"""Slopscan HF Space: drag & drop or paste an image, get AI vs human.""" import glob import os import gradio as gr import timm import torch from huggingface_hub import hf_hub_download from PIL import Image from torchvision import transforms MODEL_ID = "Dominik72/Slopscan-Anime" BACKBONE = "convnextv2_base.fcmae_ft_in22k_in1k_384" IMG_SIZE = 1024 MEAN = [0.485, 0.456, 0.406] STD = [0.229, 0.224, 0.225] HERE = os.path.dirname(os.path.abspath(__file__)) EXAMPLES_DIR = os.path.join(HERE, "examples") class Slopscan: def __init__(self, model_path): self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu") ckpt = torch.load(model_path, map_location="cpu", weights_only=False) state = ckpt.get("state_dict") or ckpt.get("ema_state") or ckpt self.model = timm.create_model(BACKBONE, pretrained=False, num_classes=2) self.model.load_state_dict(state) self.model.to(self.device).eval() self.transform = transforms.Compose([ transforms.Resize(IMG_SIZE + 32), transforms.CenterCrop(IMG_SIZE), transforms.ToTensor(), transforms.Normalize(MEAN, STD), ]) @torch.no_grad() def classify(self, image): img = image.convert("RGB") if isinstance(image, Image.Image) else Image.open(image) x = self.transform(img).unsqueeze(0).to(self.device) return torch.softmax(self.model(x), dim=1)[0, 1].item() best_pt = os.path.join(HERE, "best.pt") model = Slopscan(best_pt if os.path.exists(best_pt) else hf_hub_download(MODEL_ID, "best.pt")) def run(image, threshold): if image is None: return None prob = model.classify(image) verdict = "AI generated" if prob >= threshold else "Human drawn" return {verdict: prob} EXAMPLES = sorted( p for ext in ("png", "jpg", "jpeg", "webp") for p in glob.glob(os.path.join(EXAMPLES_DIR, f"*.{ext}")) ) with gr.Blocks(title="Slopscan") as demo: gr.Markdown( "# Slopscan\n" "Detects if an anime image was generated by AI or drawn by a human using [Dominik72/Slopscan-Anime](https://huggingface.co/Dominik72/Slopscan-Anime)." ) with gr.Row(): with gr.Column(): image = gr.Image(label="Image", sources=["upload", "clipboard"], type="pil", height=420) threshold = gr.Slider(0.0, 1.0, value=0.42, step=0.01, label="Threshold", info="Lower flags more images as AI") with gr.Row(): clear = gr.Button("Clear") submit = gr.Button("Submit", variant="primary") with gr.Column(): result = gr.Label() if EXAMPLES: gr.Examples( examples=[[p] for p in EXAMPLES], example_labels=[os.path.basename(p) for p in EXAMPLES], inputs=[image], cache_examples=False, ) submit.click(run, inputs=[image, threshold], outputs=result) clear.click(lambda: (None, None), outputs=[image, result]) image.change(lambda: None, outputs=result) if __name__ == "__main__": demo.launch()