app.py

# -*- coding: utf-8 -*-
import gradio as gr
import os
import sys
import random
import time
import uuid
from omegaconf import OmegaConf
import torch
import torchvision
from pytorch_lightning import seed_everything
from huggingface_hub import hf_hub_download
from einops import repeat
import torchvision.transforms as transforms
from utils.utils import instantiate_from_config
sys.path.insert(0, "scripts/evaluation")
from funcs import (
    batch_ddim_sampling,
    load_model_checkpoint,
    get_latent_z,
    save_videos
)
from transformers import pipeline, AutoTokenizer, AutoModelForSeq2SeqLM
from diffusers import StableDiffusionXLPipeline

print("PyTorch version:", torch.__version__)
print("CUDA available:", torch.cuda.is_available())

# 나머지 코드는 그대로 유지...

def is_tensor(x):
    return torch.is_tensor(x)

# 번역 모델 로드 (PyTorch 버전 사용)
translator = pipeline("translation", model="Helsinki-NLP/opus-mt-ko-en", device=0 if torch.cuda.is_available() else -1, framework="pt")

# 이미지 생성 모델 로드
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
pipe = StableDiffusionXLPipeline.from_pretrained(
    "SG161222/RealVisXL_V4.0",
    torch_dtype=torch.float32,
    use_safetensors=True,
    add_watermarker=False
).to(device)


os.environ['KERAS_BACKEND'] = 'pytorch'

def download_model():
    REPO_ID = 'Doubiiu/DynamiCrafter_1024'
    filename_list = ['model.ckpt']
    if not os.path.exists('./checkpoints/dynamicrafter_1024_v1/'):
        os.makedirs('./checkpoints/dynamicrafter_1024_v1/')
    for filename in filename_list:
        local_file = os.path.join('./checkpoints/dynamicrafter_1024_v1/', filename)
        if not os.path.exists(local_file):
            hf_hub_download(repo_id=REPO_ID, filename=filename, local_dir='./checkpoints/dynamicrafter_1024_v1/', force_download=True)

download_model()
ckpt_path = 'checkpoints/dynamicrafter_1024_v1/model.ckpt'
config_file = 'configs/inference_1024_v1.0.yaml'
config = OmegaConf.load(config_file)
model_config = config.pop("model", OmegaConf.create())
model_config['params']['unet_config']['params']['use_checkpoint'] = True
model = instantiate_from_config(model_config)
assert os.path.exists(ckpt_path), "Error: checkpoint Not Found!"
model = load_model_checkpoint(model, ckpt_path)
model.eval()

# 모델을 DataParallel로 감싸서 여러 GPU에서 실행 가능하게 설정
#model = torch.nn.DataParallel(model)
model = model.cuda()



def generate_image(prompt: str):
    # 한글 입력 감지 및 번역
    if any('\uac00' <= char <= '\ud7a3' for char in prompt):
        translated = translator(prompt, max_length=512)
        prompt = translated[0]['translation_text']
    
    # Hi-res와 3840x2160 스타일 적용
    prompt = f"hyper-realistic 8K image of {prompt}. ultra-detailed, lifelike, high-resolution, sharp, vibrant colors, photorealistic"
    
    # 고정된 설정값
    negative_prompt = "cartoonish, low resolution, blurry, simplistic, abstract, deformed, ugly, (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, disgusting, amputation"
    width = 1024
    height = 576
    guidance_scale = 6
    num_inference_steps = 100
    seed = random.randint(0, 2**32 - 1)
    generator = torch.Generator().manual_seed(seed)
    
    image = pipe(
        prompt=prompt,
        negative_prompt=negative_prompt,
        width=width,
        height=height,
        guidance_scale=guidance_scale,
        num_inference_steps=num_inference_steps,
        generator=generator,
    ).images[0]
    
    unique_name = str(uuid.uuid4()) + ".png"
    image.save(unique_name)
    return unique_name

# @spaces.GPU(duration=300, gpu_type="l40s")
def infer(prompt, steps=50, cfg_scale=7.5, eta=1.0, fs=3, seed=123, frames=64):
    try:
        image_path = generate_image(prompt)
        image = torchvision.io.read_image(image_path).float() / 255.0
        
        if any('\u3131' <= char <= '\u318E' or '\uAC00' <= char <= '\uD7A3' for char in prompt):
            translated = translator(prompt, max_length=512)
            prompt = translated[0]['translation_text']
        
        resolution = (576, 1024)
        save_fps = 8
        seed_everything(seed)
        transform = transforms.Compose([transforms.Resize(resolution, antialias=True)])
        
        print('start:', prompt, time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time())))
        start = time.time()
        if steps > 60:
            steps = 60 

        batch_size = 1
        channels = model.diffusion_model.out_channels  # model.module 제거
        h, w = resolution[0] // 8, resolution[1] // 8
        noise_shape = [batch_size, channels, frames, h, w]

        with torch.no_grad(), torch.cuda.amp.autocast():
            text_emb = model.get_learned_conditioning([prompt])  # model.module 제거
            
            img_tensor = image.to(torch.cuda.current_device())
            img_tensor = (img_tensor - 0.5) * 2
            image_tensor_resized = transform(img_tensor)
            videos = image_tensor_resized.unsqueeze(0)
            
            z = get_latent_z(model, videos.unsqueeze(2))  # model.module 제거
            img_tensor_repeat = repeat(z, 'b c t h w -> b c (repeat t) h w', repeat=frames)
        
            cond_images = model.embedder(img_tensor.unsqueeze(0))  # model.module 제거
            img_emb = model.image_proj_model(cond_images)  # model.module 제거
        
            imtext_cond = torch.cat([text_emb, img_emb], dim=1)
        
            fs = torch.tensor([fs], dtype=torch.long, device=torch.cuda.current_device())
            cond = {"c_crossattn": [imtext_cond], "fs": fs, "c_concat": [img_tensor_repeat]}
            
            batch_samples = batch_ddim_sampling(model, cond, noise_shape, n_samples=1, ddim_steps=steps, ddim_eta=eta, cfg_scale=cfg_scale)  # model.module 제거
        
            video_path = './output.mp4'
            save_videos(batch_samples, './', filenames=['output'], fps=save_fps)
        
        # 메모리 정리
        del text_emb, img_tensor, image_tensor_resized, videos, z, img_tensor_repeat, cond_images, img_emb, imtext_cond, cond, batch_samples
        torch.cuda.empty_cache()
        
        return video_path

    except Exception as e:
        print(f"Error occurred: {e}")
        return None
    finally:
        torch.cuda.empty_cache()


i2v_examples = [
    ['우주인 복장으로 기타를 치는 남자', 30, 7.5, 1.0, 6, 123, 64],
    ['time-lapse of a blooming flower with leaves and a stem', 30, 7.5, 1.0, 10, 123, 64],
]

css = """#output_vid {max-width: 1024px; max-height: 576px}"""

with gr.Blocks(analytics_enabled=False, css=css) as dynamicrafter_iface:
    
    with gr.Tab(label='ImageAnimation_576x1024'):
        with gr.Column():
            with gr.Row():
                with gr.Column():
                    with gr.Row():
                        i2v_input_text = gr.Textbox(label='Prompts (한글 입력 가능)')
                    with gr.Row():
                        i2v_seed = gr.Slider(label='Random Seed', minimum=0, maximum=10000, step=1, value=123)
                        i2v_eta = gr.Slider(minimum=0.0, maximum=1.0, step=0.1, label='ETA', value=1.0, elem_id="i2v_eta")
                        i2v_cfg_scale = gr.Slider(minimum=1.0, maximum=15.0, step=0.5, label='CFG Scale', value=7.5, elem_id="i2v_cfg_scale")
                    with gr.Row():
                        i2v_steps = gr.Slider(minimum=1, maximum=50, step=1, elem_id="i2v_steps", label="Sampling steps", value=50)
                        i2v_motion = gr.Slider(minimum=5, maximum=20, step=1, elem_id="i2v_motion", label="FPS", value=10)
                        i2v_frames = gr.Slider(minimum=16, maximum=128, step=16, elem_id="i2v_frames", label="Number of frames", value=64)
                    i2v_end_btn = gr.Button("Generate")
                with gr.Row():
                    i2v_output_video = gr.Video(label="Generated Video",elem_id="output_vid",autoplay=True,show_share_button=True)

            gr.Examples(examples=i2v_examples,
                        inputs=[i2v_input_text, i2v_steps, i2v_cfg_scale, i2v_eta, i2v_motion, i2v_seed, i2v_frames],
                        outputs=[i2v_output_video],
                        fn = infer,
                        cache_examples=False
            )
        i2v_end_btn.click(inputs=[i2v_input_text, i2v_steps, i2v_cfg_scale, i2v_eta, i2v_motion, i2v_seed, i2v_frames],
                        outputs=[i2v_output_video],
                        fn = infer
        )

dynamicrafter_iface.launch()