app.py

import os
import re
import spaces
import random
import string
import torch
import requests
import gradio as gr
import numpy as np
from lxml.html import fromstring
from transformers import pipeline
from torch import multiprocessing as mp, nn
#from torch.multiprocessing import Pool
#from pathos.multiprocessing import ProcessPool as Pool
from pathos.threading import ThreadPool as Pool
from diffusers.pipelines.flux import FluxPipeline
from diffusers.utils import export_to_gif, load_image
from diffusers.models.modeling_utils import ModelMixin
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file, save_file
from diffusers import DiffusionPipeline, AnimateDiffPipeline, MotionAdapter, EulerDiscreteScheduler, DDIMScheduler, StableDiffusionXLPipeline, UNet2DConditionModel, AutoencoderKL, UNet3DConditionModel
import jax
import jax.numpy as jnp

last_motion=None
fps=15
time=1
width=448
height=512
device = "cuda"
dtype = torch.float16
result=[]
step = 30
accu=7.5
#repo = "ByteDance/AnimateDiff-Lightning"
#ckpt = f"animatediff_lightning_{step}step_diffusers.safetensors"
#base = "emilianJR/epiCRealism"
base = "SG161222/Realistic_Vision_V6.0_B1_noVAE"
vae = AutoencoderKL.from_pretrained("stabilityai/sd-vae-ft-mse").to(device, dtype=dtype)
#unet = UNet2DConditionModel.from_config("emilianJR/epiCRealism",subfolder="unet").to(device, dtype).load_state_dict(load_file(hf_hub_download("emilianJR/epiCRealism", "unet/diffusion_pytorch_model.safetensors"), device=device), strict=False)
adapter = MotionAdapter.from_pretrained("guoyww/animatediff-motion-adapter-v1-5-3", torch_dtype=dtype, device=device)

css="""
input, input::placeholder {
    text-align: center !important;
}
*, *::placeholder {
    font-family: Suez One !important;
}
h1,h2,h3,h4,h5,h6 {
    width: 100%;
    text-align: center;
}
footer {
    display: none !important;
}
#col-container {
    margin: 0 auto;
    max-width: 15cm;
}
.image-container {
    aspect-ratio: """+str(width)+"/"+str(height)+""" !important;
}
.dropdown-arrow {
    display: none !important;
}
*:has(>.btn) {
    display: flex;
    justify-content: space-evenly;
    align-items: center;
}
.btn {
    display: flex;
}
"""

js="""
function custom(){
    document.querySelector("div#prompt input").setAttribute("maxlength","38")
    document.querySelector("div#prompt2 input").setAttribute("maxlength","38")
}
"""

#def forest_schnell():
#    PIPE = DiffusionPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16, token=os.getenv("hf_token")).to("cuda")
#    return PIPE
    
def translate(text,lang):
    if text == None or lang == None:
        return ""       
    text = re.sub(f'[{string.punctuation}]', '', re.sub('[\s+]', ' ', text)).lower().strip()
    lang = re.sub(f'[{string.punctuation}]', '', re.sub('[\s+]', ' ', lang)).lower().strip()    
    if text == "" or lang == "":
        return ""
    if len(text) > 38:
        raise Exception("Translation Error: Too long text!")
    user_agents = [
        'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/126.0.0.0 Safari/537.36',
        'Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/126.0.0.0 Safari/537.36',
        'Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/126.0.0.0 Safari/537.36',
        'Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/605.1.15 (KHTML, like Gecko) Version/16.1 Safari/605.1.15',
        'Mozilla/5.0 (Macintosh; Intel Mac OS X 13_1) AppleWebKit/605.1.15 (KHTML, like Gecko) Version/16.1 Safari/605.1.15'
    ]
    padded_chars = re.sub("[(^\-)(\-$)]","",text.replace("","-").replace("- -"," ")).strip()
    query_text = f'Please translate {padded_chars}, into {lang}'
    url = f'https://www.google.com/search?q={query_text}'
    resp = requests.get(
        url = url,
        headers = {
            'User-Agent': random.choice(user_agents)
        }
    )
    content = resp.content
    html = fromstring(content)
    translated = text
    try:
        src_lang = html.xpath('//*[@class="source-language"]')[0].text_content().lower().strip()
        trgt_lang = html.xpath('//*[@class="target-language"]')[0].text_content().lower().strip()
        src_text = html.xpath('//*[@id="tw-source-text"]/*')[0].text_content().lower().strip()
        trgt_text = html.xpath('//*[@id="tw-target-text"]/*')[0].text_content().lower().strip()
        if trgt_lang == lang:
            translated = trgt_text
    except:
        print(f'Translation Warning: Failed To Translate!')
    ret = re.sub(f'[{string.punctuation}]', '', re.sub('[\s+]', ' ', translated)).lower().strip()
    print(ret)
    return ret

def generate_random_string(length):
    characters = string.ascii_letters + string.digits
    return ''.join(random.choice(characters) for _ in range(length))

@spaces.GPU(duration=120)
def Piper(image,positive,negative,motion):
    global last_motion
    global ip_loaded

    if last_motion != motion:
        pipe.unload_lora_weights()
        if motion != "":
            pipe.load_lora_weights(motion, adapter_name="motion")
            pipe.fuse_lora()
            pipe.set_adapters(["motion"], [0.7])
        last_motion = motion

    pipe.to(device,dtype)

    if negative=="":
        return pipe(
            prompt=positive,
            height=height,
            width=width,
            ip_adapter_image=image.convert("RGB").resize((width,height)),
            num_inference_steps=step,
            guidance_scale=accu,
            num_frames=(fps*time)
        )
    
    return pipe(
        prompt=positive,
        negative_prompt=negative,
        height=height,
        width=width,
        ip_adapter_image=image.convert("RGB").resize((width,height)),
        num_inference_steps=step,
        guidance_scale=accu,
        num_frames=(fps*time)
    )

def infer(pm):
        print("infer: started")
    
        p1 = pm["p"]
        name = generate_random_string(12)+".png"
        neg = pm["n"]

        _do = ['beautiful', 'playful', 'photographed', 'realistic', 'dynamic poze', 'deep field', 'reasonable coloring', 'rough texture', 'best quality', 'focused']
        if p1 != "":
            _do.append(f'{p1}')
        posi = " ".join(_do)

        if pm["i"] == None:
            return None
        out = Piper(pm["i"],posi,neg,pm["m"])
        export_to_gif(out.frames[0],name,fps=fps)
        return name

def run(i,m,p1,p2,*result):
    
        p1_en = translate(p1,"english")
        p2_en = translate(p2,"english")
        pm = {"p":p1_en,"n":p2_en,"m":m,"i":i}
        ln = len(result)
        print("images: "+str(ln))
        rng = list(range(ln))
    
        arr = [pm for _ in rng]
        pool = Pool(ln)
        out = list(pool.imap(infer,arr))
        pool.close()
        pool.join()
        pool.clear()

        return out

pipe = AnimateDiffPipeline.from_pretrained(base, vae=vae, motion_adapter=adapter, torch_dtype=dtype).to(device)
pipe.scheduler = DDIMScheduler(
        clip_sample=False,
        beta_start=0.00085,
        beta_end=0.012,
        beta_schedule="linear",
        timestep_spacing="trailing",
        steps_offset=1
)
#pipe.unet.load_state_dict(load_file(hf_hub_download(repo, ckpt), device=device), strict=False)
pipe.load_ip_adapter("h94/IP-Adapter", subfolder="models", weight_name="ip-adapter_sd15.bin")
pipe.enable_vae_slicing()
pipe.enable_free_init(method="butterworth", use_fast_sampling=True)

mp.set_start_method("spawn", force=True)

with gr.Blocks(theme=gr.themes.Soft(),css=css,js=js) as demo:
    with gr.Column(elem_id="col-container"):
        gr.Markdown(f"""
            # MULTI-LANGUAGE IMAGE GENERATOR
        """)
        with gr.Row():
            img = gr.Image(type="pil")
        with gr.Row():
            prompt = gr.Textbox(
                elem_id="prompt",
                placeholder="INCLUDE",
                container=False,
                max_lines=1
            )
        with gr.Row():
            prompt2 = gr.Textbox(
                elem_id="prompt2",
                placeholder="EXCLUDE",
                container=False,
                max_lines=1
            )
        with gr.Row(visible=False):
                motion = gr.Dropdown(
                    label='Motion',
                    show_label=False,
                    choices=[
                        ("(None)", ""),
                        ("Zoom in", "guoyww/animatediff-motion-lora-zoom-in"),
                        ("Zoom out", "guoyww/animatediff-motion-lora-zoom-out"),
                        ("Tilt up", "guoyww/animatediff-motion-lora-tilt-up"),
                        ("Tilt down", "guoyww/animatediff-motion-lora-tilt-down"),
                        ("Pan left", "guoyww/animatediff-motion-lora-pan-left"),
                        ("Pan right", "guoyww/animatediff-motion-lora-pan-right"),
                        ("Roll left", "guoyww/animatediff-motion-lora-rolling-anticlockwise"),
                        ("Roll right", "guoyww/animatediff-motion-lora-rolling-clockwise"),
                    ],
                    value="",
                    interactive=True
                )
        with gr.Row():
                run_button = gr.Button("START",elem_classes="btn",scale=0)
        with gr.Row():
                result.append(gr.Image(interactive=False,elem_classes="image-container", label="Result", show_label=False, type='filepath', show_share_button=False))
                result.append(gr.Image(interactive=False,elem_classes="image-container", label="Result", show_label=False, type='filepath', show_share_button=False))
                
    gr.on(
            triggers=[run_button.click, prompt.submit, prompt2.submit],
            fn=run,inputs=[img,motion,prompt,prompt2,*result],outputs=result
    )
    demo.queue().launch()