app.py

"""
Two‑stage AI‑image detector with visual explainability
──────────────────────────────────────────────────────
Stage‑1 : haywoodsloan/ai-image-detector-deploy (Swin‑V2)  → Real vs AI
          ⟳ Grad‑CAM overlay
Stage‑2 : SuSy.pt (torchscript ResNet)                      → Generator
          ⟳ Saliency‑grad overlay (Captum)
"""

# ───────────────────── Imports ────────────────────────────────────────
import torch, numpy as np, pandas as pd, matplotlib.pyplot as plt
from PIL import Image
from torchvision import transforms
from transformers import AutoImageProcessor, AutoModelForImageClassification
from torchcam.methods import GradCAM
from captum.attr import Saliency
from skimage.feature import graycomatrix, graycoprops
import gradio as gr

# ─────────────────── Runtime / models ────────────────────────────────
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
plt.set_loglevel("ERROR")

BIN_ID   = "haywoodsloan/ai-image-detector-deploy"
bin_proc = AutoImageProcessor.from_pretrained(BIN_ID)
bin_mod  = AutoModelForImageClassification.from_pretrained(BIN_ID).to(device).eval()
CAM_LAYER_BIN = "encoder.layers.3.blocks.1.layernorm_after"

susy_mod = torch.jit.load("SuSy.pt").to(device).eval()
CAM_LAYER_SUSY = "feature_extractor.resnet_model.layer4.1.relu"

GEN_CLASSES = ["Stable Diffusion 1.x", "DALL·E 3",
               "MJ V5/V6", "Stable Diffusion XL", "MJ V1/V2"]
PATCH, TOP = 224, 5

# ─────────────── Universal overlay helper ────────────────────────────
def overlay_explanation(model, model_inputs, target_layer, class_idx, base_img):
    """Return heat‑map PIL.Image blended on top of base_img."""
    is_script = isinstance(model, torch.jit.ScriptModule)

    # clone & ensure gradients
    if torch.is_tensor(model_inputs):
        forward_inputs = model_inputs.clone().detach().requires_grad_(True)
    else:
        forward_inputs = {
            k: v.clone().detach().requires_grad_(True)
            for k, v in model_inputs.items()
        }

    if is_script:
        model.zero_grad(set_to_none=True)
        sal   = Saliency(model)
        grads = sal.attribute(forward_inputs, target=class_idx).abs().mean(1, keepdim=True)
        mask  = grads.squeeze().detach().cpu().numpy()

    else:
        mods = dict(model.named_modules())
        tgt  = mods.get(target_layer) or next(m for n, m in mods.items() if n.endswith(target_layer))
        cam  = GradCAM(model, target_layer=tgt)

        with torch.enable_grad():
            outputs = (model(forward_inputs) if torch.is_tensor(forward_inputs)
                       else model(**forward_inputs))
            logits  = outputs.logits if hasattr(outputs, "logits") else outputs

            # torchcam 0.7 → scores=…, earlier → logits=…
            try:
                cam_result = cam(class_idx, scores=logits)
            except TypeError:
                cam_result = cam(class_idx, logits=logits)

            mask = cam_result[0].detach().cpu().numpy()

        # clean up handles
        if hasattr(cam, "remove_hooks"):
            cam.remove_hooks()
        elif hasattr(cam, "clear_hooks"):
            cam.clear_hooks()

    mask = (mask - mask.min()) / (mask.max() - mask.min() + 1e-6)
    heat = Image.fromarray((plt.cm.jet(mask)[:, :, :3] * 255).astype(np.uint8))\
               .resize(base_img.size, Image.BICUBIC)
    return Image.blend(base_img.convert("RGBA"), heat.convert("RGBA"), alpha=0.45)

# ───────────── SuSy patch‑ranking helper ──────────────────────────────
to_tensor = transforms.ToTensor()
to_gray   = transforms.Compose([transforms.PILToTensor(), transforms.Grayscale()])

def susy_predict(img: Image.Image):
    w, h   = img.size
    npx, npy = max(1, w // PATCH), max(1, h // PATCH)
    patches  = np.zeros((npx * npy, PATCH, PATCH, 3), dtype=np.uint8)

    for i in range(npx):
        for j in range(npy):
            x, y = i * PATCH, j * PATCH
            patches[i*npy+j] = np.array(img.crop((x, y, x+PATCH, y+PATCH)).resize((PATCH, PATCH)))

    contrasts = []
    for p in patches:
        g = to_gray(Image.fromarray(p)).squeeze(0).numpy()
        glcm = graycomatrix(g, [5], [0], 256, symmetric=True, normed=True)
        contrasts.append(graycoprops(glcm, "contrast")[0, 0])

    idx   = np.argsort(contrasts)[::-1][:TOP]
    tens  = torch.from_numpy(patches[idx].transpose(0, 3, 1, 2)).float() / 255.0
    with torch.no_grad():
        probs = susy_mod(tens.to(device)).softmax(-1).mean(0).cpu().numpy()[1:]
    return dict(zip(GEN_CLASSES, probs))

# ───────────────────── Pipeline ───────────────────────────────────────
def pipeline(img_arr):
    img = Image.fromarray(img_arr) if isinstance(img_arr, np.ndarray) else img_arr
    heatmaps = []

    # Stage‑1 classification (no grad)
    with torch.no_grad():
        inp_bin  = bin_proc(images=img, return_tensors="pt").to(device)
        logits   = bin_mod(**inp_bin).logits.softmax(-1)[0]

    ai_conf, real_conf = logits
    winner_idx = 0 if ai_conf >= real_conf else 1

    # Stage‑1 heat‑map
    inp_bin_heat = {k: v.clone().detach().requires_grad_(True) for k, v in inp_bin.items()}
    heatmaps.append(
        overlay_explanation(bin_mod, inp_bin_heat, CAM_LAYER_BIN, winner_idx, img)
    )

    verdict = f"Authentic ({real_conf*100:.1f} %)"
    bar_df, show_bar = None, False

    # Stage‑2 if AI
    if ai_conf > real_conf:
        verdict = f"AI‑generated ({ai_conf*100:.1f} %)"
        gen_probs = susy_predict(img)
        bar_df    = pd.DataFrame({"class": gen_probs.keys(), "prob": gen_probs.values()})
        show_bar  = True

        with torch.no_grad():
            susy_in = to_tensor(img.resize((224, 224))).unsqueeze(0).to(device)
            g_idx   = susy_mod(susy_in)[0, 1:].argmax().item() + 1

        heatmaps.append(
            overlay_explanation(susy_mod, susy_in, CAM_LAYER_SUSY, g_idx, img)
        )

    return verdict, gr.update(value=bar_df, visible=show_bar), heatmaps

# ───────────────────────── UI ─────────────────────────────────────────
with gr.Blocks(theme=gr.themes.Soft()) as demo:
    gr.Markdown("## 🖼️ Two‑Stage AI Fake Detector — Explained with Heat‑maps")
    with gr.Row():
        img_in = gr.Image(type="numpy", label="Upload image")
        btn    = gr.Button("Detect")

    txt_out = gr.Textbox(label="Verdict", interactive=False)
    bar_out = gr.BarPlot(x="class", y="prob", title="Likely generator",
                         y_label="probability", visible=False)
    gal_out = gr.Gallery(label="Heat‑maps", columns=2, height=320)

    btn.click(pipeline, inputs=img_in, outputs=[txt_out, bar_out, gal_out])

demo.launch()