init
Browse files
app.py
ADDED
@@ -0,0 +1,116 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
from diffusers import DiffusionPipeline
|
3 |
+
import torch
|
4 |
+
import os
|
5 |
+
|
6 |
+
# Ensure necessary libraries are installed
|
7 |
+
# pip install diffusers --upgrade
|
8 |
+
# pip install invisible_watermark transformers accelerate safetensors gradio torch
|
9 |
+
|
10 |
+
model_id = "stabilityai/stable-diffusion-xl-base-1.0"
|
11 |
+
|
12 |
+
# Determine device and dtype
|
13 |
+
if torch.cuda.is_available():
|
14 |
+
device = "cuda"
|
15 |
+
dtype = torch.float16
|
16 |
+
print("Using CUDA (GPU).")
|
17 |
+
# elif torch.backends.mps.is_available(): # Uncomment for MacOS Metal support
|
18 |
+
# device = "mps"
|
19 |
+
# dtype = torch.float16
|
20 |
+
# print("Using MPS (Apple Silicon GPU).")
|
21 |
+
else:
|
22 |
+
device = "cpu"
|
23 |
+
dtype = torch.float32
|
24 |
+
print("Using CPU.")
|
25 |
+
|
26 |
+
# Load the Stable Diffusion XL pipeline
|
27 |
+
# Using float16 and safetensors for efficiency if on GPU
|
28 |
+
# variant="fp16" loads the fp16 weights
|
29 |
+
try:
|
30 |
+
pipe = DiffusionPipeline.from_pretrained(
|
31 |
+
model_id,
|
32 |
+
torch_dtype=dtype,
|
33 |
+
use_safetensors=True,
|
34 |
+
variant="fp16" if device!= "cpu" else None # Only use fp16 variant if not on CPU
|
35 |
+
)
|
36 |
+
pipe.to(device)
|
37 |
+
|
38 |
+
# Optional: Enable CPU offloading if VRAM is limited (only works on CUDA)
|
39 |
+
if device == "cuda":
|
40 |
+
try:
|
41 |
+
# Check VRAM - this is a rough estimate, adjust threshold as needed
|
42 |
+
total_vram_gb = torch.cuda.get_device_properties(0).total_memory / (1024**3)
|
43 |
+
if total_vram_gb < 10: # Example threshold: less than 10GB VRAM
|
44 |
+
print(f"Low VRAM ({total_vram_gb:.2f}GB detected). Enabling model CPU offload.")
|
45 |
+
pipe.enable_model_cpu_offload()
|
46 |
+
except Exception as offload_err:
|
47 |
+
print(f"Could not check VRAM or enable offload: {offload_err}")
|
48 |
+
|
49 |
+
|
50 |
+
# Optional: Use torch.compile for speedup (requires torch >= 2.0)
|
51 |
+
# if device!= "cpu" and hasattr(torch, "compile"):
|
52 |
+
# try:
|
53 |
+
# print("Attempting to compile the UNet...")
|
54 |
+
# pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
|
55 |
+
# print("UNet compiled successfully.")
|
56 |
+
# except Exception as compile_err:
|
57 |
+
# print(f"Torch compile failed: {compile_err}")
|
58 |
+
|
59 |
+
print(f"SDXL pipeline loaded successfully on {device}.")
|
60 |
+
|
61 |
+
except Exception as e:
|
62 |
+
print(f"Error loading SDXL pipeline: {e}")
|
63 |
+
pipe = None
|
64 |
+
|
65 |
+
def generate_image(prompt):
|
66 |
+
"""Generates an image based on the text prompt."""
|
67 |
+
if pipe is None:
|
68 |
+
# Handle case where pipeline failed to load
|
69 |
+
# Create a placeholder image or return an error message
|
70 |
+
from PIL import Image, ImageDraw, ImageFont
|
71 |
+
img = Image.new('RGB', (512, 512), color = (200, 200, 200))
|
72 |
+
d = ImageDraw.Draw(img)
|
73 |
+
try:
|
74 |
+
# Try to load a default font
|
75 |
+
fnt = ImageFont.truetype("arial.ttf", 15)
|
76 |
+
except IOError:
|
77 |
+
fnt = ImageFont.load_default()
|
78 |
+
d.text((10,10), "Error: Model pipeline failed to load.", fill=(255,0,0), font=fnt)
|
79 |
+
return img
|
80 |
+
|
81 |
+
if not prompt:
|
82 |
+
return None # Return nothing if prompt is empty
|
83 |
+
|
84 |
+
print(f"Generating image for prompt: '{prompt}'")
|
85 |
+
try:
|
86 |
+
# Generate the image
|
87 |
+
# Using default steps/guidance scale, can be customized
|
88 |
+
with torch.inference_mode(): # Use inference mode for efficiency
|
89 |
+
image = pipe(prompt=prompt, num_inference_steps=30).images
|
90 |
+
print("Image generated successfully.")
|
91 |
+
return image
|
92 |
+
except Exception as e:
|
93 |
+
print(f"Error during image generation: {e}")
|
94 |
+
# Return an error image or message
|
95 |
+
from PIL import Image, ImageDraw, ImageFont
|
96 |
+
img = Image.new('RGB', (512, 512), color = (200, 200, 200))
|
97 |
+
d = ImageDraw.Draw(img)
|
98 |
+
try: fnt = ImageFont.truetype("arial.ttf", 15)
|
99 |
+
except IOError: fnt = ImageFont.load_default()
|
100 |
+
d.text((10,10), f"Error generating image:\n{e}", fill=(255,0,0), font=fnt)
|
101 |
+
return img
|
102 |
+
|
103 |
+
|
104 |
+
# Create the Gradio interface
|
105 |
+
demo = gr.Interface(
|
106 |
+
fn=generate_image,
|
107 |
+
inputs=gr.Textbox(label="Enter Text Prompt", placeholder="e.g., 'An astronaut riding a green horse'"),
|
108 |
+
outputs=gr.Image(label="Generated Image", type="pil"),
|
109 |
+
title="Text-to-Image Generation with Stable Diffusion XL",
|
110 |
+
description=f"Generate images from text prompts using the {model_id} model. Loading and inference might take a moment, especially on the first run or on CPU.",
|
111 |
+
examples=["A high-tech cityscape at sunset, cinematic lighting"]
|
112 |
+
)
|
113 |
+
|
114 |
+
if __name__ == "__main__":
|
115 |
+
# Launch the Gradio app
|
116 |
+
demo.launch(debug=True)
|