ginigen-sora / app-backup1.py
openfree's picture
Update app-backup1.py
5beda4f verified
raw
history blame
48.2 kB
import gradio as gr
from gradio_toggle import Toggle
import torch
from huggingface_hub import snapshot_download
from transformers import pipeline
from xora.models.autoencoders.causal_video_autoencoder import CausalVideoAutoencoder
from xora.models.transformers.transformer3d import Transformer3DModel
from xora.models.transformers.symmetric_patchifier import SymmetricPatchifier
from xora.schedulers.rf import RectifiedFlowScheduler
from xora.pipelines.pipeline_xora_video import XoraVideoPipeline
from transformers import T5EncoderModel, T5Tokenizer
from xora.utils.conditioning_method import ConditioningMethod
from pathlib import Path
import safetensors.torch
import json
import numpy as np
import cv2
from PIL import Image
import tempfile
import os
import gc
from openai import OpenAI
import re
import time
# Load system prompts
system_prompt_t2v = """당신은 λΉ„λ””μ˜€ 생성을 μœ„ν•œ ν”„λ‘¬ν”„νŠΈ μ „λ¬Έκ°€μž…λ‹ˆλ‹€.
주어진 ν”„λ‘¬ν”„νŠΈλ₯Ό λ‹€μŒ ꡬ쑰에 맞게 κ°œμ„ ν•΄μ£Όμ„Έμš”:
1. μ£Όμš” λ™μž‘μ„ λͺ…ν™•ν•œ ν•œ λ¬Έμž₯으둜 μ‹œμž‘
2. ꡬ체적인 λ™μž‘κ³Ό 제슀처λ₯Ό μ‹œκ°„ μˆœμ„œλŒ€λ‘œ μ„€λͺ…
3. 캐릭터/객체의 μ™Έλͺ¨λ₯Ό μƒμ„Ένžˆ λ¬˜μ‚¬
4. λ°°κ²½κ³Ό ν™˜κ²½ μ„ΈλΆ€ 사항을 ꡬ체적으둜 포함
5. 카메라 각도와 μ›€μ§μž„μ„ λͺ…μ‹œ
6. μ‘°λͺ…κ³Ό 색상을 μžμ„Ένžˆ μ„€λͺ…
7. λ³€ν™”λ‚˜ κ°‘μž‘μŠ€λŸ¬μš΄ 사건을 μžμ—°μŠ€λŸ½κ²Œ 포함
λͺ¨λ“  μ„€λͺ…은 ν•˜λ‚˜μ˜ μžμ—°μŠ€λŸ¬μš΄ λ¬Έλ‹¨μœΌλ‘œ μž‘μ„±ν•˜κ³ ,
촬영 감독이 촬영 λͺ©λ‘μ„ μ„€λͺ…ν•˜λŠ” κ²ƒμ²˜λŸΌ ꡬ체적이고 μ‹œκ°μ μœΌλ‘œ μž‘μ„±ν•˜μ„Έμš”.
200단어λ₯Ό λ„˜μ§€ μ•Šλ„λ‘ ν•˜λ˜, μ΅œλŒ€ν•œ μƒμ„Έν•˜κ²Œ μž‘μ„±ν•˜μ„Έμš”."""
system_prompt_i2v = """당신은 이미지 기반 λΉ„λ””μ˜€ 생성을 μœ„ν•œ ν”„λ‘¬ν”„νŠΈ μ „λ¬Έκ°€μž…λ‹ˆλ‹€.
주어진 ν”„λ‘¬ν”„νŠΈλ₯Ό λ‹€μŒ ꡬ쑰에 맞게 κ°œμ„ ν•΄μ£Όμ„Έμš”:
1. μ£Όμš” λ™μž‘μ„ λͺ…ν™•ν•œ ν•œ λ¬Έμž₯으둜 μ‹œμž‘
2. ꡬ체적인 λ™μž‘κ³Ό 제슀처λ₯Ό μ‹œκ°„ μˆœμ„œλŒ€λ‘œ μ„€λͺ…
3. 캐릭터/객체의 μ™Έλͺ¨λ₯Ό μƒμ„Ένžˆ λ¬˜μ‚¬
4. λ°°κ²½κ³Ό ν™˜κ²½ μ„ΈλΆ€ 사항을 ꡬ체적으둜 포함
5. 카메라 각도와 μ›€μ§μž„μ„ λͺ…μ‹œ
6. μ‘°λͺ…κ³Ό 색상을 μžμ„Ένžˆ μ„€λͺ…
7. λ³€ν™”λ‚˜ κ°‘μž‘μŠ€λŸ¬μš΄ 사건을 μžμ—°μŠ€λŸ½κ²Œ 포함
λͺ¨λ“  μ„€λͺ…은 ν•˜λ‚˜μ˜ μžμ—°μŠ€λŸ¬μš΄ λ¬Έλ‹¨μœΌλ‘œ μž‘μ„±ν•˜κ³ ,
촬영 감독이 촬영 λͺ©λ‘μ„ μ„€λͺ…ν•˜λŠ” κ²ƒμ²˜λŸΌ ꡬ체적이고 μ‹œκ°μ μœΌλ‘œ μž‘μ„±ν•˜μ„Έμš”.
200단어λ₯Ό λ„˜μ§€ μ•Šλ„λ‘ ν•˜λ˜, μ΅œλŒ€ν•œ μƒμ„Έν•˜κ²Œ μž‘μ„±ν•˜μ„Έμš”."""
# Load Hugging Face token if needed
hf_token = os.getenv("HF_TOKEN")
openai_api_key = os.getenv("OPENAI_API_KEY")
client = OpenAI(api_key=openai_api_key)
# Initialize translation pipeline with device and clean_up settings
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
translator = pipeline(
"translation",
model="Helsinki-NLP/opus-mt-ko-en",
device=device,
clean_up_tokenization_spaces=True
)
# Korean text detection function
def contains_korean(text):
korean_pattern = re.compile('[γ„±-γ…Žγ…-γ…£κ°€-힣]')
return bool(korean_pattern.search(text))
def translate_korean_prompt(prompt, max_length=450):
"""
Translate Korean prompt to English if Korean text is detected
Split long text into chunks if necessary
"""
if not contains_korean(prompt):
return prompt
# Split long text into chunks
def split_text(text, max_length):
words = text.split()
chunks = []
current_chunk = []
current_length = 0
for word in words:
if current_length + len(word) + 1 > max_length:
chunks.append(' '.join(current_chunk))
current_chunk = [word]
current_length = len(word)
else:
current_chunk.append(word)
current_length += len(word) + 1
if current_chunk:
chunks.append(' '.join(current_chunk))
return chunks
try:
if len(prompt) > max_length:
chunks = split_text(prompt, max_length)
translated_chunks = []
for chunk in chunks:
translated = translator(chunk, max_length=512)[0]['translation_text']
translated_chunks.append(translated)
final_translation = ' '.join(translated_chunks)
else:
final_translation = translator(prompt, max_length=512)[0]['translation_text']
print(f"Original Korean prompt: {prompt}")
print(f"Translated English prompt: {final_translation}")
return final_translation
except Exception as e:
print(f"Translation error: {e}")
return prompt # Return original prompt if translation fails
def enhance_prompt(prompt, type="t2v"):
system_prompt = system_prompt_t2v if type == "t2v" else system_prompt_i2v
messages = [
{"role": "system", "content": system_prompt},
{"role": "user", "content": prompt},
]
try:
response = client.chat.completions.create(
model="gpt-4-1106-preview",
messages=messages,
max_tokens=2000,
)
enhanced_prompt = response.choices[0].message.content.strip()
print("\n=== ν”„λ‘¬ν”„νŠΈ 증강 κ²°κ³Ό ===")
print("Original Prompt:")
print(prompt)
print("\nEnhanced Prompt:")
print(enhanced_prompt)
print("========================\n")
return enhanced_prompt
except Exception as e:
print(f"Error during prompt enhancement: {e}")
return prompt
def update_prompt_t2v(prompt, enhance_toggle):
return update_prompt(prompt, enhance_toggle, "t2v")
def update_prompt_i2v(prompt, enhance_toggle):
return update_prompt(prompt, enhance_toggle, "i2v")
def update_prompt(prompt, enhance_toggle, type="t2v"):
if enhance_toggle:
return enhance_prompt(prompt, type)
return prompt
# Set model download directory within Hugging Face Spaces
model_path = "asset"
if not os.path.exists(model_path):
snapshot_download(
"Lightricks/LTX-Video", local_dir=model_path, repo_type="model", token=hf_token
)
# Global variables to load components
vae_dir = Path(model_path) / "vae"
unet_dir = Path(model_path) / "unet"
scheduler_dir = Path(model_path) / "scheduler"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
def load_vae(vae_dir):
vae_ckpt_path = vae_dir / "vae_diffusion_pytorch_model.safetensors"
vae_config_path = vae_dir / "config.json"
with open(vae_config_path, "r") as f:
vae_config = json.load(f)
vae = CausalVideoAutoencoder.from_config(vae_config)
vae_state_dict = safetensors.torch.load_file(vae_ckpt_path)
vae.load_state_dict(vae_state_dict)
return vae.to(device=device, dtype=torch.bfloat16)
def load_unet(unet_dir):
unet_ckpt_path = unet_dir / "unet_diffusion_pytorch_model.safetensors"
unet_config_path = unet_dir / "config.json"
transformer_config = Transformer3DModel.load_config(unet_config_path)
transformer = Transformer3DModel.from_config(transformer_config)
unet_state_dict = safetensors.torch.load_file(unet_ckpt_path)
transformer.load_state_dict(unet_state_dict, strict=True)
return transformer.to(device=device, dtype=torch.bfloat16)
def load_scheduler(scheduler_dir):
scheduler_config_path = scheduler_dir / "scheduler_config.json"
scheduler_config = RectifiedFlowScheduler.load_config(scheduler_config_path)
return RectifiedFlowScheduler.from_config(scheduler_config)
# Helper function for image processing
def center_crop_and_resize(frame, target_height, target_width):
h, w, _ = frame.shape
aspect_ratio_target = target_width / target_height
aspect_ratio_frame = w / h
if aspect_ratio_frame > aspect_ratio_target:
new_width = int(h * aspect_ratio_target)
x_start = (w - new_width) // 2
frame_cropped = frame[:, x_start : x_start + new_width]
else:
new_height = int(w / aspect_ratio_target)
y_start = (h - new_height) // 2
frame_cropped = frame[y_start : y_start + new_height, :]
frame_resized = cv2.resize(frame_cropped, (target_width, target_height))
return frame_resized
def load_image_to_tensor_with_resize(image_path, target_height=512, target_width=768):
image = Image.open(image_path).convert("RGB")
image_np = np.array(image)
frame_resized = center_crop_and_resize(image_np, target_height, target_width)
frame_tensor = torch.tensor(frame_resized).permute(2, 0, 1).float()
frame_tensor = (frame_tensor / 127.5) - 1.0
return frame_tensor.unsqueeze(0).unsqueeze(2)
# Load models
vae = load_vae(vae_dir)
unet = load_unet(unet_dir)
scheduler = load_scheduler(scheduler_dir)
patchifier = SymmetricPatchifier(patch_size=1)
text_encoder = T5EncoderModel.from_pretrained(
"PixArt-alpha/PixArt-XL-2-1024-MS", subfolder="text_encoder"
).to(device)
tokenizer = T5Tokenizer.from_pretrained(
"PixArt-alpha/PixArt-XL-2-1024-MS", subfolder="tokenizer"
)
pipeline = XoraVideoPipeline(
transformer=unet,
patchifier=patchifier,
text_encoder=text_encoder,
tokenizer=tokenizer,
scheduler=scheduler,
vae=vae,
).to(device)
# Preset options for resolution and frame configuration
# Convert frames to seconds assuming 25 FPS
preset_options = [
{"label": "[16:9 HD] 1216x704, 1.6초", "width": 1216, "height": 704, "num_frames": 41},
{"label": "[16:9] 1088x704, 2.0초", "width": 1088, "height": 704, "num_frames": 49},
{"label": "[16:9] 1056x640, 2.3초", "width": 1056, "height": 640, "num_frames": 57},
{"label": "[16:9] 992x608, 2.6초", "width": 992, "height": 608, "num_frames": 65},
{"label": "[16:9] 896x608, 2.9초", "width": 896, "height": 608, "num_frames": 73},
{"label": "[16:9] 896x544, 3.2초", "width": 896, "height": 544, "num_frames": 81},
{"label": "[16:9] 832x544, 3.6초", "width": 832, "height": 544, "num_frames": 89},
{"label": "[16:9] 800x512, 3.9초", "width": 800, "height": 512, "num_frames": 97},
{"label": "[16:9] 768x512, 3.9초", "width": 768, "height": 512, "num_frames": 97},
{"label": "[16:9] 800x480, 4.2초", "width": 800, "height": 480, "num_frames": 105},
{"label": "[16:9] 736x480, 4.5초", "width": 736, "height": 480, "num_frames": 113},
{"label": "[3:2] 704x480, 4.8초", "width": 704, "height": 480, "num_frames": 121},
{"label": "[16:9] 704x448, 5.2초", "width": 704, "height": 448, "num_frames": 129},
{"label": "[16:9] 672x448, 5.5초", "width": 672, "height": 448, "num_frames": 137},
{"label": "[16:9] 640x416, 6.1초", "width": 640, "height": 416, "num_frames": 153},
{"label": "[16:9] 672x384, 6.4초", "width": 672, "height": 384, "num_frames": 161},
{"label": "[16:9] 640x384, 6.8초", "width": 640, "height": 384, "num_frames": 169},
{"label": "[16:9] 608x384, 7.1초", "width": 608, "height": 384, "num_frames": 177},
{"label": "[16:9] 576x384, 7.4초", "width": 576, "height": 384, "num_frames": 185},
{"label": "[16:9] 608x352, 7.7초", "width": 608, "height": 352, "num_frames": 193},
{"label": "[16:9] 576x352, 8.0초", "width": 576, "height": 352, "num_frames": 201},
{"label": "[16:9] 544x352, 8.4초", "width": 544, "height": 352, "num_frames": 209},
{"label": "[3:2] 512x352, 9.3초", "width": 512, "height": 352, "num_frames": 233},
{"label": "[16:9] 544x320, 9.6초", "width": 544, "height": 320, "num_frames": 241},
{"label": "[16:9] 512x320, 10.3초", "width": 512, "height": 320, "num_frames": 257},
]
def preset_changed(preset):
selected = next((item for item in preset_options if item["label"] == preset), None)
if selected is None:
raise gr.Error("Invalid preset selected")
return [
gr.State(value=selected["height"]),
gr.State(value=selected["width"]),
gr.State(value=selected["num_frames"]),
gr.update(visible=False),
gr.update(visible=False),
gr.update(visible=False),
]
def generate_video_from_text(
prompt,
enhance_prompt_toggle,
negative_prompt,
frame_rate,
seed,
num_inference_steps,
guidance_scale,
height,
width,
num_frames,
progress=gr.Progress(),
):
if len(prompt.strip()) < 50:
raise gr.Error(
"ν”„λ‘¬ν”„νŠΈλŠ” μ΅œμ†Œ 50자 이상이어야 ν•©λ‹ˆλ‹€. 더 μžμ„Έν•œ μ„€λͺ…을 μ œκ³΅ν•΄μ£Όμ„Έμš”.",
duration=5,
)
# ν”„λ‘¬ν”„νŠΈ κ°œμ„ μ΄ ν™œμ„±ν™”λœ 경우
if enhance_prompt_toggle:
prompt = enhance_prompt(prompt, "t2v")
# Translate Korean prompts to English
prompt = translate_korean_prompt(prompt)
negative_prompt = translate_korean_prompt(negative_prompt)
# κΈ°λ³Έκ°’ μ„€μ •
height = height or 320
width = width or 512
num_frames = num_frames or 257
frame_rate = frame_rate or 25
seed = seed or 171198
num_inference_steps = num_inference_steps or 41
guidance_scale = guidance_scale or 4.0
sample = {
"prompt": prompt,
"prompt_attention_mask": None,
"negative_prompt": negative_prompt,
"negative_prompt_attention_mask": None,
"media_items": None,
}
generator = torch.Generator(device="cpu").manual_seed(seed)
def gradio_progress_callback(self, step, timestep, kwargs):
progress((step + 1) / num_inference_steps)
try:
with torch.no_grad():
images = pipeline(
num_inference_steps=num_inference_steps,
num_images_per_prompt=1,
guidance_scale=guidance_scale,
generator=generator,
output_type="pt",
height=height,
width=width,
num_frames=num_frames,
frame_rate=frame_rate,
**sample,
is_video=True,
vae_per_channel_normalize=True,
conditioning_method=ConditioningMethod.UNCONDITIONAL,
mixed_precision=True,
callback_on_step_end=gradio_progress_callback,
).images
except Exception as e:
raise gr.Error(
f"λΉ„λ””μ˜€ 생성 쀑 였λ₯˜κ°€ λ°œμƒν–ˆμŠ΅λ‹ˆλ‹€. λ‹€μ‹œ μ‹œλ„ν•΄μ£Όμ„Έμš”. 였λ₯˜: {e}",
duration=5,
)
finally:
torch.cuda.empty_cache()
gc.collect()
output_path = tempfile.mktemp(suffix=".mp4")
video_np = images.squeeze(0).permute(1, 2, 3, 0).cpu().float().numpy()
video_np = (video_np * 255).astype(np.uint8)
height, width = video_np.shape[1:3]
out = cv2.VideoWriter(
output_path, cv2.VideoWriter_fourcc(*"mp4v"), frame_rate, (width, height)
)
for frame in video_np[..., ::-1]:
out.write(frame)
out.release()
del images
del video_np
torch.cuda.empty_cache()
return output_path
def generate_video_from_image(
image_path,
prompt,
enhance_prompt_toggle,
negative_prompt,
frame_rate,
seed,
num_inference_steps,
guidance_scale,
height,
width,
num_frames,
progress=gr.Progress(),
):
if not image_path:
raise gr.Error("μž…λ ₯ 이미지λ₯Ό μ œκ³΅ν•΄μ£Όμ„Έμš”.", duration=5)
if len(prompt.strip()) < 50:
raise gr.Error(
"ν”„λ‘¬ν”„νŠΈλŠ” μ΅œμ†Œ 50자 이상이어야 ν•©λ‹ˆλ‹€. 더 μžμ„Έν•œ μ„€λͺ…을 μ œκ³΅ν•΄μ£Όμ„Έμš”.",
duration=5,
)
# ν”„λ‘¬ν”„νŠΈ κ°œμ„ μ΄ ν™œμ„±ν™”λœ 경우
if enhance_prompt_toggle:
prompt = enhance_prompt(prompt, "i2v")
# Translate Korean prompts to English
prompt = translate_korean_prompt(prompt)
negative_prompt = translate_korean_prompt(negative_prompt)
# κΈ°λ³Έκ°’ μ„€μ •
height = height or 320
width = width or 512
num_frames = num_frames or 257
frame_rate = frame_rate or 25
seed = seed or 171198
num_inference_steps = num_inference_steps or 41
guidance_scale = guidance_scale or 4.0
# 이미지 λ‘œλ“œ 및 μ „μ²˜λ¦¬
media_items = (
load_image_to_tensor_with_resize(image_path, height, width).to(device).detach()
)
sample = {
"prompt": prompt,
"prompt_attention_mask": None,
"negative_prompt": negative_prompt,
"negative_prompt_attention_mask": None,
"media_items": media_items,
}
generator = torch.Generator(device="cpu").manual_seed(seed)
def gradio_progress_callback(self, step, timestep, kwargs):
progress((step + 1) / num_inference_steps)
try:
with torch.no_grad():
images = pipeline(
num_inference_steps=num_inference_steps,
num_images_per_prompt=1,
guidance_scale=guidance_scale,
generator=generator,
output_type="pt",
height=height,
width=width,
num_frames=num_frames,
frame_rate=frame_rate,
**sample,
is_video=True,
vae_per_channel_normalize=True,
conditioning_method=ConditioningMethod.FIRST_FRAME,
mixed_precision=True,
callback_on_step_end=gradio_progress_callback,
).images
output_path = tempfile.mktemp(suffix=".mp4")
video_np = images.squeeze(0).permute(1, 2, 3, 0).cpu().float().numpy()
video_np = (video_np * 255).astype(np.uint8)
height, width = video_np.shape[1:3]
out = cv2.VideoWriter(
output_path, cv2.VideoWriter_fourcc(*"mp4v"), frame_rate, (width, height)
)
for frame in video_np[..., ::-1]:
out.write(frame)
out.release()
except Exception as e:
raise gr.Error(
f"λΉ„λ””μ˜€ 생성 쀑 였λ₯˜κ°€ λ°œμƒν–ˆμŠ΅λ‹ˆλ‹€. λ‹€μ‹œ μ‹œλ„ν•΄μ£Όμ„Έμš”. 였λ₯˜: {e}",
duration=5,
)
finally:
torch.cuda.empty_cache()
gc.collect()
if 'images' in locals():
del images
if 'video_np' in locals():
del video_np
if 'media_items' in locals():
del media_items
return output_path
def create_advanced_options():
with gr.Accordion("Step 4: Advanced Options (Optional)", open=False):
seed = gr.Slider(
label="Seed",
minimum=0,
maximum=1000000,
step=1,
value=171198
)
inference_steps = gr.Slider(
label="4.2 Inference Steps",
minimum=1,
maximum=50,
step=1,
value=41,
visible=False
)
guidance_scale = gr.Slider(
label="4.3 Guidance Scale",
minimum=1.0,
maximum=5.0,
step=0.1,
value=4.0,
visible=False
)
height_slider = gr.Slider(
label="4.4 Height",
minimum=256,
maximum=1024,
step=64,
value=320,
visible=False,
)
width_slider = gr.Slider(
label="4.5 Width",
minimum=256,
maximum=1024,
step=64,
value=512,
visible=False,
)
num_frames_slider = gr.Slider(
label="4.5 Number of Frames",
minimum=1,
maximum=200,
step=1,
value=257,
visible=False,
)
return [
seed,
inference_steps,
guidance_scale,
height_slider,
width_slider,
num_frames_slider,
]
system_prompt_scenario = """당신은 μ˜μƒ μŠ€ν¬λ¦½νŠΈμ— λ§žλŠ” λ°°κ²½ μ˜μƒμ„ μƒμ„±ν•˜κΈ° μœ„ν•œ ν”„λ‘¬ν”„νŠΈ μ „λ¬Έκ°€μž…λ‹ˆλ‹€.
주어진 슀크립트의 λΆ„μœ„κΈ°μ™€ λ§₯락을 μ‹œκ°μ  배경으둜 ν‘œν˜„ν•˜λ˜, λ‹€μŒ 원칙을 λ°˜λ“œμ‹œ μ€€μˆ˜ν•˜μ„Έμš”:
1. μ œν’ˆμ΄λ‚˜ μ„œλΉ„μŠ€λ₯Ό μ§μ ‘μ μœΌλ‘œ λ¬˜μ‚¬ν•˜μ§€ 말 것
2. 슀크립트의 감성과 ν†€μ•€λ§€λ„ˆλ₯Ό ν‘œν˜„ν•˜λŠ” λ°°κ²½ μ˜μƒμ— 집쀑할 것
3. 5개 μ„Ήμ…˜μ΄ ν•˜λ‚˜μ˜ μ΄μ•ΌκΈ°μ²˜λŸΌ μžμ—°μŠ€λŸ½κ²Œ μ—°κ²°λ˜λ„λ‘ ν•  것
4. 좔상적이고 μ€μœ μ μΈ μ‹œκ° ν‘œν˜„μ„ ν™œμš©ν•  것
각 μ„Ήμ…˜λ³„ ν”„λ‘¬ν”„νŠΈ μž‘μ„± κ°€μ΄λ“œ:
1. λ°°κ²½ 및 ν•„μš”μ„±: 주제의 μ „λ°˜μ μΈ λΆ„μœ„κΈ°λ₯Ό ν‘œν˜„ν•˜λŠ” λ°°κ²½ 씬
2. 문제 제기: κΈ΄μž₯κ°μ΄λ‚˜ κ°ˆλ“±μ„ μ•”μ‹œν•˜λŠ” λΆ„μœ„κΈ° μžˆλŠ” λ°°κ²½
3. ν•΄κ²°μ±… μ œμ‹œ: 희망적이고 밝은 ν†€μ˜ λ°°κ²½ μ „ν™˜
4. λ³Έλ‘ : μ•ˆμ •κ° 있고 신뒰도λ₯Ό λ†’μ΄λŠ” λ°°κ²½
5. κ²°λ‘ : μž„νŒ©νŠΈ μžˆλŠ” 마무리λ₯Ό μœ„ν•œ 역동적인 λ°°κ²½
λͺ¨λ“  μ„Ήμ…˜μ΄ μΌκ΄€λœ μŠ€νƒ€μΌκ³Ό 톀을 μœ μ§€ν•˜λ©΄μ„œλ„ μžμ—°μŠ€λŸ½κ²Œ 이어지도둝 κ΅¬μ„±ν•˜μ„Έμš”.
각 μ„Ήμ…˜μ˜ ν”„λ‘¬ν”„νŠΈ μž‘μ„±μ‹œ λ°˜λ“œμ‹œ λ‹€μŒ ꡬ쑰에 맞게 κ°œμ„ ν•΄μ£Όμ„Έμš”:
1. μ£Όμš” λ™μž‘μ„ λͺ…ν™•ν•œ ν•œ λ¬Έμž₯으둜 μ‹œμž‘
2. ꡬ체적인 λ™μž‘κ³Ό 제슀처λ₯Ό μ‹œκ°„ μˆœμ„œλŒ€λ‘œ μ„€λͺ…
3. 캐릭터/객체의 μ™Έλͺ¨λ₯Ό μƒμ„Ένžˆ λ¬˜μ‚¬
4. λ°°κ²½κ³Ό ν™˜κ²½ μ„ΈλΆ€ 사항을 ꡬ체적으둜 포함
5. 카메라 각도와 μ›€μ§μž„μ„ λͺ…μ‹œ
6. μ‘°λͺ…κ³Ό 색상을 μžμ„Ένžˆ μ„€λͺ…
7. λ³€ν™”λ‚˜ κ°‘μž‘μŠ€λŸ¬μš΄ 사건을 μžμ—°μŠ€λŸ½κ²Œ 포함
λͺ¨λ“  μ„€λͺ…은 ν•˜λ‚˜μ˜ μžμ—°μŠ€λŸ¬μš΄ λ¬Έλ‹¨μœΌλ‘œ μž‘μ„±ν•˜κ³ ,
촬영 감독이 촬영 λͺ©λ‘μ„ μ„€λͺ…ν•˜λŠ” κ²ƒμ²˜λŸΌ ꡬ체적이고 μ‹œκ°μ μœΌλ‘œ μž‘μ„±ν•˜μ„Έμš”.
200단어λ₯Ό λ„˜μ§€ μ•Šλ„λ‘ ν•˜λ˜, μ΅œλŒ€ν•œ μƒμ„Έν•˜κ²Œ μž‘μ„±ν•˜μ„Έμš”.
"""
def analyze_scenario(scenario):
"""μ‹œλ‚˜λ¦¬μ˜€λ₯Ό λΆ„μ„ν•˜μ—¬ 각 μ„Ήμ…˜λ³„ λ°°κ²½ μ˜μƒμš© ν”„λ‘¬ν”„νŠΈ 생성"""
try:
# 각 μ„Ήμ…˜λ³„ ν”„λ‘¬ν”„νŠΈ 생성을 μœ„ν•œ λ©”μ‹œμ§€ ꡬ성
section_prompts = []
for section_num in range(1, 6):
section_descriptions = {
1: "λ°°κ²½ 및 ν•„μš”μ„±: 주제의 μ „λ°˜μ μΈ λΆ„μœ„κΈ°λ₯Ό ν‘œν˜„ν•˜λŠ” λ°°κ²½ 씬",
2: "ν₯λ―Έ 유발: κΈ΄μž₯κ°μ΄λ‚˜ κ°ˆλ“±μ„ μ•”μ‹œν•˜λŠ” λΆ„μœ„κΈ° μžˆλŠ” λ°°κ²½",
3: "ν•΄κ²°μ±… μ œμ‹œ: 희망적이고 밝은 ν†€μ˜ λ°°κ²½ μ „ν™˜",
4: "λ³Έλ‘ : μ•ˆμ •κ° 있고 신뒰도λ₯Ό λ†’μ΄λŠ” λ°°κ²½",
5: "κ²°λ‘ : μž„νŒ©νŠΈ μžˆλŠ” 마무리λ₯Ό μœ„ν•œ 역동적인 λ°°κ²½"
}
messages = [
{"role": "system", "content": system_prompt_scenario},
{"role": "user", "content": f"""
λ‹€μŒ 슀크립트의 {section_num}번째 μ„Ήμ…˜({section_descriptions[section_num]})에 λŒ€ν•œ
λ°°κ²½ μ˜μƒ ν”„λ‘¬ν”„νŠΈλ₯Ό μƒμ„±ν•΄μ£Όμ„Έμš”.
슀크립트:
{scenario}
μ£Όμ˜μ‚¬ν•­:
1. ν•΄λ‹Ή μ„Ήμ…˜μ˜ νŠΉμ„±({section_descriptions[section_num]})에 λ§žλŠ” λΆ„μœ„κΈ°μ™€ 톀을 λ°˜μ˜ν•˜μ„Έμš”.
2. 직접적인 μ œν’ˆ/μ„œλΉ„μŠ€ λ¬˜μ‚¬λŠ” ν”Όν•˜κ³ , 감성적이고 μ€μœ μ μΈ λ°°κ²½ μ˜μƒμ— μ§‘μ€‘ν•˜μ„Έμš”.
3. λ‹€μŒ ꡬ쑰λ₯Ό λ°˜λ“œμ‹œ ν¬ν•¨ν•˜μ„Έμš”:
- μ£Όμš” λ™μž‘μ„ λͺ…ν™•ν•œ ν•œ λ¬Έμž₯으둜 μ‹œμž‘
- ꡬ체적인 λ™μž‘κ³Ό 제슀처λ₯Ό μ‹œκ°„ μˆœμ„œλŒ€λ‘œ μ„€λͺ…
- λ°°κ²½κ³Ό ν™˜κ²½ μ„ΈλΆ€ 사항을 ꡬ체적으둜 포함
- 카메라 각도와 μ›€μ§μž„μ„ λͺ…μ‹œ
- μ‘°λͺ…κ³Ό 색상을 μžμ„Ένžˆ μ„€λͺ…
- λ³€ν™”λ‚˜ κ°‘μž‘μŠ€λŸ¬μš΄ 사건을 μžμ—°μŠ€λŸ½κ²Œ 포함"""}
]
response = client.chat.completions.create(
model="gpt-4-1106-preview",
messages=messages,
max_tokens=1000,
temperature=0.7
)
section_prompt = response.choices[0].message.content.strip()
section_prompts.append(f"{section_num}. {section_prompt}")
# API μš”μ²­ 사이에 짧은 λ”œλ ˆμ΄ μΆ”κ°€
time.sleep(1)
return section_prompts
except Exception as e:
print(f"Error during scenario analysis: {e}")
return ["Error occurred during analysis"] * 5
def generate_section_video(prompt, preset, section_number=1, base_seed=171198, progress=gr.Progress()):
"""각 μ„Ήμ…˜μ˜ λΉ„λ””μ˜€ 생성"""
try:
if not prompt or len(prompt.strip()) < 50:
raise gr.Error("ν”„λ‘¬ν”„νŠΈλŠ” μ΅œμ†Œ 50자 이상이어야 ν•©λ‹ˆλ‹€.")
if not preset:
raise gr.Error("해상도 프리셋을 μ„ νƒν•΄μ£Όμ„Έμš”.")
selected = next((item for item in preset_options if item["label"] == preset), None)
if not selected:
raise gr.Error("μ˜¬λ°”λ₯΄μ§€ μ•Šμ€ ν”„λ¦¬μ…‹μž…λ‹ˆλ‹€.")
section_seed = base_seed + section_number
return generate_video_from_text(
prompt=prompt,
enhance_prompt_toggle=False, # μ„Ήμ…˜ μƒμ„±μ‹œλŠ” ν”„λ‘¬ν”„νŠΈ 증강 λΉ„ν™œμ„±ν™”
negative_prompt="low quality, worst quality, deformed, distorted, warped",
frame_rate=25,
seed=section_seed,
num_inference_steps=41,
guidance_scale=4.0,
height=selected["height"],
width=selected["width"],
num_frames=selected["num_frames"],
progress=progress
)
except Exception as e:
print(f"Error in section {section_number}: {e}")
raise gr.Error(f"μ„Ήμ…˜ {section_number} 생성 쀑 였λ₯˜: {str(e)}")
finally:
torch.cuda.empty_cache()
gc.collect()
def generate_single_section_prompt(scenario, section_number):
"""κ°œλ³„ μ„Ήμ…˜μ— λŒ€ν•œ ν”„λ‘¬ν”„νŠΈ 생성"""
section_descriptions = {
1: "λ°°κ²½ 및 ν•„μš”μ„±: 주제의 μ „λ°˜μ μΈ λΆ„μœ„κΈ°λ₯Ό ν‘œν˜„ν•˜λŠ” λ°°κ²½ 씬",
2: "ν₯λ―Έ 유발: ν₯λ―Έλ₯Ό μœ λ°œν•˜κ³  κΈ°λŒ€κ°μ„ μ¦ν­μ‹œν‚€λŠ” λ°°κ²½",
3: "ν•΄κ²°μ±… μ œμ‹œ: 희망적이고 밝은 ν†€μ˜ λ°°κ²½ μ „ν™˜",
4: "λ³Έλ‘ : μ•ˆμ •κ° 있고 신뒰도λ₯Ό λ†’μ΄λŠ” λ°°κ²½",
5: "κ²°λ‘ : μž„νŒ©νŠΈ μžˆλŠ” 마무리λ₯Ό μœ„ν•œ 역동적인 λ°°κ²½"
}
messages = [
{"role": "system", "content": system_prompt_scenario},
{"role": "user", "content": f"""
λ‹€μŒ 슀크립트의 {section_number}번째 μ„Ήμ…˜({section_descriptions[section_number]})에 λŒ€ν•œ
λ°°κ²½ μ˜μƒ ν”„λ‘¬ν”„νŠΈλ₯Ό μƒμ„±ν•΄μ£Όμ„Έμš”.
슀크립트:
{scenario}
μ£Όμ˜μ‚¬ν•­:
1. ν•΄λ‹Ή μ„Ήμ…˜μ˜ νŠΉμ„±({section_descriptions[section_number]})에 λ§žλŠ” λΆ„μœ„κΈ°μ™€ 톀을 λ°˜μ˜ν•˜μ„Έμš”.
2. 직접적인 μ œν’ˆ/μ„œλΉ„μŠ€ λ¬˜μ‚¬λŠ” ν”Όν•˜κ³ , 감성적이고 μ€μœ μ μΈ λ°°κ²½ μ˜μƒμ— μ§‘μ€‘ν•˜μ„Έμš”.
3. λ‹€μŒ ꡬ쑰λ₯Ό λ°˜λ“œμ‹œ ν¬ν•¨ν•˜μ„Έμš”:
- μ£Όμš” λ™μž‘μ„ λͺ…ν™•ν•œ ν•œ λ¬Έμž₯으둜 μ‹œμž‘
- ꡬ체적인 λ™μž‘κ³Ό 제슀처λ₯Ό μ‹œκ°„ μˆœμ„œλŒ€λ‘œ μ„€λͺ…
- λ°°κ²½κ³Ό ν™˜κ²½ μ„ΈλΆ€ 사항을 ꡬ체적으둜 포함
- 카메라 각도와 μ›€μ§μž„μ„ λͺ…μ‹œ
- μ‘°λͺ…κ³Ό 색상을 μžμ„Ένžˆ μ„€λͺ…
- λ³€ν™”λ‚˜ κ°‘μž‘μŠ€λŸ¬μš΄ 사건을 μžμ—°μŠ€λŸ½κ²Œ 포함"""}
]
try:
response = client.chat.completions.create(
model="gpt-4-1106-preview",
messages=messages,
max_tokens=1000, # 토큰 수 증가
temperature=0.7
)
generated_prompt = response.choices[0].message.content.strip()
return f"{section_number}. {generated_prompt}"
except Exception as e:
print(f"Error during prompt generation for section {section_number}: {e}")
return f"Error occurred during prompt generation for section {section_number}"
# λΉ„λ””μ˜€ κ²°ν•© ν•¨μˆ˜ μΆ”κ°€
def combine_videos(video_paths, output_path):
"""μ—¬λŸ¬ λΉ„λ””μ˜€λ₯Ό ν•˜λ‚˜λ‘œ κ²°ν•©"""
if not all(video_paths):
raise gr.Error("λͺ¨λ“  μ„Ήμ…˜μ˜ μ˜μƒμ΄ μƒμ„±λ˜μ–΄μ•Ό ν•©λ‹ˆλ‹€.")
try:
# 첫 번째 λΉ„λ””μ˜€μ˜ 속성 κ°€μ Έμ˜€κΈ°
cap = cv2.VideoCapture(video_paths[0])
fps = int(cap.get(cv2.CAP_PROP_FPS))
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
cap.release()
# 좜λ ₯ λΉ„λ””μ˜€ μ„€μ •
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
out = cv2.VideoWriter(output_path, fourcc, fps, (width, height))
# 각 λΉ„λ””μ˜€ 순차적으둜 κ²°ν•©
for video_path in video_paths:
if video_path and os.path.exists(video_path):
cap = cv2.VideoCapture(video_path)
while True:
ret, frame = cap.read()
if not ret:
break
out.write(frame)
cap.release()
out.release()
return output_path
except Exception as e:
raise gr.Error(f"λΉ„λ””μ˜€ κ²°ν•© 쀑 였λ₯˜ λ°œμƒ: {e}")
def merge_section_videos(section1, section2, section3, section4, section5):
"""μ„Ήμ…˜ λΉ„λ””μ˜€λ“€μ„ ν•˜λ‚˜λ‘œ κ²°ν•©"""
videos = []
# 각 μ„Ήμ…˜ λΉ„λ””μ˜€ 확인 및 처리
for i, video_path in enumerate([section1, section2, section3, section4, section5], 1):
if video_path:
if os.path.exists(video_path):
try:
# λΉ„λ””μ˜€ 파일 검증
cap = cv2.VideoCapture(video_path)
if cap.isOpened():
videos.append(video_path)
cap.release()
else:
raise gr.Error(f"μ„Ήμ…˜ {i}의 μ˜μƒ 파일이 μ†μƒλ˜μ—ˆκ±°λ‚˜ 읽을 수 μ—†μŠ΅λ‹ˆλ‹€.")
except Exception as e:
raise gr.Error(f"μ„Ήμ…˜ {i} μ˜μƒ 처리 쀑 였λ₯˜: {str(e)}")
else:
raise gr.Error(f"μ„Ήμ…˜ {i}의 μ˜μƒ νŒŒμΌμ„ 찾을 수 μ—†μŠ΅λ‹ˆλ‹€.")
else:
raise gr.Error(f"μ„Ήμ…˜ {i}의 μ˜μƒμ΄ μ—†μŠ΅λ‹ˆλ‹€.")
if not videos:
raise gr.Error("κ²°ν•©ν•  μ˜μƒμ΄ μ—†μŠ΅λ‹ˆλ‹€.")
try:
output_path = tempfile.mktemp(suffix=".mp4")
# 첫 번째 λΉ„λ””μ˜€μ˜ 속성 κ°€μ Έμ˜€κΈ°
cap = cv2.VideoCapture(videos[0])
fps = int(cap.get(cv2.CAP_PROP_FPS))
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
cap.release()
# 좜λ ₯ λΉ„λ””μ˜€ μ„€μ •
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
out = cv2.VideoWriter(output_path, fourcc, fps, (width, height))
# 각 λΉ„λ””μ˜€ 순차적으둜 κ²°ν•©
for video_path in videos:
cap = cv2.VideoCapture(video_path)
while True:
ret, frame = cap.read()
if not ret:
break
# ν”„λ ˆμž„ 크기가 λ‹€λ₯Έ 경우 λ¦¬μ‚¬μ΄μ¦ˆ
if frame.shape[:2] != (height, width):
frame = cv2.resize(frame, (width, height))
out.write(frame)
cap.release()
out.release()
print(f"Successfully merged {len(videos)} videos")
return output_path
except Exception as e:
raise gr.Error(f"λΉ„λ””μ˜€ κ²°ν•© 쀑 였λ₯˜ λ°œμƒ: {e}")
def generate_script(topic):
"""μ£Όμ œμ— λ§žλŠ” 슀크립트 생성"""
if not topic:
return "주제λ₯Ό μž…λ ₯ν•΄μ£Όμ„Έμš”."
messages = [
{"role": "system", "content": """당신은 μ˜μƒ 슀크립트 μž‘μ„± μ „λ¬Έκ°€μž…λ‹ˆλ‹€.
주어진 주제둜 λ‹€μŒ ꡬ쑰에 λ§žλŠ” 5개 μ„Ήμ…˜μ˜ 슀크립트λ₯Ό μž‘μ„±ν•΄μ£Όμ„Έμš”:
1. λ°°κ²½ 및 ν•„μš”μ„±: 주제 μ†Œκ°œμ™€ μ‹œμ²­μžμ˜ ν₯λ―Έ 유발
2. ν₯λ―Έ 유발: ꡬ체적인 λ‚΄μš© μ „κ°œμ™€ ν˜ΈκΈ°μ‹¬ 자극
3. ν•΄κ²°μ±… μ œμ‹œ: 핡심 λ‚΄μš©κ³Ό ν•΄κ²°λ°©μ•ˆ μ œμ‹œ
4. λ³Έλ‘ : μƒμ„Έν•œ μ„€λͺ…κ³Ό μž₯점 뢀각
5. κ²°λ‘ : 핡심 λ©”μ‹œμ§€ 강쑰와 행동 μœ λ„
각 μ„Ήμ…˜μ€ μžμ—°μŠ€λŸ½κ²Œ μ—°κ²°λ˜μ–΄μ•Ό ν•˜λ©°,
μ „μ²΄μ μœΌλ‘œ μΌκ΄€λœ 톀과 λΆ„μœ„κΈ°λ₯Ό μœ μ§€ν•˜λ©΄μ„œλ„
μ‹œμ²­μžμ˜ 관심을 λκΉŒμ§€ μœ μ§€ν•  수 μžˆλ„λ‘ μž‘μ„±ν•΄μ£Όμ„Έμš”."""},
{"role": "user", "content": f"λ‹€μŒ 주제둜 μ˜μƒ 슀크립트λ₯Ό μž‘μ„±ν•΄μ£Όμ„Έμš”: {topic}"}
]
try:
response = client.chat.completions.create(
model="gpt-4-1106-preview",
messages=messages,
max_tokens=2000,
temperature=0.7
)
return response.choices[0].message.content.strip()
except Exception as e:
print(f"Error during script generation: {e}")
return "슀크립트 생성 쀑 였λ₯˜κ°€ λ°œμƒν–ˆμŠ΅λ‹ˆλ‹€."
def cleanup():
"""λ©”λͺ¨λ¦¬ 정리 ν•¨μˆ˜"""
torch.cuda.empty_cache()
gc.collect()
with gr.Blocks(theme="Yntec/HaleyCH_Theme_Orange") as iface:
# State λ³€μˆ˜λ“€μ˜ μ΄ˆκΈ°ν™”
txt2vid_current_height = gr.State(value=320)
txt2vid_current_width = gr.State(value=512)
txt2vid_current_num_frames = gr.State(value=257)
img2vid_current_height = gr.State(value=320)
img2vid_current_width = gr.State(value=512)
img2vid_current_num_frames = gr.State(value=257)
with gr.Tabs():
# Text to Video Tab
with gr.TabItem("ν…μŠ€νŠΈλ‘œ λΉ„λ””μ˜€ λ§Œλ“€κΈ°"):
with gr.Row():
with gr.Column():
txt2vid_prompt = gr.Textbox(
label="Step 1: ν”„λ‘¬ν”„νŠΈ μž…λ ₯",
placeholder="μƒμ„±ν•˜κ³  싢은 λΉ„λ””μ˜€λ₯Ό μ„€λͺ…ν•˜μ„Έμš” (μ΅œμ†Œ 50자)...",
value="κ·€μ—¬μš΄ 고양이",
lines=5,
)
txt2vid_enhance_toggle = Toggle(
label="ν”„λ‘¬ν”„νŠΈ 증강",
value=False,
interactive=True,
)
txt2vid_negative_prompt = gr.Textbox(
label="Step 2: λ„€κ±°ν‹°λΈŒ ν”„λ‘¬ν”„νŠΈ μž…λ ₯",
placeholder="λΉ„λ””μ˜€μ—μ„œ μ›ν•˜μ§€ μ•ŠλŠ” μš”μ†Œλ₯Ό μ„€λͺ…ν•˜μ„Έμš”...",
value="low quality, worst quality, deformed, distorted, warped, motion smear, motion artifacts, fused fingers, incorrect anatomy, strange hands, unattractive",
lines=2,
visible=False
)
txt2vid_preset = gr.Dropdown(
choices=[p["label"] for p in preset_options],
value="[16:9] 512x320, 10.3초",
label="Step 2: 해상도 프리셋 선택",
)
txt2vid_frame_rate = gr.Slider(
label="Step 3: ν”„λ ˆμž„ 레이트",
minimum=21,
maximum=30,
step=1,
value=25,
visible=False
)
txt2vid_advanced = create_advanced_options()
txt2vid_generate = gr.Button(
"Step 3: λΉ„λ””μ˜€ 생성",
variant="primary",
size="lg",
)
with gr.Column():
txt2vid_output = gr.Video(label="μƒμ„±λœ λΉ„λ””μ˜€")
# Image to Video Tab
with gr.TabItem("μ΄λ―Έμ§€λ‘œ λΉ„λ””μ˜€ λ§Œλ“€κΈ°"):
with gr.Row():
with gr.Column():
img2vid_image = gr.Image(
type="filepath",
label="Step 1: μž…λ ₯ 이미지 μ—…λ‘œλ“œ",
elem_id="image_upload",
)
img2vid_prompt = gr.Textbox(
label="Step 2: ν”„λ‘¬ν”„νŠΈ μž…λ ₯",
placeholder="이미지λ₯Ό μ–΄λ–»κ²Œ μ• λ‹ˆλ©”μ΄μ…˜ν™”ν• μ§€ μ„€λͺ…ν•˜μ„Έμš” (μ΅œμ†Œ 50자)...",
value="κ·€μ—¬μš΄ 고양이",
lines=5,
)
img2vid_enhance_toggle = Toggle(
label="ν”„λ‘¬ν”„νŠΈ 증강",
value=False,
interactive=True,
)
img2vid_negative_prompt = gr.Textbox(
label="Step 3: λ„€κ±°ν‹°λΈŒ ν”„λ‘¬ν”„νŠΈ μž…λ ₯",
placeholder="λΉ„λ””μ˜€μ—μ„œ μ›ν•˜μ§€ μ•ŠλŠ” μš”μ†Œλ₯Ό μ„€λͺ…ν•˜μ„Έμš”...",
value="low quality, worst quality, deformed, distorted, warped, motion smear, motion artifacts, fused fingers, incorrect anatomy, strange hands, unattractive",
lines=2,
visible=False
)
img2vid_preset = gr.Dropdown(
choices=[p["label"] for p in preset_options],
value="[16:9] 512x320, 10.3초",
label="Step 3: 해상도 프리셋 선택",
)
img2vid_frame_rate = gr.Slider(
label="Step 4: ν”„λ ˆμž„ 레이트",
minimum=21,
maximum=30,
step=1,
value=25,
visible=False
)
img2vid_advanced = create_advanced_options()
img2vid_generate = gr.Button(
"Step 4: λΉ„λ””μ˜€ 생성",
variant="primary",
size="lg",
)
with gr.Column():
img2vid_output = gr.Video(label="μƒμ„±λœ λΉ„λ””μ˜€")
# Scenario Tab
with gr.TabItem("μ‹œλ‚˜λ¦¬μ˜€λ‘œ λΉ„λ””μ˜€ λ§Œλ“€κΈ°(숏폼)"):
with gr.Row():
with gr.Column(scale=1):
script_topic = gr.Textbox(
label="슀크립트 생성",
placeholder="겨울 일본 온천 여행을 주제둜 밝은 λŠλ‚ŒμœΌλ‘œ 슀크립트 μƒμ„±ν•˜λΌ",
lines=2
)
generate_script_btn = gr.Button("슀크립트 생성", variant="primary")
scenario_input = gr.Textbox(
label="μ˜μƒ 슀크립트 μž…λ ₯",
placeholder="전체 μ‹œλ‚˜λ¦¬μ˜€λ₯Ό μž…λ ₯ν•˜μ„Έμš”...",
lines=10
)
scenario_preset = gr.Dropdown(
choices=[p["label"] for p in preset_options],
value="[16:9] 512x320, 10.3초",
label="ν™”λ©΄ 크기 선택"
)
analyze_btn = gr.Button("μ‹œλ‚˜λ¦¬μ˜€ 뢄석 및 ν”„λ‘¬ν”„νŠΈ 생성", variant="primary")
with gr.Column(scale=2):
with gr.Row():
# μ„Ήμ…˜ 1
with gr.Column():
section1_prompt = gr.Textbox(
label="1. λ°°κ²½ 및 ν•„μš”μ„±",
lines=4
)
with gr.Row():
section1_regenerate = gr.Button("πŸ”„ ν”„λ‘¬ν”„νŠΈ 생성")
section1_generate = gr.Button("πŸ”„ μ˜μƒ 생성")
section1_video = gr.Video(label="μ„Ήμ…˜ 1 μ˜μƒ")
# μ„Ήμ…˜ 2
with gr.Column():
section2_prompt = gr.Textbox(
label="2. ν₯λ―Έ 유발",
lines=4
)
with gr.Row():
section2_regenerate = gr.Button("πŸ”„ ν”„λ‘¬ν”„νŠΈ 생성")
section2_generate = gr.Button("πŸ”„ μ˜μƒ 생성")
section2_video = gr.Video(label="μ„Ήμ…˜ 2 μ˜μƒ")
with gr.Row():
# μ„Ήμ…˜ 3
with gr.Column():
section3_prompt = gr.Textbox(
label="3. ν•΄κ²°μ±… μ œμ‹œ",
lines=4
)
with gr.Row():
section3_regenerate = gr.Button("πŸ”„ ν”„λ‘¬ν”„νŠΈ 생성")
section3_generate = gr.Button("πŸ”„ μ˜μƒ 생성")
section3_video = gr.Video(label="μ„Ήμ…˜ 3 μ˜μƒ")
# μ„Ήμ…˜ 4
with gr.Column():
section4_prompt = gr.Textbox(
label="4. λ³Έλ‘ ",
lines=4
)
with gr.Row():
section4_regenerate = gr.Button("πŸ”„ ν”„λ‘¬ν”„νŠΈ 생성")
section4_generate = gr.Button("πŸ”„ μ˜μƒ 생성")
section4_video = gr.Video(label="μ„Ήμ…˜ 4 μ˜μƒ")
with gr.Row():
# μ„Ήμ…˜ 5
with gr.Column():
section5_prompt = gr.Textbox(
label="5. κ²°λ‘  및 κ°•μ‘°",
lines=4
)
with gr.Row():
section5_regenerate = gr.Button("πŸ”„ ν”„λ‘¬ν”„νŠΈ 생성")
section5_generate = gr.Button("πŸ”„ μ˜μƒ 생성")
section5_video = gr.Video(label="μ„Ήμ…˜ 5 μ˜μƒ")
# 톡합 μ˜μƒ μ„Ήμ…˜
with gr.Row():
with gr.Column(scale=1):
merge_videos_btn = gr.Button("톡합 μ˜μƒ 생성", variant="primary", size="lg")
with gr.Column(scale=2):
with gr.Row():
merged_video_output = gr.Video(label="톡합 μ˜μƒ")
# Text to Video Tab handlers
txt2vid_preset.change(
fn=preset_changed,
inputs=[txt2vid_preset],
outputs=[
txt2vid_current_height,
txt2vid_current_width,
txt2vid_current_num_frames,
txt2vid_advanced[3], # height_slider
txt2vid_advanced[4], # width_slider
txt2vid_advanced[5], # num_frames_slider
]
)
txt2vid_enhance_toggle.change(
fn=update_prompt_t2v,
inputs=[txt2vid_prompt, txt2vid_enhance_toggle],
outputs=txt2vid_prompt
)
txt2vid_generate.click(
fn=generate_video_from_text,
inputs=[
txt2vid_prompt,
txt2vid_enhance_toggle,
txt2vid_negative_prompt,
txt2vid_frame_rate,
txt2vid_advanced[0], # seed
txt2vid_advanced[1], # inference_steps
txt2vid_advanced[2], # guidance_scale
txt2vid_current_height,
txt2vid_current_width,
txt2vid_current_num_frames,
],
outputs=txt2vid_output,
)
# Image to Video Tab handlers
img2vid_preset.change(
fn=preset_changed,
inputs=[img2vid_preset],
outputs=[
img2vid_current_height,
img2vid_current_width,
img2vid_current_num_frames,
img2vid_advanced[3], # height_slider
img2vid_advanced[4], # width_slider
img2vid_advanced[5], # num_frames_slider
]
)
img2vid_enhance_toggle.change(
fn=update_prompt_i2v,
inputs=[img2vid_prompt, img2vid_enhance_toggle],
outputs=img2vid_prompt
)
img2vid_generate.click(
fn=generate_video_from_image,
inputs=[
img2vid_image,
img2vid_prompt,
img2vid_enhance_toggle,
img2vid_negative_prompt,
img2vid_frame_rate,
img2vid_advanced[0], # seed
img2vid_advanced[1], # inference_steps
img2vid_advanced[2], # guidance_scale
img2vid_current_height,
img2vid_current_width,
img2vid_current_num_frames,
],
outputs=img2vid_output,
)
# Scenario Tab handlers
generate_script_btn.click(
fn=generate_script,
inputs=[script_topic],
outputs=[scenario_input]
)
analyze_btn.click(
fn=analyze_scenario,
inputs=[scenario_input],
outputs=[
section1_prompt, section2_prompt, section3_prompt,
section4_prompt, section5_prompt
]
)
# μ„Ήμ…˜λ³„ ν”„λ‘¬ν”„νŠΈ μž¬μƒμ„± ν•Έλ“€λŸ¬
section1_regenerate.click(
fn=lambda x: generate_single_section_prompt(x, 1),
inputs=[scenario_input],
outputs=section1_prompt
)
section2_regenerate.click(
fn=lambda x: generate_single_section_prompt(x, 2),
inputs=[scenario_input],
outputs=section2_prompt
)
section3_regenerate.click(
fn=lambda x: generate_single_section_prompt(x, 3),
inputs=[scenario_input],
outputs=section3_prompt
)
section4_regenerate.click(
fn=lambda x: generate_single_section_prompt(x, 4),
inputs=[scenario_input],
outputs=section4_prompt
)
section5_regenerate.click(
fn=lambda x: generate_single_section_prompt(x, 5),
inputs=[scenario_input],
outputs=section5_prompt
)
# μ„Ήμ…˜λ³„ λΉ„λ””μ˜€ 생성 ν•Έλ“€λŸ¬
section1_generate.click(
fn=lambda p, pr: generate_section_video(p, pr, 1),
inputs=[section1_prompt, scenario_preset],
outputs=section1_video
)
section2_generate.click(
fn=lambda p, pr: generate_section_video(p, pr, 2),
inputs=[section2_prompt, scenario_preset],
outputs=section2_video
)
section3_generate.click(
fn=lambda p, pr: generate_section_video(p, pr, 3),
inputs=[section3_prompt, scenario_preset],
outputs=section3_video
)
section4_generate.click(
fn=lambda p, pr: generate_section_video(p, pr, 4),
inputs=[section4_prompt, scenario_preset],
outputs=section4_video
)
section5_generate.click(
fn=lambda p, pr: generate_section_video(p, pr, 5),
inputs=[section5_prompt, scenario_preset],
outputs=section5_video
)
# 톡합 μ˜μƒ 생성 ν•Έλ“€λŸ¬
merge_videos_btn.click(
fn=merge_section_videos,
inputs=[
section1_video,
section2_video,
section3_video,
section4_video,
section5_video
],
outputs=merged_video_output
)
if __name__ == "__main__":
iface.queue(max_size=64, default_concurrency_limit=1, api_open=False).launch(
share=True,
show_api=False
)