ginigen-sora / app-backup1.py
openfree's picture
Create app-backup1.py
44f4dde verified
raw
history blame
30.7 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
# Load system prompts
system_prompt_t2v = """๋‹น์‹ ์€ ๋น„๋””์˜ค ์ƒ์„ฑ์„ ์œ„ํ•œ ํ”„๋กฌํ”„ํŠธ ์ „๋ฌธ๊ฐ€์ž…๋‹ˆ๋‹ค.
์ฃผ์–ด์ง„ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋‹ค์Œ ๊ตฌ์กฐ์— ๋งž๊ฒŒ ๊ฐœ์„ ํ•ด์ฃผ์„ธ์š”:
1. ์ฃผ์š” ๋™์ž‘์„ ๋ช…ํ™•ํ•œ ํ•œ ๋ฌธ์žฅ์œผ๋กœ ์‹œ์ž‘
2. ๊ตฌ์ฒด์ ์ธ ๋™์ž‘๊ณผ ์ œ์Šค์ฒ˜๋ฅผ ์‹œ๊ฐ„ ์ˆœ์„œ๋Œ€๋กœ ์„ค๋ช…
3. ์บ๋ฆญํ„ฐ/๊ฐ์ฒด์˜ ์™ธ๋ชจ๋ฅผ ์ƒ์„ธํžˆ ๋ฌ˜์‚ฌ
4. ๋ฐฐ๊ฒฝ๊ณผ ํ™˜๊ฒฝ ์„ธ๋ถ€ ์‚ฌํ•ญ์„ ๊ตฌ์ฒด์ ์œผ๋กœ ํฌํ•จ
5. ์นด๋ฉ”๋ผ ๊ฐ๋„์™€ ์›€์ง์ž„์„ ๋ช…์‹œ
6. ์กฐ๋ช…๊ณผ ์ƒ‰์ƒ์„ ์ž์„ธํžˆ ์„ค๋ช…
7. ๋ณ€ํ™”๋‚˜ ๊ฐ‘์ž‘์Šค๋Ÿฌ์šด ์‚ฌ๊ฑด์„ ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ํฌํ•จ
๋ชจ๋“  ์„ค๋ช…์€ ํ•˜๋‚˜์˜ ์ž์—ฐ์Šค๋Ÿฌ์šด ๋ฌธ๋‹จ์œผ๋กœ ์ž‘์„ฑํ•˜๊ณ ,
์ดฌ์˜ ๊ฐ๋…์ด ์ดฌ์˜ ๋ชฉ๋ก์„ ์„ค๋ช…ํ•˜๋Š” ๊ฒƒ์ฒ˜๋Ÿผ ๊ตฌ์ฒด์ ์ด๊ณ  ์‹œ๊ฐ์ ์œผ๋กœ ์ž‘์„ฑํ•˜์„ธ์š”.
200๋‹จ์–ด๋ฅผ ๋„˜์ง€ ์•Š๋„๋ก ํ•˜๋˜, ์ตœ๋Œ€ํ•œ ์ƒ์„ธํ•˜๊ฒŒ ์ž‘์„ฑํ•˜์„ธ์š”."""
system_prompt_i2v = """๋‹น์‹ ์€ ์ด๋ฏธ์ง€ ๊ธฐ๋ฐ˜ ๋น„๋””์˜ค ์ƒ์„ฑ์„ ์œ„ํ•œ ํ”„๋กฌํ”„ํŠธ ์ „๋ฌธ๊ฐ€์ž…๋‹ˆ๋‹ค.
์ฃผ์–ด์ง„ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋‹ค์Œ ๊ตฌ์กฐ์— ๋งž๊ฒŒ ๊ฐœ์„ ํ•ด์ฃผ์„ธ์š”:
1. ์ฃผ์š” ๋™์ž‘์„ ๋ช…ํ™•ํ•œ ํ•œ ๋ฌธ์žฅ์œผ๋กœ ์‹œ์ž‘
2. ๊ตฌ์ฒด์ ์ธ ๋™์ž‘๊ณผ ์ œ์Šค์ฒ˜๋ฅผ ์‹œ๊ฐ„ ์ˆœ์„œ๋Œ€๋กœ ์„ค๋ช…
3. ์บ๋ฆญํ„ฐ/๊ฐ์ฒด์˜ ์™ธ๋ชจ๋ฅผ ์ƒ์„ธํžˆ ๋ฌ˜์‚ฌ
4. ๋ฐฐ๊ฒฝ๊ณผ ํ™˜๊ฒฝ ์„ธ๋ถ€ ์‚ฌํ•ญ์„ ๊ตฌ์ฒด์ ์œผ๋กœ ํฌํ•จ
5. ์นด๋ฉ”๋ผ ๊ฐ๋„์™€ ์›€์ง์ž„์„ ๋ช…์‹œ
6. ์กฐ๋ช…๊ณผ ์ƒ‰์ƒ์„ ์ž์„ธํžˆ ์„ค๋ช…
7. ๋ณ€ํ™”๋‚˜ ๊ฐ‘์ž‘์Šค๋Ÿฌ์šด ์‚ฌ๊ฑด์„ ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ํฌํ•จ
๋ชจ๋“  ์„ค๋ช…์€ ํ•˜๋‚˜์˜ ์ž์—ฐ์Šค๋Ÿฌ์šด ๋ฌธ๋‹จ์œผ๋กœ ์ž‘์„ฑํ•˜๊ณ ,
์ดฌ์˜ ๊ฐ๋…์ด ์ดฌ์˜ ๋ชฉ๋ก์„ ์„ค๋ช…ํ•˜๋Š” ๊ฒƒ์ฒ˜๋Ÿผ ๊ตฌ์ฒด์ ์ด๊ณ  ์‹œ๊ฐ์ ์œผ๋กœ ์ž‘์„ฑํ•˜์„ธ์š”.
200๋‹จ์–ด๋ฅผ ๋„˜์ง€ ์•Š๋„๋ก ํ•˜๋˜, ์ตœ๋Œ€ํ•œ ์ƒ์„ธํ•˜๊ฒŒ ์ž‘์„ฑํ•˜์„ธ์š”."""
# 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
translator = pipeline("translation", model="Helsinki-NLP/opus-mt-ko-en")
# Korean text detection function
def contains_korean(text):
korean_pattern = re.compile('[ใ„ฑ-ใ…Žใ…-ใ…ฃ๊ฐ€-ํžฃ]')
return bool(korean_pattern.search(text))
def translate_korean_prompt(prompt):
"""
Translate Korean prompt to English if Korean text is detected
"""
if contains_korean(prompt):
translated = translator(prompt)[0]['translation_text']
print(f"Original Korean prompt: {prompt}")
print(f"Translated English prompt: {translated}")
return translated
return prompt
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=== ํ”„๋กฌํ”„ํŠธ ์ฆ๊ฐ• ๊ฒฐ๊ณผ ===")
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
preset_options = [
{"label": "1216x704, 41 frames", "width": 1216, "height": 704, "num_frames": 41},
{"label": "1088x704, 49 frames", "width": 1088, "height": 704, "num_frames": 49},
{"label": "1056x640, 57 frames", "width": 1056, "height": 640, "num_frames": 57},
{"label": "992x608, 65 frames", "width": 992, "height": 608, "num_frames": 65},
{"label": "896x608, 73 frames", "width": 896, "height": 608, "num_frames": 73},
{"label": "896x544, 81 frames", "width": 896, "height": 544, "num_frames": 81},
{"label": "832x544, 89 frames", "width": 832, "height": 544, "num_frames": 89},
{"label": "800x512, 97 frames", "width": 800, "height": 512, "num_frames": 97},
{"label": "768x512, 97 frames", "width": 768, "height": 512, "num_frames": 97},
{"label": "800x480, 105 frames", "width": 800, "height": 480, "num_frames": 105},
{"label": "736x480, 113 frames", "width": 736, "height": 480, "num_frames": 113},
{"label": "704x480, 121 frames", "width": 704, "height": 480, "num_frames": 121},
{"label": "704x448, 129 frames", "width": 704, "height": 448, "num_frames": 129},
{"label": "672x448, 137 frames", "width": 672, "height": 448, "num_frames": 137},
{"label": "640x416, 153 frames", "width": 640, "height": 416, "num_frames": 153},
{"label": "672x384, 161 frames", "width": 672, "height": 384, "num_frames": 161},
{"label": "640x384, 169 frames", "width": 640, "height": 384, "num_frames": 169},
{"label": "608x384, 177 frames", "width": 608, "height": 384, "num_frames": 177},
{"label": "576x384, 185 frames", "width": 576, "height": 384, "num_frames": 185},
{"label": "608x352, 193 frames", "width": 608, "height": 352, "num_frames": 193},
{"label": "576x352, 201 frames", "width": 576, "height": 352, "num_frames": 201},
{"label": "544x352, 209 frames", "width": 544, "height": 352, "num_frames": 209},
{"label": "512x352, 225 frames", "width": 512, "height": 352, "num_frames": 225},
{"label": "512x352, 233 frames", "width": 512, "height": 352, "num_frames": 233},
{"label": "544x320, 241 frames", "width": 544, "height": 320, "num_frames": 241},
{"label": "512x320, 249 frames", "width": 512, "height": 320, "num_frames": 249},
{"label": "512x320, 257 frames", "width": 512, "height": 320, "num_frames": 257},
]
def preset_changed(preset):
if preset != "Custom":
selected = next(item for item in preset_options if item["label"] == preset)
return (
selected["height"],
selected["width"],
selected["num_frames"],
gr.update(visible=False),
gr.update(visible=False),
gr.update(visible=False),
)
else:
return (
None,
None,
None,
gr.update(visible=True),
gr.update(visible=True),
gr.update(visible=True),
)
def generate_video_from_text(
prompt="",
enhance_prompt_toggle=False,
negative_prompt="",
frame_rate=25,
seed=171198,
num_inference_steps=30,
guidance_scale=3,
height=512,
width=768,
num_frames=121,
progress=gr.Progress(),
):
if len(prompt.strip()) < 50:
raise gr.Error(
"ํ”„๋กฌํ”„ํŠธ๋Š” ์ตœ์†Œ 50์ž ์ด์ƒ์ด์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋” ์ž์„ธํ•œ ์„ค๋ช…์„ ์ œ๊ณตํ•ด์ฃผ์„ธ์š”.",
duration=5,
)
# Translate Korean prompts to English
prompt = translate_korean_prompt(prompt)
negative_prompt = translate_korean_prompt(negative_prompt)
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"๋น„๋””์˜ค ์ƒ์„ฑ ์ค‘ ์˜ค๋ฅ˜๊ฐ€ ๋ฐœ์ƒํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์‹œ ์‹œ๋„ํ•ด์ฃผ์„ธ์š”. ์˜ค๋ฅ˜: {e}",
duration=5,
)
finally:
torch.cuda.empty_cache()
gc.collect()
output_path = tempfile.mktemp(suffix=".mp4")
print(images.shape)
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=False,
negative_prompt="",
frame_rate=25,
seed=171198,
num_inference_steps=30,
guidance_scale=3,
height=512,
width=768,
num_frames=121,
progress=gr.Progress(),
):
print("Height: ", height)
print("Width: ", width)
print("Num Frames: ", num_frames)
if len(prompt.strip()) < 50:
raise gr.Error(
"ํ”„๋กฌํ”„ํŠธ๋Š” ์ตœ์†Œ 50์ž ์ด์ƒ์ด์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋” ์ž์„ธํ•œ ์„ค๋ช…์„ ์ œ๊ณตํ•ด์ฃผ์„ธ์š”.",
duration=5,
)
if not image_path:
raise gr.Error("์ž…๋ ฅ ์ด๋ฏธ์ง€๋ฅผ ์ œ๊ณตํ•ด์ฃผ์„ธ์š”.", duration=5)
# Translate Korean prompts to English
prompt = translate_korean_prompt(prompt)
negative_prompt = translate_korean_prompt(negative_prompt)
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"๋น„๋””์˜ค ์ƒ์„ฑ ์ค‘ ์˜ค๋ฅ˜๊ฐ€ ๋ฐœ์ƒํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์‹œ ์‹œ๋„ํ•ด์ฃผ์„ธ์š”. ์˜ค๋ฅ˜: {e}",
duration=5,
)
finally:
torch.cuda.empty_cache()
gc.collect()
return output_path
def create_advanced_options():
with gr.Accordion("Step 4: Advanced Options (Optional)", open=False):
seed = gr.Slider(
label="4.1 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=30
)
guidance_scale = gr.Slider(
label="4.3 Guidance Scale", minimum=1.0, maximum=5.0, step=0.1, value=3.0
)
height_slider = gr.Slider(
label="4.4 Height",
minimum=256,
maximum=1024,
step=64,
value=512,
visible=False,
)
width_slider = gr.Slider(
label="4.5 Width",
minimum=256,
maximum=1024,
step=64,
value=768,
visible=False,
)
num_frames_slider = gr.Slider(
label="4.5 Number of Frames",
minimum=1,
maximum=200,
step=1,
value=121,
visible=False,
)
return [
seed,
inference_steps,
guidance_scale,
height_slider,
width_slider,
num_frames_slider,
]
# Gradio Interface Definition
with gr.Blocks(theme=gr.themes.Soft()) as iface:
with gr.Tabs():
# Text to Video Tab
with gr.TabItem("ํ…์ŠคํŠธ๋กœ ๋น„๋””์˜ค ๋งŒ๋“ค๊ธฐ"):
with gr.Row():
with gr.Column():
txt2vid_prompt = gr.Textbox(
label="Step 1: ํ”„๋กฌํ”„ํŠธ ์ž…๋ ฅ",
placeholder="์ƒ์„ฑํ•˜๊ณ  ์‹ถ์€ ๋น„๋””์˜ค๋ฅผ ์„ค๋ช…ํ•˜์„ธ์š” (์ตœ์†Œ 50์ž)...",
value="๊ฐˆ์ƒ‰ ๊ธด ๋จธ๋ฆฌ๋ฅผ ๊ฐ€์ง„ ์—ฌ์„ฑ์ด ๊ธˆ๋ฐœ์˜ ๊ธด ๋จธ๋ฆฌ๋ฅผ ๊ฐ€์ง„ ๋‹ค๋ฅธ ์—ฌ์„ฑ์„ ํ–ฅํ•ด ๋ฏธ์†Œ์ง“์Šต๋‹ˆ๋‹ค. ๊ฐˆ์ƒ‰ ๋จธ๋ฆฌ์˜ ์—ฌ์„ฑ์€ ๊ฒ€์€์ƒ‰ ์ž์ผ“์„ ์ž…๊ณ  ์žˆ์œผ๋ฉฐ ์˜ค๋ฅธ์ชฝ ๋บจ์— ์ž‘์€ ์ ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์นด๋ฉ”๋ผ ๊ฐ๋„๋Š” ๊ฐˆ์ƒ‰ ๋จธ๋ฆฌ ์—ฌ์„ฑ์˜ ์–ผ๊ตด์— ํด๋กœ์ฆˆ์—…๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์กฐ๋ช…์€ ์ž์—ฐ์Šค๋Ÿฝ๊ณ  ๋”ฐ๋œปํ•˜๋ฉฐ, ์„์–‘์—์„œ ์˜ค๋Š” ๋“ฏํ•œ ๋ถ€๋“œ๋Ÿฌ์šด ๋น›์ด ์žฅ๋ฉด์„ ๋น„์ถฅ๋‹ˆ๋‹ค. ์žฅ๋ฉด์€ ์‹ค์ œ ์˜์ƒ์ฒ˜๋Ÿผ ๋ณด์ž…๋‹ˆ๋‹ค.",
lines=5,
)
txt2vid_enhance_toggle = Toggle(
label="ํ”„๋กฌํ”„ํŠธ ๊ฐœ์„ ",
value=False,
interactive=True,
)
txt2vid_negative_prompt = gr.Textbox(
label="Step 2: ๋„ค๊ฑฐํ‹ฐ๋ธŒ ํ”„๋กฌํ”„ํŠธ ์ž…๋ ฅ",
placeholder="๋น„๋””์˜ค์—์„œ ์›ํ•˜์ง€ ์•Š๋Š” ์š”์†Œ๋ฅผ ์„ค๋ช…ํ•˜์„ธ์š”...",
value="low quality, worst quality, deformed, distorted, warped, motion smear, motion artifacts, fused fingers, incorrect anatomy, strange hands, unattractive",
lines=2,
)
txt2vid_preset = gr.Dropdown(
choices=[p["label"] for p in preset_options],
value="768x512, 97 frames",
label="Step 3.1: ํ•ด์ƒ๋„ ํ”„๋ฆฌ์…‹ ์„ ํƒ",
)
txt2vid_frame_rate = gr.Slider(
label="Step 3.2: ํ”„๋ ˆ์ž„ ๋ ˆ์ดํŠธ",
minimum=21,
maximum=30,
step=1,
value=25,
)
txt2vid_advanced = create_advanced_options()
txt2vid_generate = gr.Button(
"Step 5: ๋น„๋””์˜ค ์ƒ์„ฑ",
variant="primary",
size="lg",
)
with gr.Column():
txt2vid_output = gr.Video(label="์ƒ์„ฑ๋œ ๋น„๋””์˜ค")
with gr.Row():
gr.Examples(
examples=[
[
"์ „ํ†ต์ ์ธ ๋ชฝ๊ณจ ๋“œ๋ ˆ์Šค๋ฅผ ์ž…์€ ์ Š์€ ์—ฌ์„ฑ์ด ์–‡์€ ํฐ์ƒ‰ ์ปคํŠผ์„ ํ†ตํ•ด ํ˜ธ๊ธฐ์‹ฌ๊ณผ ๊ธด์žฅ์ด ์„ž์ธ ํ‘œ์ •์œผ๋กœ ๋“ค์—ฌ๋‹ค๋ณด๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์—ฌ์„ฑ์€ ํฐ ๊ตฌ์Šฌ๋กœ ์žฅ์‹๋œ ๋‘ ๊ฐœ์˜ ๋•‹์€ ๋จธ๋ฆฌ๋กœ ์Šคํƒ€์ผ๋ง๋œ ๊ธด ๊ฒ€์€ ๋จธ๋ฆฌ๋ฅผ ํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ๋ˆˆ์€ ๋†€๋žŒ์„ ๋„๋ฉฐ ํฌ๊ฒŒ ๋– ์ ธ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€์˜ ๋“œ๋ ˆ์Šค๋Š” ํ™”๋ คํ•œ ๊ธˆ์ƒ‰ ์ž์ˆ˜๊ฐ€ ์ƒˆ๊ฒจ์ง„ ์„ ๋ช…ํ•œ ํŒŒ๋ž€์ƒ‰์ด๋ฉฐ, ๋น„์Šทํ•œ ๋””์ž์ธ์˜ ๋จธ๋ฆฌ๋ ๋ฅผ ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐฐ๊ฒฝ์€ ์‹ ๋น„๋กœ์›€๊ณผ ํ˜ธ๊ธฐ์‹ฌ์„ ์ž์•„๋‚ด๋Š” ๋‹จ์ˆœํ•œ ํฐ์ƒ‰ ์ปคํŠผ์ž…๋‹ˆ๋‹ค.",
"low quality, worst quality, deformed, distorted, warped, motion smear, motion artifacts, fused fingers, incorrect anatomy, strange hands, unattractive",
"assets/t2v_2.mp4",
],
[
"๋…ธ๋ž€์ƒ‰ ์žฌํ‚ท์„ ์ž…์€ ๊ธˆ๋ฐœ ๋จธ๋ฆฌ์˜ ์ Š์€ ๋‚จ์ž๊ฐ€ ์ˆฒ์— ์„œ์„œ ์ฃผ์œ„๋ฅผ ๋‘˜๋Ÿฌ๋ด…๋‹ˆ๋‹ค. ๊ทธ๋Š” ๋ฐ์€ ํ”ผ๋ถ€๋ฅผ ๊ฐ€์กŒ๊ณ  ๋จธ๋ฆฌ๋Š” ๊ฐ€์šด๋ฐ ๊ฐ€๋ฅด๋งˆ๋กœ ์Šคํƒ€์ผ๋ง๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Š” ์™ผ์ชฝ์„ ๋ณด๊ณ  ๋‚œ ํ›„ ์˜ค๋ฅธ์ชฝ์„ ๋ณด๋ฉฐ, ๊ฐ ๋ฐฉํ–ฅ์„ ์ž ์‹œ ์‘์‹œํ•ฉ๋‹ˆ๋‹ค. ์นด๋ฉ”๋ผ๋Š” ๋‚ฎ์€ ๊ฐ๋„์—์„œ ๋‚จ์ž๋ฅผ ์˜ฌ๋ ค๋‹ค๋ณด๋ฉฐ ๊ณ ์ •๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐฐ๊ฒฝ์€ ์•ฝ๊ฐ„ ํ๋ฆฟํ•˜๋ฉฐ, ๋…น์ƒ‰ ๋‚˜๋ฌด๋“ค๊ณผ ๋‚จ์ž์˜ ๋’ค์—์„œ ๋ฐ๊ฒŒ ๋น„์น˜๋Š” ํƒœ์–‘์ด ๋ณด์ž…๋‹ˆ๋‹ค. ์กฐ๋ช…์€ ์ž์—ฐ์Šค๋Ÿฝ๊ณ  ๋”ฐ๋œปํ•˜๋ฉฐ, ํƒœ์–‘ ๋น›์ด ๋‚จ์ž์˜ ์–ผ๊ตด์„ ๊ฐ€๋กœ์ง€๋ฅด๋Š” ๋ Œ์ฆˆ ํ”Œ๋ ˆ์–ด๋ฅผ ๋งŒ๋“ญ๋‹ˆ๋‹ค. ์žฅ๋ฉด์€ ์‹ค์ œ ์˜์ƒ์ฒ˜๋Ÿผ ์ดฌ์˜๋˜์—ˆ์Šต๋‹ˆ๋‹ค.",
"low quality, worst quality, deformed, distorted, warped, motion smear, motion artifacts, fused fingers, incorrect anatomy, strange hands, unattractive",
"assets/t2v_1.mp4",
],
[
"ํ•œ ์‚ฌ์ดํด๋ฆฌ์ŠคํŠธ๊ฐ€ ๊ตฝ์ด์ง„ ์‚ฐ๊ธธ์„ ๋”ฐ๋ผ ๋‹ฌ๋ฆฝ๋‹ˆ๋‹ค. ๊ณต๊ธฐ์—ญํ•™์ ์ธ ์žฅ๋น„๋ฅผ ์ž…์€ ๊ทธ๋Š” ๊ฐ•ํ•˜๊ฒŒ ํŽ˜๋‹ฌ์„ ๋ฐŸ๊ณ  ์žˆ์œผ๋ฉฐ, ์ด๋งˆ์—๋Š” ๋•€๋ฐฉ์šธ์ด ๋ฐ˜์ง์ž…๋‹ˆ๋‹ค. ์นด๋ฉ”๋ผ๋Š” ๊ทธ์˜ ๊ฒฐ์—ฐํ•œ ํ‘œ์ •๊ณผ ์ˆจ ๋ง‰ํžˆ๋Š” ํ’๊ฒฝ์„ ๋ฒˆ๊ฐˆ์•„๊ฐ€๋ฉฐ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์†Œ๋‚˜๋ฌด๋“ค์ด ์Šค์ณ ์ง€๋‚˜๊ฐ€๊ณ , ํ•˜๋Š˜์€ ์„ ๋ช…ํ•œ ํŒŒ๋ž€์ƒ‰์ž…๋‹ˆ๋‹ค. ์ด ์žฅ๋ฉด์€ ํ™œ๊ธฐ์ฐจ๊ณ  ๊ฒฝ์Ÿ์ ์ธ ๋ถ„์œ„๊ธฐ๋ฅผ ์ž์•„๋ƒ…๋‹ˆ๋‹ค.",
"low quality, worst quality, deformed, distorted, warped, motion smear, motion artifacts, fused fingers, incorrect anatomy, strange hands, unattractive",
"assets/t2v_0.mp4",
],
],
inputs=[txt2vid_prompt, txt2vid_negative_prompt, txt2vid_output],
label="ํ…์ŠคํŠธ-๋น„๋””์˜ค ์ƒ์„ฑ ์˜ˆ์‹œ",
)
# Image to Video Tab
with gr.TabItem("์ด๋ฏธ์ง€๋กœ ๋น„๋””์˜ค ๋งŒ๋“ค๊ธฐ"):
with gr.Row():
with gr.Column():
img2vid_image = gr.Image(
type="filepath",
label="Step 1: ์ž…๋ ฅ ์ด๋ฏธ์ง€ ์—…๋กœ๋“œ",
elem_id="image_upload",
)
img2vid_prompt = gr.Textbox(
label="Step 2: ํ”„๋กฌํ”„ํŠธ ์ž…๋ ฅ",
placeholder="์ด๋ฏธ์ง€๋ฅผ ์–ด๋–ป๊ฒŒ ์• ๋‹ˆ๋ฉ”์ด์…˜ํ™”ํ• ์ง€ ์„ค๋ช…ํ•˜์„ธ์š” (์ตœ์†Œ 50์ž)...",
value="๊ฐˆ์ƒ‰ ๊ธด ๋จธ๋ฆฌ๋ฅผ ๊ฐ€์ง„ ์—ฌ์„ฑ์ด ๊ธˆ๋ฐœ์˜ ๊ธด ๋จธ๋ฆฌ๋ฅผ ๊ฐ€์ง„ ๋‹ค๋ฅธ ์—ฌ์„ฑ์„ ํ–ฅํ•ด ๋ฏธ์†Œ์ง“์Šต๋‹ˆ๋‹ค. ๊ฐˆ์ƒ‰ ๋จธ๋ฆฌ์˜ ์—ฌ์„ฑ์€ ๊ฒ€์€์ƒ‰ ์ž์ผ“์„ ์ž…๊ณ  ์žˆ์œผ๋ฉฐ ์˜ค๋ฅธ์ชฝ ๋บจ์— ์ž‘์€ ์ ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์นด๋ฉ”๋ผ ๊ฐ๋„๋Š” ๊ฐˆ์ƒ‰ ๋จธ๋ฆฌ ์—ฌ์„ฑ์˜ ์–ผ๊ตด์— ํด๋กœ์ฆˆ์—…๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์กฐ๋ช…์€ ์ž์—ฐ์Šค๋Ÿฝ๊ณ  ๋”ฐ๋œปํ•˜๋ฉฐ, ์„์–‘์—์„œ ์˜ค๋Š” ๋“ฏํ•œ ๋ถ€๋“œ๋Ÿฌ์šด ๋น›์ด ์žฅ๋ฉด์„ ๋น„์ถฅ๋‹ˆ๋‹ค. ์žฅ๋ฉด์€ ์‹ค์ œ ์˜์ƒ์ฒ˜๋Ÿผ ๋ณด์ž…๋‹ˆ๋‹ค.",
lines=5,
)
img2vid_enhance_toggle = Toggle(
label="ํ”„๋กฌํ”„ํŠธ ๊ฐœ์„ ",
value=False,
interactive=True,
)
img2vid_negative_prompt = gr.Textbox(
label="Step 3: ๋„ค๊ฑฐํ‹ฐ๋ธŒ ํ”„๋กฌํ”„ํŠธ ์ž…๋ ฅ",
placeholder="๋น„๋””์˜ค์—์„œ ์›ํ•˜์ง€ ์•Š๋Š” ์š”์†Œ๋ฅผ ์„ค๋ช…ํ•˜์„ธ์š”...",
value="low quality, worst quality, deformed, distorted, warped, motion smear, motion artifacts, fused fingers, incorrect anatomy, strange hands, unattractive",
lines=2,
)
img2vid_preset = gr.Dropdown(
choices=[p["label"] for p in preset_options],
value="768x512, 97 frames",
label="Step 3.1: ํ•ด์ƒ๋„ ํ”„๋ฆฌ์…‹ ์„ ํƒ",
)
img2vid_frame_rate = gr.Slider(
label="Step 3.2: ํ”„๋ ˆ์ž„ ๋ ˆ์ดํŠธ",
minimum=21,
maximum=30,
step=1,
value=25,
)
img2vid_advanced = create_advanced_options()
img2vid_generate = gr.Button(
"Step 6: ๋น„๋””์˜ค ์ƒ์„ฑ", variant="primary", size="lg"
)
with gr.Column():
img2vid_output = gr.Video(label="์ƒ์„ฑ๋œ ๋น„๋””์˜ค")
with gr.Row():
gr.Examples(
examples=[
[
"assets/i2v_i2.png",
"์—ฌ์„ฑ์ด ํฐ์ƒ‰ ์ „๊ธฐ ๋ฒ„๋„ˆ ์œ„์—์„œ ๋“๋Š” ๋ฌผ์ด ๋‹ด๊ธด ๋ƒ„๋น„๋ฅผ ์ “๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋ณด๋ผ์ƒ‰ ๋งค๋‹ˆํ์–ด๋ฅผ ๋ฐ”๋ฅธ ๊ทธ๋…€์˜ ์†์ด ํ•˜์–€ ๋ƒ„๋น„ ์•ˆ์—์„œ ๋‚˜๋ฌด ์ˆŸ๊ฐ€๋ฝ์„ ์›ํ˜•์œผ๋กœ ์›€์ง์ž…๋‹ˆ๋‹ค. ๋ƒ„๋น„๋Š” ๊ฒ€์€์ƒ‰ ๋ฒ„ํŠผ๊ณผ ๋””์ง€ํ„ธ ๋””์Šคํ”Œ๋ ˆ์ด๊ฐ€ ์žˆ๋Š” ํฐ์ƒ‰ ์ „๊ธฐ ๋ฒ„๋„ˆ ์œ„์— ๋†“์—ฌ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฒ„๋„ˆ๋Š” ์˜ค๋ฅธ์ชฝ ์•„๋ž˜ ๋ชจ์„œ๋ฆฌ์— ๋นจ๊ฐ„์ƒ‰๊ณผ ํฐ์ƒ‰ ์ฒดํฌ๋ฌด๋Šฌ ์ฒœ์ด ๋ถ€๋ถ„์ ์œผ๋กœ ๋ณด์ด๋Š” ํฐ์ƒ‰ ์กฐ๋ฆฌ๋Œ€ ์œ„์— ๋†“์—ฌ ์žˆ์Šต๋‹ˆ๋‹ค. ์นด๋ฉ”๋ผ ๊ฐ๋„๋Š” ์ •ํ™•ํžˆ ์œ„์—์„œ ๋‚ด๋ ค๋‹ค๋ณด๋Š” ๊ฐ๋„์ด๋ฉฐ ์žฅ๋ฉด ๋‚ด๋‚ด ๊ณ ์ •๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์กฐ๋ช…์€ ๋ฐ๊ณ  ๊ณ ๋ฅธ ์ค‘์„ฑ์ ์ธ ํฐ์ƒ‰ ๋น›์œผ๋กœ ์žฅ๋ฉด์„ ๋น„์ถฅ๋‹ˆ๋‹ค. ์žฅ๋ฉด์€ ์‹ค์ œ ์˜์ƒ์ฒ˜๋Ÿผ ๋ณด์ž…๋‹ˆ๋‹ค.",
"low quality, worst quality, deformed, distorted, warped, motion smear, motion artifacts, fused fingers, incorrect anatomy, strange hands, unattractive",
"assets/i2v_2.mp4",
],
[
"assets/i2v_i0.png",
"๊ธด ํ๋ฅด๋Š” ๋“œ๋ ˆ์Šค๋ฅผ ์ž…์€ ์—ฌ์„ฑ์ด ๋“คํŒ์— ์„œ์„œ ๋“ฑ์„ ์นด๋ฉ”๋ผ๋ฅผ ํ–ฅํ•œ ์ฑ„ ์ง€ํ‰์„ ์„ ๋ฐ”๋ผ๋ณด๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€์˜ ๋จธ๋ฆฌ์นด๋ฝ์€ ๊ธธ๊ณ  ๋ฐ์œผ๋ฉฐ ๋“ฑ ์•„๋ž˜๋กœ ํ˜๋Ÿฌ๋‚ด๋ฆฝ๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ํฐ ์ฐธ๋‚˜๋ฌด์˜ ๋„“๊ฒŒ ํผ์ง„ ๊ฐ€์ง€ ์•„๋ž˜์— ์„œ ์žˆ์Šต๋‹ˆ๋‹ค. ์™ผ์ชฝ์œผ๋กœ๋Š” ๋ง๋ผ๋ถ™์€ ์ž”๋”” ์œ„์— ํด๋ž˜์‹ํ•œ ๋ฏธ๊ตญ ์ž๋™์ฐจ๊ฐ€ ์ฃผ์ฐจ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฉ€๋ฆฌ์„œ๋Š” ํ•œ ๋Œ€์˜ ๋ถ€์„œ์ง„ ์ž๋™์ฐจ๊ฐ€ ์˜†์œผ๋กœ ๋ˆ„์›Œ ์žˆ์Šต๋‹ˆ๋‹ค. ์œ„์˜ ํ•˜๋Š˜์€ ์–ด๋‘์šด ํ•˜๋Š˜์„ ๋ฐฐ๊ฒฝ์œผ๋กœ ๋ฐ์€ ํฐ ๊ตฌ๋ฆ„์ด ๊ทน์ ์ธ ์บ”๋ฒ„์Šค๋ฅผ ์ด๋ฃจ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ „์ฒด ์ด๋ฏธ์ง€๋Š” ํ‘๋ฐฑ์œผ๋กœ, ๋น›๊ณผ ๊ทธ๋ฆผ์ž์˜ ๋Œ€๋น„๋ฅผ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ์—ฌ์„ฑ์ด ์ฒœ์ฒœํžˆ ์ž๋™์ฐจ๋ฅผ ํ–ฅํ•ด ๊ฑธ์–ด๊ฐ€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.",
"low quality, worst quality, deformed, distorted, warped, motion smear, motion artifacts, fused fingers, incorrect anatomy, strange hands, unattractive",
"assets/i2v_0.mp4",
],
[
"assets/i2v_i1.png",
"ํ•œ ์Œ์˜ ์†์ด ๋„์ž๊ธฐ ๋ฌผ๋ ˆ ์œ„์—์„œ ์ ํ†  ์กฐ๊ฐ์„ ๋ชจ์–‘ ์žก์•„ ์ ์ฐจ์ ์œผ๋กœ ์›๋ฟ” ๋ชจ์–‘์„ ๋งŒ๋“ค์–ด๊ฐ€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํ”„๋ ˆ์ž„ ๋ฐ–์˜ ์‚ฌ๋žŒ์˜ ์†์ด ์ ํ† ๋กœ ๋ฎ์—ฌ ์žˆ์œผ๋ฉฐ, ํšŒ์ „ํ•˜๋Š” ๋„์ž๊ธฐ ๋ฌผ๋ ˆ ์ค‘์•™์— ์ ํ†  ๋ฉ์–ด๋ฆฌ๋ฅผ ๋ถ€๋“œ๋Ÿฝ๊ฒŒ ๋ˆ„๋ฅด๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์†์€ ์›ํ˜•์œผ๋กœ ์›€์ง์ด๋ฉฐ, ์ ํ†  ์œ„์ชฝ์— ์ ์ฐจ์ ์œผ๋กœ ์›๋ฟ” ๋ชจ์–‘์„ ๋งŒ๋“ค์–ด๊ฐ‘๋‹ˆ๋‹ค. ์นด๋ฉ”๋ผ๋Š” ๋„์ž๊ธฐ ๋ฌผ๋ ˆ ๋ฐ”๋กœ ์œ„์— ์œ„์น˜ํ•˜์—ฌ ์ ํ† ๊ฐ€ ๋ชจ์–‘ ์žกํ˜€๊ฐ€๋Š” ๊ฒƒ์„ ์กฐ๊ฐ๋„๋กœ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์กฐ๋ช…์€ ๋ฐ๊ณ  ๊ณ ๋ฅด๋ฉฐ, ์ ํ† ์™€ ๊ทธ๊ฒƒ์„ ๋‹ค๋ฃจ๋Š” ์†์„ ๋ฐ๊ฒŒ ๋น„์ถฅ๋‹ˆ๋‹ค. ์žฅ๋ฉด์€ ์‹ค์ œ ์˜์ƒ์ฒ˜๋Ÿผ ์ดฌ์˜๋˜์—ˆ์Šต๋‹ˆ๋‹ค.",
"low quality, worst quality, deformed, distorted, warped, motion smear, motion artifacts, fused fingers, incorrect anatomy, strange hands, unattractive",
"assets/i2v_1.mp4",
],
],
inputs=[
img2vid_image,
img2vid_prompt,
img2vid_negative_prompt,
img2vid_output,
],
label="์ด๋ฏธ์ง€-๋น„๋””์˜ค ์ƒ์„ฑ ์˜ˆ์‹œ",
)
# Event handlers
# Event handlers
txt2vid_preset.change(
fn=preset_changed,
inputs=[txt2vid_preset],
outputs=txt2vid_advanced[3:]
)
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,
],
outputs=txt2vid_output,
concurrency_limit=1,
concurrency_id="generate_video",
queue=True,
)
img2vid_preset.change(
fn=preset_changed,
inputs=[img2vid_preset],
outputs=img2vid_advanced[3:]
)
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,
],
outputs=img2vid_output,
concurrency_limit=1,
concurrency_id="generate_video",
queue=True,
)
if __name__ == "__main__":
iface.queue(max_size=64, default_concurrency_limit=1, api_open=False).launch(
share=True, show_api=False
)