= num_samples:
break
except Exception as e:
print(f" Candidate {i} error: {e}")
continue
except Exception as e:
print(f"Generation Error: {e}")
import traceback
traceback.print_exc()
if progress_callback:
progress_callback(0.95, f"Generated {len(all_candidates)} valid candidates")
return all_candidates
@spaces.GPU(duration=120)
def gradio_text_to_svg(text_description, model_size, num_candidates, temperature, top_p, top_k,
repetition_penalty, max_length, progress=gr.Progress()):
"""Gradio interface - text-to-svg with model selection"""
if not text_description or text_description.strip() == "":
return 'Please enter a description
', "", f"Ready (no model loaded)"
print("\n" + "="*60)
print(f"[TASK] text-to-svg")
print(f"[MODEL] Requested: {model_size}")
print(f"[INPUT] {text_description[:100]}{'...' if len(text_description) > 100 else ''}")
print(f"[PARAMS] candidates={num_candidates}, temp={temperature}, top_p={top_p}, top_k={top_k}, rep_penalty={repetition_penalty}, max_length={max_length}")
print("="*60)
progress(0, "Initializing...")
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
start_time = time.time()
# Ensure model is loaded
progress(0.02, f"Loading {model_size} model (first time may take a while)...")
ensure_model_loaded(model_size)
model_status = f"✅ Using {model_size} model"
progress(0.05, "Model ready, starting generation...")
subtype = detect_text_subtype(text_description)
print(f"[SUBTYPE] Detected: {subtype}")
progress(0.08, f"Detected: {subtype}")
inputs = prepare_inputs("text-to-svg", text_description.strip())
def update_progress(val, msg):
progress(val, msg)
all_candidates = generate_candidates(
inputs, "text-to-svg", subtype,
temperature, top_p, int(top_k), repetition_penalty,
int(max_length), int(num_candidates),
progress_callback=update_progress
)
elapsed = time.time() - start_time
print(f"[RESULT] Generated {len(all_candidates)} valid candidates in {elapsed:.2f}s")
if not all_candidates:
print("[WARNING] No valid SVG generated")
return (
'No valid SVG generated. Try different parameters or rephrase your prompt.
',
f"",
model_status
)
svg_codes = []
for i, cand in enumerate(all_candidates):
svg_codes.append(f"\n{cand['svg']}")
combined_svg = "\n\n".join(svg_codes)
gallery_html = create_gallery_html(all_candidates)
progress(1.0, f"Done! {len(all_candidates)} candidates in {elapsed:.1f}s")
print(f"[COMPLETE] text-to-svg finished\n")
return gallery_html, combined_svg, model_status
@spaces.GPU(duration=120)
def gradio_image_to_svg(image, model_size, num_candidates, temperature, top_p, top_k, repetition_penalty,
max_length, replace_background, progress=gr.Progress()):
"""Gradio interface - image-to-svg with model selection"""
if image is None:
return (
'Please upload an image
',
"",
None,
f"Ready (no model loaded)"
)
print("\n" + "="*60)
print(f"[TASK] image-to-svg")
print(f"[MODEL] Requested: {model_size}")
print(f"[INPUT] Image size: {image.size if hasattr(image, 'size') else 'unknown'}, mode: {image.mode if hasattr(image, 'mode') else 'unknown'}")
print(f"[PARAMS] candidates={num_candidates}, temp={temperature}, top_p={top_p}, top_k={top_k}, rep_penalty={repetition_penalty}, max_length={max_length}, replace_bg={replace_background}")
print("="*60)
progress(0, "Initializing...")
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
start_time = time.time()
# Ensure model is loaded
progress(0.02, f"Loading {model_size} model (first time may take a while)...")
ensure_model_loaded(model_size)
model_status = f"✅ Using {model_size} model"
progress(0.05, "Processing input image...")
img_processed, was_modified = preprocess_image_for_svg(
image,
replace_background=replace_background,
target_size=TARGET_IMAGE_SIZE
)
if was_modified:
print("[PREPROCESS] Background processed/replaced")
progress(0.08, "Background processed")
with tempfile.NamedTemporaryFile(suffix='.png', delete=False) as tmp_file:
img_processed.save(tmp_file.name, format='PNG', quality=100)
tmp_path = tmp_file.name
try:
progress(0.1, "Preparing model inputs...")
inputs = prepare_inputs("image-to-svg", tmp_path)
def update_progress(val, msg):
progress(val, msg)
all_candidates = generate_candidates(
inputs, "image-to-svg", "image",
temperature, top_p, int(top_k), repetition_penalty,
int(max_length), int(num_candidates),
progress_callback=update_progress
)
elapsed = time.time() - start_time
print(f"[RESULT] Generated {len(all_candidates)} valid candidates in {elapsed:.2f}s")
if not all_candidates:
print("[WARNING] No valid SVG generated")
return (
'No valid SVG generated. Try adjusting parameters.
',
f"",
img_processed,
model_status
)
svg_codes = []
for i, cand in enumerate(all_candidates):
svg_codes.append(f"\n{cand['svg']}")
combined_svg = "\n\n".join(svg_codes)
gallery_html = create_gallery_html(all_candidates)
progress(1.0, f"Done! {len(all_candidates)} candidates in {elapsed:.1f}s")
print(f"[COMPLETE] image-to-svg finished\n")
return gallery_html, combined_svg, img_processed, model_status
finally:
if os.path.exists(tmp_path):
os.unlink(tmp_path)
def get_example_images():
"""Get example images from the examples directory"""
example_dir = "./examples"
example_images = []
if os.path.exists(example_dir):
for ext in SUPPORTED_FORMATS:
pattern = os.path.join(example_dir, f"*{ext}")
example_images.extend(glob.glob(pattern))
example_images.sort()
return example_images
def create_interface():
"""Create Gradio interface"""
# Simplified example prompts - 6 key examples
example_texts = [
"A red heart shape with smooth curved edges, centered.",
"Circular avatar: person with short black hair, round face with two dot eyes and small curved smile, wearing blue collar shirt. Minimal style.",
"Sunset beach: orange gradient sky at top, yellow semicircle sun on horizon, dark blue wavy ocean below, tan beach at bottom.",
"Cute cat: orange round head with two triangular ears, oval orange body, curved tail. Simple cartoon style with black outlines, sitting pose.",
"Simple black bird: oval body, small round head, pointed triangular beak facing right, triangular tail, two stick legs. Silhouette style.",
"Simple house: red triangular roof on top, beige rectangular wall, brown rectangular door in center, two small blue square windows. Green ground at bottom.",
]
example_images = get_example_images()
with gr.Blocks(title="OmniSVG Demo Page") as demo:
# Header
gr.HTML("""
""")
# GPU quota warning
gr.HTML("""
⚠️
GPU Quota Notice: Free users have 300 seconds/day of GPU time.
Each generation uses ~20-60s. Start with 1 candidate and lower max_length (512) to save quota!
🎲
Not satisfied with the result? This is normal! Just click generate again to re-roll. Each generation is different.
""")
with gr.Tabs():
# ==================== Text-to-SVG Tab ====================
with gr.TabItem("Text-to-SVG", id="text-tab"):
with gr.Row(equal_height=False):
with gr.Column(scale=1, min_width=300):
gr.Markdown("### Description")
text_input = gr.Textbox(
label="",
placeholder="Describe your SVG with geometric shapes and colors...\n\nExample: A red heart shape with smooth curved edges, centered.",
lines=4
)
with gr.Group(elem_classes=["settings-group"]):
gr.Markdown("### Settings")
# Model Selection
text_model_size = gr.Dropdown(
choices=AVAILABLE_MODEL_SIZES,
value=DEFAULT_MODEL_SIZE,
label="Model Size",
info="8B: Higher quality | 4B: Faster, less GPU"
)
text_num_candidates = gr.Slider(
minimum=1, maximum=MAX_NUM_CANDIDATES, value=DEFAULT_NUM_CANDIDATES, step=1,
label="Number of Candidates",
info="Start with 1 to save GPU quota"
)
# Max Length slider
text_max_length = gr.Slider(
minimum=MAX_LENGTH_MIN,
maximum=MAX_LENGTH_MAX,
value=MAX_LENGTH_DEFAULT,
step=64,
label="Max Length",
info="Lower = faster & saves GPU quota"
)
with gr.Accordion("Advanced Parameters", open=False):
text_temperature = gr.Slider(
minimum=0.1, maximum=1.0,
value=TASK_CONFIGS["text-to-svg-icon"].get("default_temperature", 0.5),
step=0.05,
label="Temperature",
info="Lower = more accurate, Higher = more creative"
)
text_top_p = gr.Slider(
minimum=0.5, maximum=1.0,
value=TASK_CONFIGS["text-to-svg-icon"].get("default_top_p", 0.90),
step=0.02,
label="Top-P"
)
text_top_k = gr.Slider(
minimum=10, maximum=100,
value=TASK_CONFIGS["text-to-svg-icon"].get("default_top_k", 60),
step=5,
label="Top-K"
)
text_rep_penalty = gr.Slider(
minimum=1.0, maximum=1.3,
value=TASK_CONFIGS["text-to-svg-icon"].get("default_repetition_penalty", 1.03),
step=0.01,
label="Repetition Penalty"
)
text_generate_btn = gr.Button(
"Generate SVG",
variant="primary",
size="lg",
elem_classes=["primary-btn"]
)
# Model status display
text_model_status = gr.Textbox(
label="Status",
value="Ready",
interactive=False
)
gr.Markdown("### Example Prompts")
gr.Examples(
examples=[[text] for text in example_texts],
inputs=[text_input],
label=""
)
with gr.Column(scale=2, min_width=500):
gr.Markdown("### Generated SVG")
text_gallery = gr.HTML(
value='Generated SVGs will appear here
🎲 Not satisfied? Click generate again to re-roll!
'
)
gr.Markdown("### SVG Code")
text_svg_output = gr.Code(label="", language="html", lines=10, elem_classes=["code-output"])
text_generate_btn.click(
fn=gradio_text_to_svg,
inputs=[text_input, text_model_size, text_num_candidates, text_temperature, text_top_p,
text_top_k, text_rep_penalty, text_max_length],
outputs=[text_gallery, text_svg_output, text_model_status],
queue=True
)
# ==================== Image-to-SVG Tab ====================
with gr.TabItem("Image-to-SVG", id="image-tab"):
gr.HTML(IMAGE_TIPS_HTML)
with gr.Row(equal_height=False):
with gr.Column(scale=1, min_width=300):
gr.Markdown("### Upload Image")
image_input = gr.Image(
label="Drag, upload, or Ctrl+V to paste",
type="pil",
image_mode="RGBA",
height=200,
sources=["upload", "clipboard"],
elem_classes=["input-image"]
)
with gr.Group(elem_classes=["settings-group"]):
gr.Markdown("### Settings")
# Model Selection
img_model_size = gr.Dropdown(
choices=AVAILABLE_MODEL_SIZES,
value=DEFAULT_MODEL_SIZE,
label="Model Size",
info="8B: Higher quality | 4B: Faster, less GPU"
)
img_num_candidates = gr.Slider(
minimum=1, maximum=MAX_NUM_CANDIDATES, value=DEFAULT_NUM_CANDIDATES, step=1,
label="Number of Candidates",
info="Start with 1 to save GPU quota"
)
img_replace_bg = gr.Checkbox(
label="Replace non-white background",
value=True
)
# Max Length slider
img_max_length = gr.Slider(
minimum=MAX_LENGTH_MIN,
maximum=MAX_LENGTH_MAX,
value=MAX_LENGTH_DEFAULT,
step=64,
label="Max Length",
info="Lower = faster & saves GPU quota"
)
with gr.Accordion("Advanced Parameters", open=False):
img_temperature = gr.Slider(
minimum=0.1, maximum=1.0,
value=TASK_CONFIGS["image-to-svg"].get("default_temperature", 0.3),
step=0.05,
label="Temperature"
)
img_top_p = gr.Slider(
minimum=0.5, maximum=1.0,
value=TASK_CONFIGS["image-to-svg"].get("default_top_p", 0.90),
step=0.02,
label="Top-P"
)
img_top_k = gr.Slider(
minimum=10, maximum=100,
value=TASK_CONFIGS["image-to-svg"].get("default_top_k", 50),
step=5,
label="Top-K"
)
img_rep_penalty = gr.Slider(
minimum=1.0, maximum=1.3,
value=TASK_CONFIGS["image-to-svg"].get("default_repetition_penalty", 1.05),
step=0.01,
label="Repetition Penalty"
)
image_generate_btn = gr.Button(
"Generate SVG",
variant="primary",
size="lg",
elem_classes=["primary-btn"]
)
# Model status display
img_model_status = gr.Textbox(
label="Status",
value="Ready",
interactive=False
)
if example_images:
gr.Markdown("### Examples")
gr.Examples(examples=example_images, inputs=[image_input], label="")
with gr.Column(scale=2, min_width=500):
gr.Markdown("### Processed Input")
image_processed = gr.Image(label="", type="pil", height=100)
gr.Markdown("### Generated SVG")
image_gallery = gr.HTML(
value='Generated SVGs will appear here
🎲 Not satisfied? Click generate again to re-roll!
'
)
gr.Markdown("### SVG Code")
image_svg_output = gr.Code(label="", language="html", lines=8, elem_classes=["code-output"])
image_generate_btn.click(
fn=gradio_image_to_svg,
inputs=[image_input, img_model_size, img_num_candidates, img_temperature, img_top_p,
img_top_k, img_rep_penalty, img_max_length, img_replace_bg],
outputs=[image_gallery, image_svg_output, image_processed, img_model_status],
queue=True
)
# Tips section at the bottom (not as a tab)
gr.HTML(TIPS_HTML_BOTTOM)
# Footer
gr.HTML(f"""
""")
return demo
if __name__ == "__main__":
os.environ["TOKENIZERS_PARALLELISM"] = "false"
args = parse_args()
print("="*60)
print("OmniSVG Demo Page - Gradio App (HuggingFace Space)")
print("="*60)
print(f"Available model sizes: {AVAILABLE_MODEL_SIZES}")
print(f"Default model size: {DEFAULT_MODEL_SIZE}")
print(f"Device: {device}")
print(f"Precision: {DTYPE}")
print("Models will be loaded on-demand when first used.")
print("="*60)
# Print loaded config values
print("\n[CONFIG] Shared settings:")
print(f" - TARGET_IMAGE_SIZE: {TARGET_IMAGE_SIZE}")
print(f" - RENDER_SIZE: {RENDER_SIZE}")
print(f" - BLACK_COLOR_TOKEN: {BLACK_COLOR_TOKEN}")
print(f" - MAX_LENGTH (default): {MAX_LENGTH_DEFAULT}")
print(f" - MAX_LENGTH_MIN: {MAX_LENGTH_MIN}")
print(f" - MAX_LENGTH_MAX: {MAX_LENGTH_MAX}")
print(f" - BOS_TOKEN_ID: {BOS_TOKEN_ID}")
print(f" - EOS_TOKEN_ID: {EOS_TOKEN_ID}")
print(f" - PAD_TOKEN_ID: {PAD_TOKEN_ID}")
print(f" - GPU Duration: 60s per call")
print("="*60)
demo = create_interface()
demo.queue(default_concurrency_limit=1, max_size=20)
demo.launch(
server_name=args.listen,
server_port=args.port,
share=args.share,
debug=args.debug,
)