import spaces import gradio as gr import os import random import json import time import uuid from PIL import Image from huggingface_hub import snapshot_download from diffusers import AutoencoderKL from diffusers import StableDiffusionXLPipeline, EulerDiscreteScheduler, AutoPipelineForText2Image, DiffusionPipeline from diffusers import EulerAncestralDiscreteScheduler, DPMSolverMultistepScheduler, DPMSolverSDEScheduler from diffusers.models.attention_processor import AttnProcessor2_0 import torch from typing import Tuple from datetime import datetime import requests import torch from diffusers import DiffusionPipeline import importlib import re from urllib.parse import urlparse random.seed(time.time()) MAX_SEED = 12211231 CACHE_EXAMPLES = "1" MAX_IMAGE_SIZE = int(os.getenv("MAX_IMAGE_SIZE", "4192")) USE_TORCH_COMPILE = os.getenv("USE_TORCH_COMPILE", "0") == "1" ENABLE_CPU_OFFLOAD = os.getenv("ENABLE_CPU_OFFLOAD", "0") == "1" NUM_IMAGES_PER_PROMPT = 1 # Define the regular expression child_related_regex = re.compile( r'(child|children|kid|kids|baby|babies|toddler|infant|juvenile|minor|underage|preteen|adolescent|youngster|youth|son|daughter|young|kindergarten|preschool|' r'([1-9]|1[0-7])[\s_\-|\.\,]*year(s)?[\s_\-|\.\,]*old|' # Matches 1 to 17 years old with various separators r'little|small|tiny|short|young|new[\s_\-|\.\,]*born[\s_\-|\.\,]*(boy|girl|male|man|bro|brother|sis|sister))', re.IGNORECASE ) # Function to remove child-related content from a prompt def remove_child_related_content(prompt): cleaned_prompt = re.sub(child_related_regex, '', prompt) return cleaned_prompt.strip() # Function to check if a prompt contains child-related content def contains_child_related_content(prompt): if child_related_regex.search(prompt): return True return False cfg = json.load(open("app.conf")) def load_pipeline_and_scheduler(): clip_skip = cfg.get("clip_skip", 0) # Download the model files ckpt_dir = snapshot_download(repo_id=cfg["model_id"]) # Load the models vae = AutoencoderKL.from_pretrained(os.path.join(ckpt_dir, "vae"), torch_dtype=torch.float16) pipe = StableDiffusionXLPipeline.from_pretrained( ckpt_dir, vae=vae, torch_dtype=torch.float16, use_safetensors=True, variant="fp16" ) pipe = pipe.to("cuda") pipe.unet.set_attn_processor(AttnProcessor2_0()) # Define samplers samplers = { "Euler a": EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config), "DPM++ SDE Karras": DPMSolverSDEScheduler.from_config(pipe.scheduler.config, use_karras_sigmas=True) } # Set the scheduler based on the selected sampler pipe.scheduler = samplers[cfg.get("sampler","DPM++ SDE Karras")] # Set clip skip pipe.text_encoder.config.num_hidden_layers -= (clip_skip - 1) if USE_TORCH_COMPILE: pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True) print("Model Compiled!") return pipe pipe = load_pipeline_and_scheduler() css = ''' .gradio-container{max-width: 560px !important} body { background-color: rgb(3, 7, 18); color: white; } .gradio-container { background-color: rgb(3, 7, 18) !important; border: none !important; } .gradio-container footer { display: block !important; } ''' js = ''' ''' desc_html='''

For the full version and more exciting NSFW AI apps, visit nsfwais.io!

''' def save_image(img): # Generate a unique filename unique_name = str(uuid.uuid4()) + ".webp" # Convert the image to WebP format webp_img = img.convert("RGB") # Ensure the image is in RGB mode # Save the image in WebP format with high quality webp_img.save(unique_name, "WEBP", quality=90) # Open the saved WebP file and return it as a PIL Image object with Image.open(unique_name) as webp_file: webp_image = webp_file.copy() return webp_image, 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) def generate(p, progress=gr.Progress(track_tqdm=True)): negative_prompt = cfg.get("negative_prompt", "") style_selection = "" use_negative_prompt = True seed = 0 width = cfg.get("width", 1024) height = cfg.get("width", 768) inference_steps = cfg.get("inference_steps", 30) randomize_seed = True guidance_scale = cfg.get("guidance_scale", 7.5) p = remove_child_related_content(p) prompt_str = cfg.get("prompt", "{prompt}").replace("{prompt}", p) seed = int(randomize_seed_fn(seed, randomize_seed)) generator = torch.Generator(pipe.device).manual_seed(seed) images = pipe( prompt=prompt_str, negative_prompt=negative_prompt, width=width, height=height, guidance_scale=guidance_scale, num_inference_steps=inference_steps, generator=generator, num_images_per_prompt=NUM_IMAGES_PER_PROMPT, output_type="pil", ).images images = [save_image(img) for img in images] image_paths = [i[1] for i in images] print(prompt_str, image_paths) return [i[0] for i in images] default_image = cfg.get("cover_path", None) if default_image: if isinstance(default_image, list): # Filter out non-existent paths existing_images = [img for img in default_image if os.path.exists(img)] #print(f"found cover files: {existing_images}") if existing_images: default_image = existing_images[int(time.time()*1000)%len(existing_images)] else: default_image = None elif not os.path.exists(default_image): print(f"cover file not existed, {default_image}") default_image = None else: default_image = None with gr.Blocks(css=css,head=js,fill_height=True) as demo: with gr.Row(equal_height=False): with gr.Group(): gr.HTML(value=desc_html, elem_id='desc_html_code') result = gr.Gallery(value=[default_image], label="Result", show_label=False, columns=1, rows=1, show_share_button=True,elem_id=cfg["model_id"].replace("/", "-"), show_download_button=True,allow_preview=False,interactive=False, min_width=cfg.get("window_min_width", 340),height=360 ) with gr.Row(): prompt = gr.Text( show_label=False, max_lines=2, lines=2, placeholder="Enter your fantasy or click ->", container=False, scale=5, min_width=100, elem_id="prompt_input_box" ) random_button = gr.Button("Surprise Me", scale=1, min_width=10) run_button = gr.Button( "GO!", scale=1, min_width=20, variant="primary",icon="https://huggingface.co/spaces/nsfwalex/sd_card/resolve/main/hot.svg") def on_demo_load(request: gr.Request): current_domain = request.request.headers.get("Host", "") # Get the potential iframe parent domain from the Referer header referer = request.request.headers.get("Referer", "") iframe_parent_domain = "" if referer: try: parsed_referer = urlparse(referer) iframe_parent_domain = parsed_referer.netloc except: iframe_parent_domain = "Unable to parse referer" params = dict(request.query_params) print(f"load_demo, urlparams={params},cover={default_image},domain={current_domain},iframe={iframe_parent_domain}") session_data = { "params": params, "client_ip": request.client.host, "refer": referer, "host": current_domain } if params.get("e", "0") == "1" or "nsfwais.io" in current_domain or "nsfwais.io" in iframe_parent_domain or "127.0.0.1" in current_domain or "127.0.0.1" in iframe_parent_domain: #update the image #bind events #return [Image.open(default_image)], session_data return session_data return session_data #return [], session_data session_state = gr.State() result.change(fn=lambda x,y:None , inputs=[prompt,result], outputs=[], js=f'''(p,img)=>window.uploadImage(p, img,"process_finished","demo_hf_{cfg.get("name")}_card", "{cfg["model_id"]}")''') run_button.click(generate, inputs=[prompt], outputs=[result],trigger_mode="once",js=f'''(p)=>window.postMessageToParent(p,"process_started","demo_hf_{cfg.get("name")}_card", "click_go")''') random_button.click(fn=lambda x:x, inputs=[prompt], outputs=[prompt], js='''(p)=>window.g(p)''') demo.load(fn=on_demo_load, inputs=[], outputs=[session_state], js='''()=>onDemoLoad()''') if __name__ == "__main__": demo.queue(max_size=100).launch(show_api=True,show_error=False)