update

app.py

@@ -0,0 +1,77 @@
+from diffusers import DiffusionPipeline, DPMSolverMultistepScheduler
+from utils import write_video, dummy
+from PIL import Image
+import numpy as np
+import os
+os.environ["CUDA_VISIBLE_DEVICES"]="0"
+import torch
+import gradio as gr
+
+
+def stable_diffusion_zoom_out(
+  repo_id="stabilityai/stable-diffusion-2-inpainting",
+  original_prompt="a dog",
+  negative_prompt="a cat",
+  steps=32,
+  num_frames=10,
+):
+  pipe = DiffusionPipeline.from_pretrained(repo_id, torch_dtype=torch.float16, revision="fp16")
+  pipe.set_use_memory_efficient_attention_xformers(True)
+  pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
+  pipe = pipe.to("cuda")
+  pipe.safety_checker = dummy
+
+  current_image = Image.new(mode="RGBA", size=(512, 512))
+  mask_image = np.array(current_image)[:,:,3] # assume image has alpha mask (use .mode to check for "RGBA")
+  mask_image = Image.fromarray(255-mask_image).convert("RGB")
+  current_image = current_image.convert("RGB")
+
+  num_images = 1
+  prompt = [original_prompt] * num_images
+  negative_prompt = [negative_prompt] * num_images
+
+  images = pipe(prompt=prompt, negative_prompt=negative_prompt, image=current_image, mask_image=mask_image, num_inference_steps=25)[0]
+  current_image = images[0]
+
+  all_frames = []
+  all_frames.append(current_image)
+
+  for i in range(num_frames):
+    next_image = np.array(current_image.convert("RGBA"))*0
+    prev_image = current_image.resize((512-2*steps,512-2*steps))
+    prev_image = prev_image.convert("RGBA")
+    prev_image = np.array(prev_image)
+    next_image[:, :, 3] = 1
+    next_image[steps:512-steps,steps:512-steps,:] = prev_image
+    prev_image = Image.fromarray(next_image)
+    current_image = prev_image
+    mask_image = np.array(current_image)[:,:,3] # assume image has alpha mask (use .mode to check for "RGBA")
+    mask_image = Image.fromarray(255-mask_image).convert("RGB")
+    current_image = current_image.convert("RGB")
+    images = pipe(prompt=prompt, negative_prompt=negative_prompt, image=current_image, mask_image=mask_image, num_inference_steps=25)[0]
+    current_image = images[0]
+    current_image.paste(prev_image, mask=prev_image)
+    all_frames.append(current_image)
+    
+  save_path = "infinite_zoom_out.mp4"  
+  write_video(save_path, all_frames, fps=16)
+  return save_path
+
+inputs = [
+  gr.Dropdown(["stabilityai/stable-diffusion-2-inpainting"], label="Model"),
+  gr.inputs.Textbox(lines=1, default="a dog", label="Prompt"),
+  gr.inputs.Textbox(lines=1, default="a cat", label="Negative Prompt"),
+  gr.inputs.Slider(minimum=1, maximum=64, default=32, label="Steps"),
+  gr.inputs.Slider(minimum=1, maximum=100, default=10, label="Frames"),
+]
+output = gr.outputs.Video()
+title = "Stable Diffusion Infinite Zoom Out"
+
+demo_app = gr.Interface(
+    fn=stable_diffusion_zoom_out,
+    inputs=inputs,
+    outputs=output,
+    title=title,
+    theme='huggingface',
+)
+demo_app.launch(debug=True, enable_queue=True)

utils.py

@@ -0,0 +1,25 @@
+import numpy as np
+import cv2
+
+def write_video(file_path, frames, fps):
+    """
+    Writes frames to an mp4 video file
+    :param file_path: Path to output video, must end with .mp4
+    :param frames: List of PIL.Image objects
+    :param fps: Desired frame rate
+    """
+
+    w, h = frames[0].size
+    fourcc = cv2.VideoWriter_fourcc('m', 'p', '4', 'v')
+    writer = cv2.VideoWriter(file_path, fourcc, fps, (w, h))
+
+    for frame in frames:
+        np_frame = np.array(frame.convert('RGB'))
+        cv_frame = cv2.cvtColor(np_frame, cv2.COLOR_RGB2BGR)
+        writer.write(cv_frame)
+
+    writer.release() 
+
+
+def dummy(images, **kwargs):
+    return images, False