import gradio as gr from diffusers import StableDiffusionImg2ImgPipeline import torch from PIL import Image import io import random # ---- STYLE PROMPTS ---- STYLE_PROMPTS = { "no style change": "", "tattoo": "blackwork tattoo style, bold clean lines, vector look, high contrast", "street art": "street art mural style, graffiti, spray paint, bold outlines, urban, vector look", "pop art": "pop art style, halftone dots, bold primary colors, clean outlines", "minimalism": "minimalist vector illustration, simple flat colors, lots of negative space", "geometric": "geometric vector art, sharp shapes, clean lines, abstract patterns", "y2k": "Y2K digital art style, chrome, gradients, early 2000s web aesthetics, shiny, vector look", "abstract": "abstract vector art, bold shapes, experimental, high contrast", "trippy": "psychedelic trippy art, warped shapes, vivid colors, surreal, vector look", "retro": "retro vintage poster style, muted colors, worn texture, clean outlines", "grunge": "grunge poster art, distressed textures, rough edges, dark tones" } TARGET_MAX_SIDE = 6000 # pixels (good for shirts, posters, etc.) DPI = 350 # ---- LOAD PIPELINE ---- MODEL_ID = "runwayml/stable-diffusion-v1-5" device = "cuda" if torch.cuda.is_available() else "cpu" pipe = StableDiffusionImg2ImgPipeline.from_pretrained( MODEL_ID, torch_dtype=torch.float16 if device == "cuda" else torch.float32 ).to(device) def resize_for_model(image: Image.Image, max_side: int = 1024) -> Image.Image: w, h = image.size scale = max_side / max(w, h) if scale < 1: new_size = (int(w * scale), int(h * scale)) return image.resize(new_size, Image.LANCZOS) return image def upscale_to_target(image: Image.Image, target_max_side: int = TARGET_MAX_SIDE) -> Image.Image: w, h = image.size max_side = max(w, h) if max_side >= target_max_side: return image scale = target_max_side / max_side new_size = (int(w * scale), int(h * scale)) return image.resize(new_size, Image.LANCZOS) def generate_pod_art(input_image, style_name, num_variants=3, strength=0.7, guidance_scale=7.5): if input_image is None: return [] base_image = input_image.convert("RGB") # Prepare for model model_input = resize_for_model(base_image, max_side=1024) style_prompt = STYLE_PROMPTS.get(style_name, "") if style_name == "no style change": # Just try to keep content; no heavy style prompt = "high resolution, clean, vector-style illustration, print-on-demand ready" strength_val = 0.3 # stay close to original else: prompt = f"{style_prompt}, high resolution, vector-style illustration, print-on-demand ready" strength_val = strength outputs = [] for _ in range(num_variants): seed = random.randint(0, 2**32 - 1) generator = torch.Generator(device=device).manual_seed(seed) if device == "cuda": with torch.autocast(device): result = pipe( prompt=prompt, image=model_input, strength=strength_val, guidance_scale=guidance_scale, num_inference_steps=30, generator=generator ) else: with torch.no_grad(): result = pipe( prompt=prompt, image=model_input, strength=strength_val, guidance_scale=guidance_scale, num_inference_steps=30, generator=generator ) out_image = result.images[0] # Upscale for POD out_image = upscale_to_target(out_image, target_max_side=TARGET_MAX_SIDE) # Set DPI metadata when saving to bytes buf = io.BytesIO() out_image.save(buf, format="PNG", dpi=(DPI, DPI)) buf.seek(0) pil_out = Image.open(buf) outputs.append(pil_out) return outputs def process(input_image, style): images = generate_pod_art(input_image, style) return images with gr.Blocks() as demo: gr.Markdown("# POD-Ready Art Generator\nUpload an image, pick a style, get 3 high-res 350 DPI PNGs for print-on-demand.") with gr.Row(): with gr.Column(): input_img = gr.Image(label="Input Image", type="pil") style_dropdown = gr.Dropdown( label="Style", choices=list(STYLE_PROMPTS.keys()), value="no style change" ) generate_btn = gr.Button("Generate POD Art") with gr.Column(): gallery = gr.Gallery( label="Generated Images", show_label=True, columns=3, height=400, preview=True ) generate_btn.click( fn=process, inputs=[input_img, style_dropdown], outputs=[gallery] ) if __name__ == "__main__": demo.launch()