InstaVideo / app_14B.py
rahul7star's picture
Update app_14B.py
23b71a7 verified
import spaces
import os
import uuid
import torch
import logging
import tempfile
import numpy as np
import gradio as gr
from datetime import datetime
from diffusers import WanImageToVideoPipeline
from diffusers.utils import export_to_video
from huggingface_hub import upload_file
from PIL import Image
# ----------------- Setup -----------------
logging.basicConfig(level=logging.INFO)
HF_MODEL = "rahul7star/rahulAI"
dtype = torch.bfloat16
device = "cuda"
model_id = "FastDM/Wan2.2-I2V-A14B-Merge-Lightning-V1.0-Diffusers"
pipe = WanImageToVideoPipeline.from_pretrained(model_id, torch_dtype=dtype)
pipe.to(device)
default_negative_prompt = (
"色调艳丽,过曝,静态,细节模糊不清,字幕,风格,作品,画作,画面,静止,整体发灰,最差质量,低质量,"
"JPEG压缩残留,丑陋的,残缺的,多余的手指,画得不好的手部,画得不好的脸部,畸形的,毁容的,形态畸形的肢体,"
"手指融合,静止不动的画面,杂乱的背景,三条腿,背景人很多,倒着走"
)
# ----------------- Upload helper -----------------
def upscale_and_upload_4k(input_video_path: str, input_image, summary_text: str) -> str:
"""
Upload video (4K), input image, and summary text to HF.
"""
logging.info(f"Upscaling video to 4K for upload: {input_video_path}")
# Upscale video
with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmp_upscaled:
upscaled_path = tmp_upscaled.name
cmd = [
"ffmpeg", "-i", input_video_path,
"-vf", "scale=3840:2160:flags=lanczos",
"-c:v", "libx264", "-crf", "18", "-preset", "slow", "-y", upscaled_path,
]
os.system(" ".join(cmd)) # safer: subprocess.run, but HF Spaces sometimes picky
# Create HF folder
today_str = datetime.now().strftime("%Y-%m-%d")
unique_subfolder = f"upload_{uuid.uuid4().hex[:8]}"
hf_folder = f"{today_str}-WAN-I2V/{unique_subfolder}"
# Upload video
video_filename = os.path.basename(input_video_path)
video_hf_path = f"{hf_folder}/{video_filename}"
upload_file(upscaled_path, video_hf_path, repo_id=HF_MODEL, repo_type="model",
token=os.environ.get("HUGGINGFACE_HUB_TOKEN"))
# Upload image
with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as tmp_img:
if isinstance(input_image, str):
import shutil
shutil.copy(input_image, tmp_img.name)
else:
input_image.save(tmp_img.name, format="PNG")
tmp_img_path = tmp_img.name
image_hf_path = f"{hf_folder}/input_image.png"
upload_file(tmp_img_path, image_hf_path, repo_id=HF_MODEL, repo_type="model",
token=os.environ.get("HUGGINGFACE_HUB_TOKEN"))
# Upload summary
summary_file = tempfile.NamedTemporaryFile(delete=False, suffix=".txt").name
with open(summary_file, "w", encoding="utf-8") as f:
f.write(summary_text)
summary_hf_path = f"{hf_folder}/summary.txt"
upload_file(summary_file, summary_hf_path, repo_id=HF_MODEL, repo_type="model",
token=os.environ.get("HUGGINGFACE_HUB_TOKEN"))
# Cleanup
os.remove(upscaled_path)
os.remove(tmp_img_path)
os.remove(summary_file)
return hf_folder
# ----------------- Video generation -----------------
def get_duration(
input_image,
prompt,
negative_prompt,
duration_seconds,
guidance_scale,
guidance_scale_2,
steps,
seed,
randomize_seed,
progress,
):
return steps * 15
@spaces.GPU(duration=70)
def generate_video(input_image, prompt, negative_prompt=default_negative_prompt,
duration_seconds=2, guidance_scale=3.5, steps=40, seed=0):
if input_image is None:
return None, "Please upload an image!"
# Ensure divisible by patch size
max_area = 480 * 832
aspect_ratio = input_image.height / input_image.width
mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1]
height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
input_image = input_image.resize((width, height))
generator = torch.Generator(device=device).manual_seed(int(seed))
with torch.inference_mode():
output_frames_list = pipe(
image=input_image,
prompt=prompt,
negative_prompt=negative_prompt,
height=height,
width=width,
num_frames=int(duration_seconds * 16), # 16 fps
guidance_scale=float(guidance_scale),
num_inference_steps=int(steps),
generator=generator,
).frames[0]
# Save temp video
with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmpfile:
video_path = tmpfile.name
export_to_video(output_frames_list, video_path, fps=16)
# Upload to HF
#hf_folder = upscale_and_upload_4k(video_path, input_image, prompt)
return video_path, f"✅ Uploaded to HF: {hf_folder}"
# ----------------- Gradio UI -----------------
with gr.Blocks() as demo:
gr.Markdown("# 🖼️➡️🎥 Image to Video with Wan 2.2 I2V (14B Lightning)")
with gr.Row():
with gr.Column():
input_image = gr.Image(type="pil", label="Upload an Image")
prompt = gr.Textbox(lines=4, label="Prompt")
negative_prompt = gr.Textbox(value=default_negative_prompt, lines=3, label="Negative Prompt")
duration = gr.Slider(1, 4, value=2, step=1, label="Duration (seconds)")
guidance_scale = gr.Slider(0, 10, value=3.5, step=0.5, label="Guidance Scale")
steps = gr.Slider(10, 50, value=40, step=1, label="Inference Steps")
seed = gr.Number(value=0, precision=0, label="Seed")
generate_btn = gr.Button("🚀 Generate Video")
with gr.Column():
output_video = gr.Video(label="Generated Video")
upload_status = gr.Textbox(label="Upload Status", interactive=False)
generate_btn.click(
generate_video,
inputs=[input_image, prompt, negative_prompt, duration, guidance_scale, steps, seed],
outputs=[output_video, upload_status],
)
if __name__ == "__main__":
demo.launch()