import gradio as gr import torch from transformers import Sam3Processor, Sam3Model from PIL import Image import numpy as np import matplotlib # Scegli il device device = "cuda" if torch.cuda.is_available() else "cpu" print("Using device:", device) # Carichiamo modello e processor UNA SOLA VOLTA (global) model = Sam3Model.from_pretrained("facebook/sam3").to(device) processor = Sam3Processor.from_pretrained("facebook/sam3") def overlay_masks(image: Image.Image, masks: torch.Tensor) -> Image.Image: """Applica le maschere colorate sopra l'immagine.""" image = image.convert("RGBA") masks_np = 255 * masks.cpu().numpy().astype(np.uint8) n_masks = masks_np.shape[0] if n_masks == 0: return image cmap = matplotlib.colormaps.get_cmap("rainbow").resampled(n_masks) colors = [ tuple(int(c * 255) for c in cmap(i)[:3]) for i in range(n_masks) ] for mask, color in zip(masks_np, colors): mask_img = Image.fromarray(mask) overlay = Image.new("RGBA", image.size, color + (0,)) # 0.4 = trasparenza 40% alpha = mask_img.point(lambda v: int(v * 0.4)) overlay.putalpha(alpha) image = Image.alpha_composite(image, overlay) return image def segment(image, text_prompt): """Funzione chiamata da Gradio: prende immagine + testo e ritorna immagine con maschere.""" if image is None or not text_prompt: return image # Pre-process inputs = processor( images=image, text=text_prompt, return_tensors="pt" ).to(device) # Forward with torch.no_grad(): outputs = model(**inputs) # Post-process: maschere alla risoluzione originale results = processor.post_process_instance_segmentation( outputs, threshold=0.5, mask_threshold=0.5, target_sizes=inputs["original_sizes"].tolist() )[0] masks = results["masks"] # [N, H, W] print("Found", len(masks), "objects for prompt:", text_prompt) if len(masks) == 0: # Nessun oggetto trovato → ritorno immagine originale return image # Applica overlay out_img = overlay_masks(image, masks) return out_img # Interfaccia Gradio demo = gr.Interface( fn=segment, inputs=[ gr.Image(type="pil", label="Immagine"), gr.Textbox(label="Prompt (es. 'cat', 'dog', 'ear')", value="cat"), ], outputs=gr.Image(type="pil", label="Risultato"), title="SAM3 – Promptable Concept Segmentation", description="Carica un'immagine, scrivi un concetto (es. 'cat', 'dog', 'ear') e SAM3 segmenterà tutte le istanze trovate." ) if __name__ == "__main__": demo.launch()