app.py

from diffusers import DiffusionPipeline, DPMSolverMultistepScheduler
from utils import write_video, dummy, preprocess_image, preprocess_mask_image
from PIL import Image
import gradio as gr
import torch
import os
os.environ["CUDA_VISIBLE_DEVICES"]="0"


orig_prompt = "Ancient underground architectural ruins of Hong Kong in a flooded apocalypse landscape of dead skyscrapers"
orig_negative_prompt = "lurry, bad art, blurred, text, watermark"
model_list = ["stabilityai/stable-diffusion-2-inpainting", "runwayml/stable-diffusion-inpainting"]

def stable_diffusion_zoom_out(
  repo_id,
  original_prompt,
  negative_prompt,
  step_size,
  num_frames,
  fps,
  num_inference_steps
    ):
    
    pipe = DiffusionPipeline.from_pretrained(repo_id, torch_dtype=torch.float16)
    pipe.set_use_memory_efficient_attention_xformers(True)
    pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
    pipe = pipe.to("cuda")
    pipe.safety_checker = dummy

    new_image = Image.new(mode="RGBA", size=(512,512)) 
    current_image, mask_image = preprocess_mask_image(new_image)
    
    current_image = pipe(prompt=[original_prompt], negative_prompt=[negative_prompt], image=current_image, mask_image=mask_image, num_inference_steps=num_inference_steps).images[0]

    
    all_frames = []
    all_frames.append(current_image)

    for i in range(num_frames):
        prev_image = preprocess_image(current_image, step_size, 512)
        current_image = prev_image
        current_image, mask_image = preprocess_mask_image(current_image)
        current_image = pipe(prompt=[original_prompt], negative_prompt=[negative_prompt], image=current_image, mask_image=mask_image, num_inference_steps=num_inference_steps).images[0]
        
        current_image.paste(prev_image, mask=prev_image)
        all_frames.append(current_image)

    save_path = "output.mp4"  
    write_video(save_path, all_frames, fps=fps)
    return save_path


inputs = [
  gr.Dropdown(model_list, value=model_list[0], label="Model"),
  gr.inputs.Textbox(lines=5, default=orig_prompt, label="Prompt"),
  gr.inputs.Textbox(lines=1, default=orig_negative_prompt, label="Negative Prompt"),
  gr.inputs.Slider(minimum=1, maximum=120, default=25, step=5, label="Steps"),
  gr.inputs.Slider(minimum=1, maximum=100, default=10, step=1, label="Frames"),
  gr.inputs.Slider(minimum=1, maximum=100, default=16, step=1, label="FPS"),
  gr.inputs.Slider(minimum=1, maximum=100, default=15, step=1, label="Inference Steps")
]

output = gr.outputs.Video()
examples = [
  ["stabilityai/stable-diffusion-2-inpainting", orig_prompt, orig_negative_prompt, 25, 10, 16, 15],
]

title = "Stable Diffusion Infinite Zoom Out"
description = """<p>For faster inference without waiting in queue, you may duplicate the space and upgrade to GPU in settings.
<br/>
<a href="https://huggingface.co/spaces/kadirnar/stable-diffusion-2-infinite-zoom-out?duplicate=true">
<img style="margin-top: 0em; margin-bottom: 0em" src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a>
<p/>"""

demo_app = gr.Interface(
    fn=stable_diffusion_zoom_out,
    description=description,
    inputs=inputs,
    outputs=output,
    title=title,
    theme='huggingface',
    examples=examples,
    cache_examples=True
)
demo_app.launch(debug=True, enable_queue=True)