Update app.py

app.py

@@ -1,10 +1,10 @@
 # -*- coding: utf-8 -*-
-import spaces
 import gradio as gr
 import os
 import sys
 import random
 import time
+import uuid
 from omegaconf import OmegaConf
 import torch
 import torchvision
@@ -21,6 +21,19 @@ from funcs import (
     save_videos
 )
 from transformers import pipeline
+from diffusers import StableDiffusionXLPipeline
+#import spaces
+import tensorflow as tf
+print(tf.__version__)
+
+print("GPU available:", len(tf.config.list_physical_devices('GPU')) > 0)
+
+
+
+def is_tensor(x):
+    return tf.is_tensor(x)
+    
+os.environ['KERAS_BACKEND'] = 'tensorflow'
 
 def download_model():
     REPO_ID = 'Doubiiu/DynamiCrafter_1024'
@@ -45,11 +58,57 @@ model.eval()
 model = model.cuda()
 
 # 번역 모델 로드
-translator = pipeline("translation", model="Helsinki-NLP/opus-mt-ko-en")
+translator = pipeline("translation", model="Helsinki-NLP/opus-mt-ko-en", device=0 if torch.cuda.is_available() else -1)
+
+# 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)
+
+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):
 
-@spaces.GPU(duration=300, gpu_type="l40s")
-def infer(image, prompt, steps=50, cfg_scale=7.5, eta=1.0, fs=3, seed=123):
     try:
+        # 이미지 생성
+        image_path = generate_image(prompt)
+        image = torchvision.io.read_image(image_path).float() / 255.0
+        
         # 한글 입력 확인 및 번역
         if any('\u3131' <= char <= '\u318E' or '\uAC00' <= char <= '\uD7A3' for char in prompt):
             translated = translator(prompt, max_length=512)
@@ -59,8 +118,7 @@ def infer(image, prompt, steps=50, cfg_scale=7.5, eta=1.0, fs=3, seed=123):
         save_fps = 8
         seed_everything(seed)
         transform = transforms.Compose([
-            transforms.Resize(min(resolution), antialias=True),
-            transforms.CenterCrop(resolution),
+            transforms.Resize(resolution, antialias=True),
         ])
         
         print('start:', prompt, time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time())))
@@ -70,15 +128,14 @@ def infer(image, prompt, steps=50, cfg_scale=7.5, eta=1.0, fs=3, seed=123):
 
         batch_size = 1
         channels = model.model.diffusion_model.out_channels
-        frames = model.temporal_length
         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])
             
-            img_tensor = torch.from_numpy(image).permute(2, 0, 1).float().to(model.device)
-            img_tensor = (img_tensor / 255. - 0.5) * 2
+            img_tensor = image.to(model.device)
+            img_tensor = (img_tensor - 0.5) * 2
             image_tensor_resized = transform(img_tensor)
             videos = image_tensor_resized.unsqueeze(0)
             
@@ -110,11 +167,11 @@ def infer(image, prompt, steps=50, cfg_scale=7.5, eta=1.0, fs=3, seed=123):
         torch.cuda.empty_cache()
 
 i2v_examples = [
-    ['prompts/1024/astronaut04.png', '우주인 복장으로 기타를 치는 남자', 30, 7.5, 1.0, 6, 123],
-    ['prompts/1024/bloom01.png', 'time-lapse of a blooming flower with leaves and a stem', 30, 7.5, 1.0, 10, 123],
+    ['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 = """#input_img {max-width: 1024px !important} #output_vid {max-width: 1024px; max-height: 576px}"""
+css = """#output_vid {max-width: 1024px; max-height: 576px}"""
 
 with gr.Blocks(analytics_enabled=False, css=css) as dynamicrafter_iface:
     
@@ -122,8 +179,6 @@ with gr.Blocks(analytics_enabled=False, css=css) as dynamicrafter_iface:
         with gr.Column():
             with gr.Row():
                 with gr.Column():
-                    with gr.Row():
-                        i2v_input_image = gr.Image(label="Input Image",elem_id="input_img")
                     with gr.Row():
                         i2v_input_text = gr.Textbox(label='Prompts (한글 입력 가능)')
                     with gr.Row():
@@ -133,19 +188,21 @@ with gr.Blocks(analytics_enabled=False, css=css) as dynamicrafter_iface:
                     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_image, i2v_input_text, i2v_steps, i2v_cfg_scale, i2v_eta, i2v_motion, i2v_seed],
+                        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  # 이 부분을 False로 설정하여 캐시를 비활성화
+                        cache_examples=False
             )
-        i2v_end_btn.click(inputs=[i2v_input_image, i2v_input_text, i2v_steps, i2v_cfg_scale, i2v_eta, i2v_motion, i2v_seed],
+        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(server_port=7890, server_name="0.0.0.0", share=True)
+dynamicrafter_iface.launch(server_port=7930, server_name="0.0.0.0", share=True)
+