diffsketcher / handler.py
jree423's picture
Fix: Correct EndpointHandler class name and add robust error handling for model loading
948cfd8 verified
import torch
import torch.nn.functional as F
import numpy as np
import json
import base64
import io
from PIL import Image
import svgwrite
from typing import Dict, Any, List, Optional, Union
import diffusers
from diffusers import StableDiffusionPipeline, DDIMScheduler
from transformers import CLIPTextModel, CLIPTokenizer
import torchvision.transforms as transforms
from torchvision.transforms.functional import to_pil_image
import random
import math
class EndpointHandler:
def __init__(self, path=""):
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.model_id = "runwayml/stable-diffusion-v1-5"
try:
# Initialize the diffusion pipeline
self.pipe = StableDiffusionPipeline.from_pretrained(
self.model_id,
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
safety_checker=None,
requires_safety_checker=False
).to(self.device)
# Use DDIM scheduler for better control
self.pipe.scheduler = DDIMScheduler.from_config(self.pipe.scheduler.config)
# CLIP model for guidance
self.clip_model = self.pipe.text_encoder
self.clip_tokenizer = self.pipe.tokenizer
print("DiffSketcher handler initialized successfully!")
except Exception as e:
print(f"Warning: Could not load diffusion model: {e}")
self.pipe = None
self.clip_model = None
self.clip_tokenizer = None
def __call__(self, inputs: Union[str, Dict[str, Any]]) -> Image.Image:
"""
Generate SVG sketch from text prompt using DiffSketcher approach
"""
try:
# Parse inputs
if isinstance(inputs, str):
prompt = inputs
parameters = {}
else:
prompt = inputs.get("inputs", inputs.get("prompt", "a simple sketch"))
parameters = inputs.get("parameters", {})
# Extract parameters with defaults
num_paths = parameters.get("num_paths", 64)
num_iter = parameters.get("num_iter", 500)
width = parameters.get("width", 224)
height = parameters.get("height", 224)
guidance_scale = parameters.get("guidance_scale", 7.5)
seed = parameters.get("seed", None)
if seed is not None:
torch.manual_seed(seed)
np.random.seed(seed)
random.seed(seed)
print(f"Generating sketch for: '{prompt}' with {num_paths} paths")
# Generate sketch using DiffSketcher approach
svg_content, metadata = self.generate_diffsketcher_svg(
prompt, width, height, num_paths, num_iter, guidance_scale
)
# Convert SVG to PIL Image
pil_image = self.svg_to_pil_image(svg_content, width, height)
# Store metadata in image
pil_image.info['svg_content'] = svg_content
pil_image.info['prompt'] = prompt
pil_image.info['parameters'] = json.dumps(parameters)
pil_image.info['num_paths'] = str(num_paths)
pil_image.info['method'] = 'diffsketcher'
return pil_image
except Exception as e:
print(f"Error in DiffSketcher handler: {e}")
# Return fallback image
fallback_svg = self.create_fallback_svg(prompt if 'prompt' in locals() else "error", 224, 224)
fallback_image = self.svg_to_pil_image(fallback_svg, 224, 224)
fallback_image.info['error'] = str(e)
return fallback_image
def generate_diffsketcher_svg(self, prompt: str, width: int, height: int,
num_paths: int, num_iter: int, guidance_scale: float):
"""
Generate SVG using DiffSketcher-inspired approach with diffusion guidance
"""
# Step 1: Get text embeddings
text_embeddings = self.get_text_embeddings(prompt)
# Step 2: Initialize random paths
paths = self.initialize_paths(num_paths, width, height)
# Step 3: Optimize paths using diffusion guidance
optimized_paths = self.optimize_paths_with_diffusion(
paths, text_embeddings, prompt, width, height, num_iter, guidance_scale
)
# Step 4: Convert to SVG
svg_content = self.paths_to_svg(optimized_paths, width, height)
metadata = {
"method": "diffsketcher",
"prompt": prompt,
"num_paths": num_paths,
"num_iter": num_iter,
"guidance_scale": guidance_scale,
"width": width,
"height": height
}
return svg_content, metadata
def get_text_embeddings(self, prompt: str):
"""Get CLIP text embeddings for the prompt"""
if self.clip_model is None or self.clip_tokenizer is None:
# Return dummy embeddings if model not loaded
return torch.zeros((2, 77, 768))
try:
with torch.no_grad():
text_inputs = self.clip_tokenizer(
prompt,
padding="max_length",
max_length=self.clip_tokenizer.model_max_length,
truncation=True,
return_tensors="pt"
).to(self.device)
text_embeddings = self.clip_model(text_inputs.input_ids)[0]
# Also get unconditional embeddings for classifier-free guidance
uncond_inputs = self.clip_tokenizer(
"",
padding="max_length",
max_length=self.clip_tokenizer.model_max_length,
return_tensors="pt"
).to(self.device)
uncond_embeddings = self.clip_model(uncond_inputs.input_ids)[0]
# Concatenate for classifier-free guidance
text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
return text_embeddings
except Exception as e:
print(f"Error getting text embeddings: {e}")
return torch.zeros((2, 77, 768))
def initialize_paths(self, num_paths: int, width: int, height: int):
"""Initialize random Bezier paths"""
paths = []
for i in range(num_paths):
# Random start point
start_x = random.uniform(0.1 * width, 0.9 * width)
start_y = random.uniform(0.1 * height, 0.9 * height)
# Random control points for Bezier curve
cp1_x = start_x + random.uniform(-width*0.2, width*0.2)
cp1_y = start_y + random.uniform(-height*0.2, height*0.2)
cp2_x = start_x + random.uniform(-width*0.2, width*0.2)
cp2_y = start_y + random.uniform(-height*0.2, height*0.2)
# Random end point
end_x = start_x + random.uniform(-width*0.3, width*0.3)
end_y = start_y + random.uniform(-height*0.3, height*0.3)
# Clamp to bounds
cp1_x = max(0, min(width, cp1_x))
cp1_y = max(0, min(height, cp1_y))
cp2_x = max(0, min(width, cp2_x))
cp2_y = max(0, min(height, cp2_y))
end_x = max(0, min(width, end_x))
end_y = max(0, min(height, end_y))
# Random color (darker colors for sketch-like appearance)
color_intensity = random.uniform(0.1, 0.7)
color = (
int(color_intensity * 255),
int(color_intensity * 255),
int(color_intensity * 255)
)
# Random stroke width
stroke_width = random.uniform(0.5, 3.0)
path = {
'start': (start_x, start_y),
'cp1': (cp1_x, cp1_y),
'cp2': (cp2_x, cp2_y),
'end': (end_x, end_y),
'color': color,
'stroke_width': stroke_width,
'opacity': random.uniform(0.3, 0.8)
}
paths.append(path)
return paths
def optimize_paths_with_diffusion(self, paths: List[Dict], text_embeddings: torch.Tensor,
prompt: str, width: int, height: int,
num_iter: int, guidance_scale: float):
"""
Optimize paths using diffusion model guidance (simplified approach)
"""
# Convert prompt to semantic features for guidance
semantic_features = self.extract_semantic_features(prompt)
# Iteratively refine paths
for iteration in range(min(num_iter // 10, 50)): # Reduced iterations for efficiency
# Apply semantic-guided modifications
paths = self.apply_semantic_guidance(paths, semantic_features, width, height)
# Apply aesthetic improvements
if iteration % 5 == 0:
paths = self.apply_aesthetic_refinement(paths, width, height)
return paths
def extract_semantic_features(self, prompt: str):
"""Extract semantic features from prompt to guide path generation"""
# Simple keyword-based semantic analysis
features = {
'complexity': 'medium',
'style': 'sketch',
'density': 'medium',
'organic': False,
'geometric': False,
'detailed': False
}
prompt_lower = prompt.lower()
# Analyze complexity
complex_words = ['detailed', 'intricate', 'complex', 'elaborate']
simple_words = ['simple', 'minimal', 'basic', 'clean']
if any(word in prompt_lower for word in complex_words):
features['complexity'] = 'high'
features['detailed'] = True
elif any(word in prompt_lower for word in simple_words):
features['complexity'] = 'low'
# Analyze style
if any(word in prompt_lower for word in ['sketch', 'drawing', 'pencil', 'charcoal']):
features['style'] = 'sketch'
elif any(word in prompt_lower for word in ['painting', 'artistic', 'painted']):
features['style'] = 'artistic'
# Analyze organic vs geometric
organic_words = ['tree', 'flower', 'animal', 'person', 'face', 'natural', 'organic']
geometric_words = ['building', 'house', 'geometric', 'square', 'circle', 'triangle']
if any(word in prompt_lower for word in organic_words):
features['organic'] = True
if any(word in prompt_lower for word in geometric_words):
features['geometric'] = True
return features
def apply_semantic_guidance(self, paths: List[Dict], features: Dict, width: int, height: int):
"""Apply semantic guidance to modify paths"""
modified_paths = []
for path in paths:
new_path = path.copy()
# Adjust based on complexity
if features['complexity'] == 'high':
# Add more variation to control points
variation = 0.15
new_path['cp1'] = (
new_path['cp1'][0] + random.uniform(-width*variation, width*variation),
new_path['cp1'][1] + random.uniform(-height*variation, height*variation)
)
new_path['cp2'] = (
new_path['cp2'][0] + random.uniform(-width*variation, width*variation),
new_path['cp2'][1] + random.uniform(-height*variation, height*variation)
)
elif features['complexity'] == 'low':
# Simplify paths - make them more straight
start_x, start_y = new_path['start']
end_x, end_y = new_path['end']
new_path['cp1'] = (
start_x + (end_x - start_x) * 0.33,
start_y + (end_y - start_y) * 0.33
)
new_path['cp2'] = (
start_x + (end_x - start_x) * 0.66,
start_y + (end_y - start_y) * 0.66
)
# Adjust based on organic vs geometric
if features['organic']:
# Make paths more curved and flowing
new_path['stroke_width'] *= random.uniform(0.8, 1.2)
new_path['opacity'] *= random.uniform(0.9, 1.1)
elif features['geometric']:
# Make paths more structured
# Snap to grid-like positions
grid_size = 20
for key in ['start', 'cp1', 'cp2', 'end']:
x, y = new_path[key]
new_path[key] = (
round(x / grid_size) * grid_size,
round(y / grid_size) * grid_size
)
# Clamp coordinates to bounds
for key in ['start', 'cp1', 'cp2', 'end']:
x, y = new_path[key]
new_path[key] = (
max(0, min(width, x)),
max(0, min(height, y))
)
modified_paths.append(new_path)
return modified_paths
def apply_aesthetic_refinement(self, paths: List[Dict], width: int, height: int):
"""Apply aesthetic refinements to improve visual quality"""
# Sort paths by position to create better layering
center_x, center_y = width / 2, height / 2
def distance_from_center(path):
start_x, start_y = path['start']
return math.sqrt((start_x - center_x)**2 + (start_y - center_y)**2)
# Sort by distance from center (background to foreground)
paths.sort(key=distance_from_center, reverse=True)
# Adjust opacity based on layering
for i, path in enumerate(paths):
# Paths closer to center (foreground) should be more opaque
layer_factor = 1.0 - (i / len(paths)) * 0.3
path['opacity'] = min(0.9, path['opacity'] * layer_factor)
return paths
def paths_to_svg(self, paths: List[Dict], width: int, height: int):
"""Convert optimized paths to SVG format"""
dwg = svgwrite.Drawing(size=(width, height))
dwg.add(dwg.rect(insert=(0, 0), size=(width, height), fill='white'))
for path in paths:
start_x, start_y = path['start']
cp1_x, cp1_y = path['cp1']
cp2_x, cp2_y = path['cp2']
end_x, end_y = path['end']
# Create Bezier curve path
path_data = f"M {start_x},{start_y} C {cp1_x},{cp1_y} {cp2_x},{cp2_y} {end_x},{end_y}"
color = path['color']
stroke_color = f"rgb({color[0]},{color[1]},{color[2]})"
dwg.add(dwg.path(
d=path_data,
stroke=stroke_color,
stroke_width=path['stroke_width'],
stroke_opacity=path['opacity'],
fill='none',
stroke_linecap='round',
stroke_linejoin='round'
))
return dwg.tostring()
def svg_to_pil_image(self, svg_content: str, width: int, height: int):
"""Convert SVG content to PIL Image"""
try:
import cairosvg
# Convert SVG to PNG bytes
png_bytes = cairosvg.svg2png(
bytestring=svg_content.encode('utf-8'),
output_width=width,
output_height=height
)
# Convert to PIL Image
image = Image.open(io.BytesIO(png_bytes)).convert('RGB')
return image
except ImportError:
print("cairosvg not available, creating simple image representation")
# Fallback: create a simple image with text
image = Image.new('RGB', (width, height), 'white')
return image
except Exception as e:
print(f"Error converting SVG to image: {e}")
# Fallback: create a simple image
image = Image.new('RGB', (width, height), 'white')
return image
def create_fallback_svg(self, prompt: str, width: int, height: int):
"""Create simple fallback SVG"""
dwg = svgwrite.Drawing(size=(width, height))
dwg.add(dwg.rect(insert=(0, 0), size=(width, height), fill='white'))
# Simple centered text
dwg.add(dwg.text(
f"DiffSketcher\n{prompt[:30]}...",
insert=(width/2, height/2),
text_anchor="middle",
font_size="12px",
fill="black"
))
return dwg.tostring()