt, negative_prompt = apply_style(style_selection, prompt, negative_prompt) options = { "prompt": [prompt] * num_images, "negative_prompt": [negative_prompt] * num_images if use_negative_prompt else None, "width": width, "height": height, "guidance_scale": guidance_scale, "num_inference_steps": num_inference_steps, "generator": generator, "output_type": "pil", } if use_resolution_binning: options["use_resolution_binning"] = True images = [] for i in range(0, num_images, BATCH_SIZE): batch_options = options.copy() batch_options["prompt"] = options["prompt"][i:i+BATCH_SIZE] if "negative_prompt" in batch_options: batch_options["negative_prompt"] = options["negative_prompt"][i:i+BATCH_SIZE] images.extend(pipe(**batch_options).images) image_paths = [save_image(img) for img in images] return image_paths, seed with gr.Blocks(css=css, theme="bethecloud/storj_theme") as demo: gr.Markdown(DESCRIPTIONXX) with gr.Row(): prompt = gr.Text( label="Prompt", show_label=False, max_lines=1, placeholder="Enter your prompt", container=False, ) run_button = gr.Button("Run", scale=0) result = gr.Gallery(label="Result", columns=1, show_label=False) with gr.Row(): model_choice = gr.Dropdown( label="Model Selection🔻", choices=list(MODEL_OPTIONS.keys()), value="REALVISXL V5.0" ) with gr.Accordion("Advanced options", open=False, visible=True): style_selection = gr.Radio( show_label=True, container=True, interactive=True, choices=STYLE_NAMES, value=DEFAULT_STYLE_NAME, label="Quality Style", ) num_images = gr.Slider( label="Number of Images", minimum=1, maximum=5, step=1, value=1, ) with gr.Row(): with gr.Column(scale=1): use_negative_prompt = gr.Checkbox(label="Use negative prompt", value=True) negative_prompt = gr.Text( label="Negative prompt", max_lines=5, lines=4, placeholder="Enter a negative prompt", value="(deformed, distorted, disfigured:1.3), poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, (mutated hands and fingers:1.4), disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation", visible=True, ) seed = gr.Slider( label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0, ) randomize_seed = gr.Checkbox(label="Randomize seed", value=True) with gr.Row(): width = gr.Slider( label="Width", minimum=512, maximum=MAX_IMAGE_SIZE, step=64, value=1024, ) height = gr.Slider( label="Height", minimum=512, maximum=MAX_IMAGE_SIZE, step=64, value=1024, ) with gr.Row(): guidance_scale = gr.Slider( label="Guidance Scale", minimum=0.1, maximum=6, step=0.1, value=3.0, ) num_inference_steps = gr.Slider( label="Number of inference steps", minimum=1, maximum=60, step=1, value=32, ) gr.Examples( examples=examples, inputs=prompt, cache_examples=False ) use_negative_prompt.change( fn=lambda x: gr.update(visible=x), inputs=use_negative_prompt, outputs=negative_prompt, api_name=False, ) gr.on( triggers=[ prompt.submit, negative_prompt.submit, run_button.click, ], fn=generate, inputs=[ model_choice, prompt, negative_prompt, use_negative_prompt, style_selection, seed, width, height, guidance_scale, num_inference_steps, randomize_seed, gr.State(value=True), num_images, ], outputs=[result, seed], ) demo.queue(concurrency_count=3).launch(debug=True)