Upload agents/modality_detector.py with huggingface_hub

agents/modality_detector.py

@@ -113,6 +113,51 @@ def detect_modality(img: Image.Image) -> ModalityResult:
         scores["PORTRAIT_MODE"] = portrait_score
         indicators["portrait_detected"] = True
     
+    # ── Macro/DSLR shallow DoF detection ─────────────────────────────
+    # Macro photos have: extreme sharpness ratio (center vs edge), very high
+    # bimodal ratio, large uniform blur region, and Bayer pattern present.
+    # Key difference from portrait mode: sharpness ratio is much higher (>5)
+    # because macro lenses produce more extreme DoF than phone portrait mode.
+    
+    # Compute center-vs-edge sharpness ratio
+    ch, cw = gray.shape
+    center_region = sharpness[ch//4:3*ch//4, cw//4:3*cw//4]
+    edge_region = np.concatenate([
+        sharpness[:ch//4, :].ravel(),
+        sharpness[3*ch//4:, :].ravel(),
+        sharpness[:, :cw//4].ravel(),
+        sharpness[:, 3*cw//4:].ravel(),
+    ])
+    center_sharp = float(np.percentile(center_region, 90))
+    edge_sharp = float(np.mean(edge_region))
+    sharpness_ratio = center_sharp / (edge_sharp + 1e-9)
+    
+    # Background uniformity: how uniform is the blurred region's color?
+    blur_region_pixels = gray[blur_region] if np.any(blur_region) else np.array([128])
+    bg_color_std = float(np.std(blur_region_pixels))
+    
+    indicators["sharpness_ratio"] = round(sharpness_ratio, 2)
+    indicators["bg_color_std"] = round(bg_color_std, 2)
+    
+    macro_score = 0.0
+    if has_detail and sharpness_ratio > 3.0:
+        macro_score += 0.25  # Extreme center/edge sharpness difference
+    if has_detail and blur_frac > 0.35:
+        macro_score += 0.2   # Large blur region
+    if has_detail and bimodal_ratio > 1.5:
+        macro_score += 0.2   # Strong bimodal sharpness
+    if has_detail and bg_color_std < 40:
+        macro_score += 0.15  # Uniform background color (bokeh'd)
+    
+    # Macro requires Bayer OR high EXIF evidence (Unsplash strips some Bayer)
+    # Allow macro without Bayer if other signals are very strong
+    if macro_score >= 0.6:
+        scores["MACRO_DSLR"] = macro_score
+        indicators["macro_detected"] = True
+    elif macro_score >= 0.4 and has_bayer:
+        scores["MACRO_DSLR"] = macro_score
+        indicators["macro_detected"] = True
+    
     # ── Screenshot detection ──────────────────────────────────────────
     edge_mag = np.hypot(sobel(gray, 0), sobel(gray, 1))
     strong = edge_mag > np.percentile(edge_mag, 95)
@@ -192,8 +237,12 @@ def detect_modality(img: Image.Image) -> ModalityResult:
         modality = max(scores, key=scores.get)
         conf = min(1.0, scores[modality])
     
-    # Portrait mode always wins when detected (it's the most specific modality)
-    if scores.get("PORTRAIT_MODE", 0) > 0.3:
+    # Macro DSLR wins over portrait mode when detected (more specific)
+    if scores.get("MACRO_DSLR", 0) >= 0.4:
+        modality = "MACRO_DSLR"
+        conf = min(1.0, scores["MACRO_DSLR"])
+    # Portrait mode wins when detected and no macro
+    elif scores.get("PORTRAIT_MODE", 0) > 0.3:
         modality = "PORTRAIT_MODE"
         conf = min(1.0, scores["PORTRAIT_MODE"])
     
@@ -204,12 +253,17 @@ def detect_modality(img: Image.Image) -> ModalityResult:
         indicators["safety_override"] = "Low-detail image — suppression disabled"
     
     # SAFETY GUARD 2: No Bayer pattern = not from a real camera sensor.
-    # AI-generated images lack Bayer demosaicing traces.
-    # Never apply camera-related suppression without Bayer evidence.
+    # Exception: MACRO_DSLR with very strong signals can bypass this
+    # (Unsplash CDN processing can strip Bayer traces from real DSLR photos)
     if not has_bayer and modality in ("PORTRAIT_MODE", "SMARTPHONE", "MESSAGING"):
         modality = "UNKNOWN"
         conf = 0.2
         indicators["safety_override"] = f"No Bayer CFA pattern (margin={bayer_margin:.3f}) — not from a real camera sensor. All suppression disabled."
+    elif not has_bayer and modality == "MACRO_DSLR" and scores.get("MACRO_DSLR", 0) < 0.6:
+        # Weak macro signal without Bayer — don't trust it
+        modality = "UNKNOWN"
+        conf = 0.2
+        indicators["safety_override"] = f"Macro signal weak + no Bayer — suppression disabled"
     
     # ═══ BUILD ADJUSTMENTS ════════════════════════════════════════════
     
@@ -222,13 +276,46 @@ def detect_modality(img: Image.Image) -> ModalityResult:
             adjustments[k] = min(adjustments.get(k, 1.0), v)
         indicators["dual_modality"] = "PORTRAIT_MODE + MESSAGING"
     
+    # Merge messaging adjustments when macro + messaging both detected
+    if modality == "MACRO_DSLR" and scores.get("MESSAGING", 0) > 0.15:
+        msg_adj = _get_modality_adjustments("MESSAGING")
+        for k, v in msg_adj.items():
+            adjustments[k] = min(adjustments.get(k, 1.0), v)
+        indicators["dual_modality"] = "MACRO_DSLR + MESSAGING"
+    
     indicators["modality_scores"] = {k: round(v, 3) for k, v in scores.items()}
     
     return ModalityResult(modality, round(conf, 3), indicators, adjustments)
 
 
 def _get_modality_adjustments(modality: str) -> dict:
-    if modality == "PORTRAIT_MODE":
+    if modality == "MACRO_DSLR":
+        return {
+            # Extreme shallow DoF creates uniform bokeh → high autocorrelation
+            "Autocorrelation Peak": 0.1,
+            "Texture Repetition": 0.1,
+            "DoF Consistency": 0.1,
+            # JPEG delivery pipeline removes sensor traces
+            "Bayer CFA Pattern": 0.3,
+            "CFA Nyquist": 0.3,
+            "PRNU Uniformity": 0.2,
+            "Demosaic Interpolation": 0.4,
+            # Bimodal content (sharp subject + smooth bokeh) creates extreme kurtosis
+            # This is physics, not AI sharpening. Suppress the ceiling trigger.
+            "DCT Kurtosis": 0.1,
+            "Wavelet Kurtosis": 0.1,
+            "Spectral Slope 1/f²": 0.4,
+            "Spectral Symmetry": 0.3,
+            "Phase Coherence": 0.4,
+            # Noise inconsistency between sharp subject and bokeh
+            "Noise Spatial Frequency": 0.2,
+            "Poisson-Gaussian Model": 0.2,
+            "HF Noise Structure": 0.3,
+            "Pixel Response Linearity": 0.4,
+            "Saturation Clipping": 0.4,
+            "VAE Patch Boundaries": 0.3,
+        }
+    elif modality == "PORTRAIT_MODE":
         return {
             "Autocorrelation Peak": 0.1,
             "Texture Repetition": 0.1,