Anurag181011 commited on
Commit
358c39a
·
1 Parent(s): b5cdc6f
Files changed (1) hide show
  1. app.py +24 -21
app.py CHANGED
@@ -4,50 +4,53 @@ import gradio as gr
4
  from diffusers import StableDiffusionImg2ImgPipeline
5
  from PIL import Image
6
 
7
- # Force CUDA usage
8
  os.environ["CUDA_VISIBLE_DEVICES"] = "0"
9
  torch.backends.cudnn.benchmark = True
10
  torch.backends.cuda.matmul.allow_tf32 = True
11
 
12
- device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
13
- print(f"Using device: {device}")
 
14
 
15
- # Ensure torch is properly installed
16
  try:
17
  torch.zeros(1).to(device)
18
- print("Torch initialized successfully on", device)
19
  except Exception as e:
20
- print("Torch initialization error:", e)
21
 
22
  # Load the optimized Stable Diffusion model
23
  model_id = "nitrosocke/Ghibli-Diffusion"
24
 
25
  pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
26
  model_id,
27
- torch_dtype=torch.float16,
28
  use_safetensors=True,
29
  low_cpu_mem_usage=True
30
- ).to("cuda")
31
 
32
- # Try enabling xformers, but fail gracefully
33
  try:
34
  pipe.enable_xformers_memory_efficient_attention()
35
  print("✅ xFormers enabled!")
36
  except Exception as e:
37
  print(f"⚠️ xFormers not available: {e}")
38
- pipe.enable_xformers_memory_efficient_attention()
 
39
  pipe.enable_model_cpu_offload()
40
  pipe.enable_vae_slicing()
41
  pipe.enable_attention_slicing()
42
 
43
- # Enhanced prompt for Studio Ghibli-style transformation
44
  prompt = (
45
- "Beautiful Studio Ghibli anime-style portrait, breathtaking landscape background, "
46
- "soft pastel colors, hand-painted texture, cinematic lighting, dreamy atmosphere, "
47
- "vibrant and rich details, Miyazaki aesthetic, magical realism, watercolor effect, "
48
- "warm sunlight, stunning composition, high detail, fantasy world."
49
  )
50
 
 
51
  def transform_image(input_image):
52
  input_image = input_image.resize((512, 512))
53
 
@@ -55,8 +58,8 @@ def transform_image(input_image):
55
  prompt=prompt,
56
  image=input_image,
57
  strength=0.65,
58
- guidance_scale=4.5,
59
- num_inference_steps=20,
60
  )
61
 
62
  return output.images[0]
@@ -64,10 +67,10 @@ def transform_image(input_image):
64
  # Gradio UI
65
  demo = gr.Interface(
66
  fn=transform_image,
67
- inputs=gr.Image(type="pil", label="Upload your portrait/photo"),
68
- outputs=gr.Image(type="pil", label="Studio Ghibli Style Output"),
69
- title="Studio Ghibli AI Converter",
70
- description="Upload a portrait or photo to transform it into a Studio Ghibli-style image.",
71
  )
72
 
73
  if __name__ == "__main__":
 
4
  from diffusers import StableDiffusionImg2ImgPipeline
5
  from PIL import Image
6
 
7
+ # Force CUDA usage if available
8
  os.environ["CUDA_VISIBLE_DEVICES"] = "0"
9
  torch.backends.cudnn.benchmark = True
10
  torch.backends.cuda.matmul.allow_tf32 = True
11
 
12
+ # Check if GPU is available
13
+ device = "cuda" if torch.cuda.is_available() else "cpu"
14
+ print(f"🚀 Using device: {device}")
15
 
16
+ # Ensure Torch is correctly initialized
17
  try:
18
  torch.zeros(1).to(device)
19
+ print("Torch initialized successfully on", device)
20
  except Exception as e:
21
+ print("⚠️ Torch initialization error:", e)
22
 
23
  # Load the optimized Stable Diffusion model
24
  model_id = "nitrosocke/Ghibli-Diffusion"
25
 
26
  pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
27
  model_id,
28
+ torch_dtype=torch.float16 if device == "cuda" else torch.float32,
29
  use_safetensors=True,
30
  low_cpu_mem_usage=True
31
+ ).to(device)
32
 
33
+ # Try enabling xFormers for memory efficiency
34
  try:
35
  pipe.enable_xformers_memory_efficient_attention()
36
  print("✅ xFormers enabled!")
37
  except Exception as e:
38
  print(f"⚠️ xFormers not available: {e}")
39
+
40
+ # Apply additional optimizations for performance
41
  pipe.enable_model_cpu_offload()
42
  pipe.enable_vae_slicing()
43
  pipe.enable_attention_slicing()
44
 
45
+ # Enhanced Studio Ghibli-style transformation prompt
46
  prompt = (
47
+ "Studio Ghibli anime-style illustration, magical landscape, soft pastel colors, "
48
+ "hand-painted textures, cinematic lighting, dreamy atmosphere, vibrant and rich details, "
49
+ "Miyazaki-inspired fantasy world, watercolor aesthetic, warm sunlight, intricate composition, "
50
+ "high detail, whimsical and nostalgic beauty."
51
  )
52
 
53
+ # Image transformation function
54
  def transform_image(input_image):
55
  input_image = input_image.resize((512, 512))
56
 
 
58
  prompt=prompt,
59
  image=input_image,
60
  strength=0.65,
61
+ guidance_scale=5.0, # Slightly increased for better stylization
62
+ num_inference_steps=25, # More steps for higher quality output
63
  )
64
 
65
  return output.images[0]
 
67
  # Gradio UI
68
  demo = gr.Interface(
69
  fn=transform_image,
70
+ inputs=gr.Image(type="pil", label="Upload a Portrait/Photo"),
71
+ outputs=gr.Image(type="pil", label="Studio Ghibli-Style Output"),
72
+ title="🎨 Studio Ghibli AI Art Generator",
73
+ description="Upload a portrait or a photo and transform it into a breathtaking Studio Ghibli-style masterpiece!",
74
  )
75
 
76
  if __name__ == "__main__":