Update processing/two_stage/two_stage_processor.py

processing/two_stage/two_stage_processor.py

@@ -1,658 +1,346 @@
 #!/usr/bin/env python3
 """
-Two-Stage Green-Screen Processing System  ✅ 2025-08-29
-Stage 1: Original → keyed background (auto-selected colour)
-Stage 2: Keyed video → final composite  (hybrid chroma + segmentation rescue)
-
-Aligned with current project layout:
-  * uses helpers from utils.cv_processing (segment_person_hq, refine_mask_hq)
-  * safe local create_video_writer (no core.app dependency)
-  * cancel support via stop_event
-  * progress_callback(pct, desc)
-  * fully self-contained – just drop in and import TwoStageProcessor
-
-Additional safety:
-  * Ensures MatAnyone receives a valid first-frame mask (bootstraps the session
-    with the first SAM2 mask). This prevents "First frame arrived without a mask"
-    warnings and shape mismatches inside the stateful refiner.
-
-Quality profiles (set via env BFX_QUALITY = speed | balanced | max):
-  * refine cadence, spill suppression, edge softness
-  * hybrid matte mix (segmentation vs chroma), small dilate/blur on mask
-  * optional tiny background blur to hide seams on very flat backgrounds
+UI Components for BackgroundFX Pro (forced Two-Stage)
+-----------------------------------------------------
+* Pure layout (tiny wrapper to set env for quality)
+* All heavy logic stays in ui/callbacks.py
+* Two-stage mode is always active (checkbox removed)
 """
+
 from __future__ import annotations
+import os
+import gradio as gr
+
+from ui.callbacks import (
+    cb_load_models,
+    cb_process_video,
+    cb_cancel,
+    cb_status,
+    cb_clear,
+    cb_generate_bg,
+    cb_use_gen_bg,
+    cb_preset_bg_preview,
+)
+
+# Typography & UI polish: sharper text + cleaner cards
+CSS = """
+:root {
+  --radius: 16px;
+  --font-sans: 'Inter', system-ui, -apple-system, 'Segoe UI', Roboto,
+               'Helvetica Neue', Arial, sans-serif;
+}
 
-import cv2, numpy as np, os, gc, pickle, logging, tempfile, traceback, threading
-from pathlib import Path
-from typing import Optional, Dict, Any, Callable, Tuple, List
-
-from utils.cv_processing import segment_person_hq, refine_mask_hq
-
-# Project logger if available
-try:
-    from utils.logger import get_logger
-    logger = get_logger(__name__)
-except Exception:
-    logger = logging.getLogger(__name__)
-
-# ---------------------------------------------------------------------------
-# Local video-writer helper
-# ---------------------------------------------------------------------------
-def create_video_writer(output_path: str, fps: float, width: int, height: int, prefer_mp4: bool = True):
-    try:
-        ext = ".mp4" if prefer_mp4 else ".avi"
-        if not output_path:
-            output_path = tempfile.mktemp(suffix=ext)
-        else:
-            base, curr_ext = os.path.splitext(output_path)
-            if curr_ext.lower() not in [".mp4", ".avi", ".mov", ".mkv"]:
-                output_path = base + ext
-
-        fourcc = cv2.VideoWriter_fourcc(*("mp4v" if prefer_mp4 else "XVID"))
-        writer = cv2.VideoWriter(output_path, fourcc, float(fps), (int(width), int(height)))
-        if writer is None or not writer.isOpened():
-            alt_ext = ".avi" if prefer_mp4 else ".mp4"
-            alt_fourcc = cv2.VideoWriter_fourcc(*("XVID" if prefer_mp4 else "mp4v"))
-            alt_path = os.path.splitext(output_path)[0] + alt_ext
-            writer = cv2.VideoWriter(alt_path, alt_fourcc, float(fps), (int(width), int(height)))
-            if writer is None or not writer.isOpened():
-                return None, output_path
-            return writer, alt_path
-        return writer, output_path
-    except Exception as e:
-        logger.error(f"create_video_writer failed: {e}")
-        return None, output_path
-
-# ---------------------------------------------------------------------------
-# Key-colour helpers (fast, no external deps)
-# ---------------------------------------------------------------------------
-def _bgr_to_hsv_hue_deg(bgr: np.ndarray) -> np.ndarray:
-    hsv = cv2.cvtColor(bgr, cv2.COLOR_BGR2HSV)
-    # OpenCV H is 0-180; scale to degrees 0-360
-    return hsv[..., 0].astype(np.float32) * 2.0
-
-
-def _hue_distance(a_deg: float, b_deg: float) -> float:
-    """Circular distance on the hue wheel (degrees)."""
-    d = abs(a_deg - b_deg) % 360.0
-    return min(d, 360.0 - d)
-
-
-def _key_candidates_bgr() -> dict:
-    return {
-        "green":   {"bgr": np.array([  0,255,  0], dtype=np.uint8), "hue": 120.0},
-        "blue":    {"bgr": np.array([255,  0,  0], dtype=np.uint8), "hue": 240.0},
-        "cyan":    {"bgr": np.array([255,255,  0], dtype=np.uint8), "hue": 180.0},
-        "magenta": {"bgr": np.array([255,  0,255], dtype=np.uint8), "hue": 300.0},
-    }
-
-
-def _choose_best_key_color(frame_bgr: np.ndarray, mask_uint8: np.ndarray) -> dict:
-    """Pick the candidate colour farthest from the actor's dominant hues."""
-    try:
-        fg = frame_bgr[mask_uint8 > 127]
-        if fg.size < 1_000:
-            return _key_candidates_bgr()["green"]
-
-        fg_hue = _bgr_to_hsv_hue_deg(fg.reshape(-1, 1, 3)).reshape(-1)
-        hist, edges = np.histogram(fg_hue, bins=36, range=(0.0, 360.0))
-        top_idx = np.argsort(hist)[-3:]
-        top_hues = [(edges[i] + edges[i+1]) * 0.5 for i in top_idx]
-
-        best_name, best_score = None, -1.0
-        for name, info in _key_candidates_bgr().items():
-            cand_hue = info["hue"]
-            score = min(abs((cand_hue - th + 180) % 360 - 180) for th in top_hues)
-            if score > best_score:
-                best_name, best_score = name, score
-        return _key_candidates_bgr().get(best_name, _key_candidates_bgr()["green"])
-    except Exception:
-        return _key_candidates_bgr()["green"]
-
-
-# ---------------------------------------------------------------------------
-# Chroma presets
-# ---------------------------------------------------------------------------
-CHROMA_PRESETS: Dict[str, Dict[str, Any]] = {
-    'standard': {'key_color': [0,255,0], 'tolerance': 38, 'edge_softness': 2, 'spill_suppression': 0.35},
-    'studio':   {'key_color': [0,255,0], 'tolerance': 30, 'edge_softness': 1, 'spill_suppression': 0.45},
-    'outdoor':  {'key_color': [0,255,0], 'tolerance': 50, 'edge_softness': 3, 'spill_suppression': 0.25},
+/* Global crisp text */
+html, body, .gradio-container, .gradio-container * {
+  font-family: var(--font-sans) !important;
+  -webkit-font-smoothing: antialiased !important;
+  -moz-osx-font-smoothing: grayscale !important;
+  text-rendering: optimizeLegibility !important;
+  font-synthesis-weight: none;
 }
 
-# ---------------------------------------------------------------------------
-# Quality profiles (env: BFX_QUALITY = speed | balanced | max)
-# ---------------------------------------------------------------------------
-QUALITY_PROFILES: Dict[str, Dict[str, Any]] = {
-    "speed":    dict(refine_stride=4, spill=0.30, edge_softness=2, mix=0.60, dilate=0, blur=0, bg_sigma=0.0),
-    "balanced": dict(refine_stride=2, spill=0.40, edge_softness=2, mix=0.75, dilate=1, blur=1, bg_sigma=0.6),
-    "max":      dict(refine_stride=1, spill=0.45, edge_softness=3, mix=0.85, dilate=2, blur=2, bg_sigma=1.0),
+/* Headings tighter & bolder */
+.gradio-container h1, .gradio-container h2, .gradio-container h3 {
+  letter-spacing: -0.01em;
+  font-weight: 700;
 }
 
-# ---------------------------------------------------------------------------
-# Two-Stage Processor
-# ---------------------------------------------------------------------------
-class TwoStageProcessor:
-    def __init__(self, sam2_predictor=None, matanyone_model=None):
-        self.sam2  = self._unwrap_sam2(sam2_predictor)
-        self.matanyone = matanyone_model
-        self.mask_cache_dir = Path("/tmp/mask_cache")
-        self.mask_cache_dir.mkdir(parents=True, exist_ok=True)
-
-        # Internal flags/state
-        self._mat_bootstrapped = False
-        self._alpha_prev: Optional[np.ndarray] = None  # temporal smoothing
-
-        # Quality selection
-        qname = os.getenv("BFX_QUALITY", "balanced").strip().lower()
-        if qname not in QUALITY_PROFILES:
-            qname = "balanced"
-        self.quality = qname
-        self.q = QUALITY_PROFILES[qname]
-        logger.info(f"TwoStageProcessor quality='{self.quality}' ⇒ {self.q}")
-
-        logger.info(f"TwoStageProcessor init – SAM2: {self.sam2 is not None} | MatAnyOne: {self.matanyone is not None}")
-
-    # --------------------------- internal utils ---------------------------
-
-    def _unwrap_sam2(self, predictor):
-        """Unwrap the SAM2 predictor if needed."""
-        if predictor is None:
-            return None
-        if hasattr(predictor, 'sam_predictor'):
-            return predictor.sam_predictor
-        return predictor
-
-    def _get_mask(self, frame: np.ndarray) -> np.ndarray:
-        """Get segmentation mask using SAM2 (delegates to project helper)."""
-        if self.sam2 is None:
-            # Fallback: simple luminance threshold (kept to avoid breaking callers)
-            gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
-            _, mask = cv2.threshold(gray, 127, 255, cv2.THRESH_BINARY)
-            return mask
-
-        try:
-            mask = segment_person_hq(frame, self.sam2)
-            # segment_person_hq returns either uint8(0..255) or float(0..1) in most builds
-            return mask
-        except Exception as e:
-            logger.warning(f"SAM2 segmentation failed: {e}")
-            gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
-            _, mask = cv2.threshold(gray, 127, 255, cv2.THRESH_BINARY)
-            return mask
-
-    @staticmethod
-    def _to_binary_mask(mask: np.ndarray) -> Optional[np.ndarray]:
-        """Convert mask to uint8(0..255)."""
-        if mask is None:
-            return None
-        if mask.dtype == bool:
-            return mask.astype(np.uint8) * 255
-        if np.issubdtype(mask.dtype, np.floating):
-            m = np.clip(mask, 0.0, 1.0)
-            return (m * 255.0 + 0.5).astype(np.uint8)
-        return mask
-
-    @staticmethod
-    def _to_float01(mask: np.ndarray, h: int = None, w: int = None) -> Optional[np.ndarray]:
-        """Float [0,1] mask, optionally resized to (h,w)."""
-        if mask is None:
-            return None
-        m = mask.astype(np.float32)
-        if m.max() > 1.0:
-            m = m / 255.0
-        if h is not None and w is not None and (m.shape[0] != h or m.shape[1] != w):
-            m = cv2.resize(m, (w, h), interpolation=cv2.INTER_LINEAR)
-        return np.clip(m, 0.0, 1.0)
-
-    def _apply_greenscreen_hard(self, frame: np.ndarray, mask: np.ndarray, bg: np.ndarray) -> np.ndarray:
-        """Apply hard greenscreen compositing."""
-        mask_3ch = cv2.cvtColor(mask, cv2.COLOR_GRAY2BGR) if mask.ndim == 2 else mask
-        mask_norm = mask_3ch.astype(np.float32) / 255.0
-        result = frame * mask_norm + bg * (1 - mask_norm)
-        return result.astype(np.uint8)
-
-    # -------- improved spill suppression (preserves luminance & skin) --------
-    def _suppress_green_spill(self, frame: np.ndarray, amount: float = 0.35) -> np.ndarray:
-        """
-        Desaturate green dominance while preserving luminance and red skin hues.
-        amount: 0..1
-        """
-        b, g, r = cv2.split(frame.astype(np.float32))
-        green_dom = (g > r) & (g > b)
-        avg_rb = (r + b) * 0.5
-        g2 = np.where(green_dom, g*(1.0-amount) + avg_rb*amount, g)
-        skin = (r > g + 12)  # protect skin tones
-        g2 = np.where(skin, g, g2)
-        out = cv2.merge([np.clip(b,0,255), np.clip(g2,0,255), np.clip(r,0,255)]).astype(np.uint8)
-        return out
-
-    # -------- edge-aware alpha refinement (guided-like) --------
-    def _refine_alpha_edges(self, frame_bgr: np.ndarray, alpha_u8: np.ndarray, radius: int = 3, iters: int = 1) -> np.ndarray:
-        """
-        Fast, dependency-free, guided-like refinement on the alpha border.
-        Returns: uint8 alpha
-        """
-        a = alpha_u8.astype(np.uint8)
-        if radius <= 0:
-            return a
-
-        band = cv2.Canny(a, 32, 64)
-        if band.max() == 0:
-            return a
-
-        for _ in range(max(1, iters)):
-            a_blur = cv2.GaussianBlur(a, (radius*2+1, radius*2+1), 0)
-            b,g,r = cv2.split(frame_bgr.astype(np.float32))
-            green_dom = (g > r) & (g > b)
-            spill_mask = (green_dom & (a > 96) & (a < 224)).astype(np.uint8)*255
-            u = cv2.bitwise_or(band, spill_mask)
-            a = np.where(u>0, a_blur, a).astype(np.uint8)
-
-        return a
-
-    # -------- soft key based on chosen color (robust to blue/cyan/magenta) --------
-    def _soft_key_mask(self, frame_bgr: np.ndarray, key_bgr: np.ndarray, tol: int = 40) -> np.ndarray:
-        """
-        Soft chroma mask (uint8 0..255, 255=keep subject) using CbCr distance.
-        """
-        if key_bgr is None:
-            return np.full(frame_bgr.shape[:2], 255, np.uint8)
-
-        ycbcr = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2YCrCb).astype(np.float32)
-        kycbcr = cv2.cvtColor(key_bgr.reshape(1,1,3).astype(np.uint8), cv2.COLOR_BGR2YCrCb).astype(np.float32)[0,0]
-        d = np.linalg.norm((ycbcr[...,1:] - kycbcr[1:]), axis=-1)
-        d = cv2.GaussianBlur(d, (5,5), 0)
-        alpha = 255.0 * np.clip((d - tol) / (tol*1.7), 0.0, 1.0)  # far from key = keep (255)
-        return alpha.astype(np.uint8)
-
-    # --------------------- NEW: MatAnyone bootstrap ----------------------
-    def _bootstrap_matanyone_if_needed(self, frame_bgr: np.ndarray, coarse_mask: np.ndarray):
-        """
-        Call the MatAnyone session ONCE with the first coarse mask to initialize
-        its memory. This guarantees downstream calls never hit "first frame without a mask".
-        """
-        if self.matanyone is None or self._mat_bootstrapped:
-            return
-        try:
-            h, w = frame_bgr.shape[:2]
-            mask_f = self._to_float01(coarse_mask, h, w)
-            rgb = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2RGB)
-            _ = self.matanyone(rgb, mask_f)  # boot only; ignore returned alpha
-            self._mat_bootstrapped = True
-            logger.info("MatAnyone session bootstrapped with first-frame mask.")
-        except Exception as e:
-            logger.warning(f"MatAnyone bootstrap failed (continuing without): {e}")
-
-    # ---------------------------------------------------------------------
-    # Stage 1 – Original → keyed (green/blue/…)  -- chooses colour on 1st frame
-    # ---------------------------------------------------------------------
-    def stage1_extract_to_greenscreen(
-        self,
-        video_path: str,
-        output_path: str,
-        *,
-        key_color_mode: str = "auto",           # "auto" | "green" | "blue" | "cyan" | "magenta"
-        progress_callback: Optional[Callable[[float, str], None]] = None,
-        stop_event: Optional["threading.Event"] = None,
-    ) -> Tuple[Optional[dict], str]:
-
-        def _prog(p, d):
-            if progress_callback:
-                try:
-                    progress_callback(float(p), str(d))
-                except Exception:
-                    pass
-
-        try:
-            _prog(0.0, "Stage 1: opening video…")
-            cap = cv2.VideoCapture(video_path)
-            if not cap.isOpened():
-                return None, "Could not open input video"
-
-            fps   = cap.get(cv2.CAP_PROP_FPS) or 25.0
-            total = int(cap.get(cv2.CAP_PROP_FRAME_COUNT)) or 0
-            w     = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
-            h     = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
-
-            writer, out_path = create_video_writer(output_path, fps, w, h)
-            if writer is None:
-                cap.release()
-                return None, "Could not create output writer"
-
-            key_info: dict | None = None
-            chosen_bgr = np.array([0, 255, 0], np.uint8)  # default
-            probe_done = False
-            masks: List[np.ndarray] = []
-            frame_idx = 0
-
-            solid_bg = np.zeros((h, w, 3), np.uint8)  # overwritten per-frame
-
-            while True:
-                if stop_event and stop_event.is_set():
-                    _prog(1.0, "Stage 1: cancelled")
-                    break
-
-                ok, frame = cap.read()
-                if not ok:
-                    break
-
-                # --- SAM2 segmentation ---
-                mask = self._get_mask(frame)
-
-                # --- MatAnyone bootstrap exactly once (first frame) ---
-                if frame_idx == 0 and self.matanyone is not None:
-                    try:
-                        self._bootstrap_matanyone_if_needed(frame, mask)
-                    except Exception as e:
-                        logger.warning(f"Bootstrap error (non-fatal): {e}")
-
-                # --- Decide key colour once ---
-                if not probe_done:
-                    if key_color_mode.lower() == "auto":
-                        key_info = _choose_best_key_color(frame, self._to_binary_mask(mask))
-                        chosen_bgr = key_info["bgr"]
-                    else:
-                        cand = _key_candidates_bgr().get(key_color_mode.lower())
-                        if cand is not None:
-                            chosen_bgr = cand["bgr"]
-                    probe_done = True
-                    logger.info(f"[TwoStage] Using key colour: {key_color_mode} → {chosen_bgr.tolist()}")
-
-                # --- Optional refinement via MatAnyone (profile cadence) ---
-                stride = int(self.q.get("refine_stride", 3))
-                if self.matanyone and (frame_idx % max(1, stride) == 0):
-                    try:
-                        mask = refine_mask_hq(frame, mask, self.matanyone, fallback_enabled=True)
-                    except Exception as e:
-                        logger.warning(f"MatAnyOne refine fail f={frame_idx}: {e}")
-
-                # --- Composite onto solid key colour ---
-                solid_bg[:] = chosen_bgr
-                mask_u8 = self._to_binary_mask(mask)
-                gs = self._apply_greenscreen_hard(frame, mask_u8, solid_bg)
-                writer.write(gs)
-                masks.append(mask_u8)
-
-                frame_idx += 1
-                pct = 0.05 + 0.9 * (frame_idx / total) if total else min(0.95, 0.05 + frame_idx * 0.002)
-                _prog(pct, f"Stage 1: {frame_idx}/{total or '?'}")
-
-            cap.release()
-            writer.release()
-
-            # save mask cache
-            try:
-                cache_file = self.mask_cache_dir / (Path(out_path).stem + "_masks.pkl")
-                with open(cache_file, "wb") as f:
-                    pickle.dump(masks, f)
-            except Exception as e:
-                logger.warning(f"mask cache save fail: {e}")
-
-            _prog(1.0, "Stage 1: complete")
-            return (
-                {"path": out_path, "frames": frame_idx, "key_bgr": chosen_bgr.tolist()},
-                f"Green-screen video created ({frame_idx} frames)"
+/* Body copy slightly tighter */
+#hero .prose, .gr-markdown, .gr-text {
+  letter-spacing: -0.003em;
+}
+
+/* Card look */
+.card {
+  border-radius: var(--radius);
+  border: 1px solid rgba(0,0,0,.08);
+  padding: 16px;
+  background: linear-gradient(180deg, rgba(255,255,255,.94), rgba(248,250,252,.94));
+  box-shadow: 0 10px 30px rgba(0,0,0,.06);
+}
+
+.footer-note { opacity: 0.7; font-size: 12px; }
+.sm { font-size: 13px; opacity: 0.85; }
+#statusbox { min-height: 120px; }
+.preview-img { border-radius: var(--radius); border: 1px solid rgba(0,0,0,.08); }
+
+/* Buttons get a tiny weight bump for clarity */
+button, .gr-button { font-weight: 600; }
+
+/* Inline Quality select between buttons */
+.inline-quality .wrap-inner { min-width: 170px; }
+"""
+
+# Keep in sync with utils/cv_processing.PROFESSIONAL_BACKGROUNDS
+_BG_CHOICES = [
+    "minimalist",
+    "office_modern",
+    "studio_blue",
+    "studio_green",
+    "warm_gradient",
+    "tech_dark",
+]
+PRO_IMAGE_CHOICES = ["minimalist", "office_modern", "studio_blue", "studio_green"]
+GRADIENT_COLOR_CHOICES = ["warm_gradient", "tech_dark"]
+
+
+def create_interface() -> gr.Blocks:
+    with gr.Blocks(
+        title="🎬 BackgroundFX Pro",
+        css=CSS,
+        analytics_enabled=False,
+        theme=gr.themes.Soft()
+    ) as demo:
+
+        # ------------------------------------------------------------------
+        # HERO
+        # ------------------------------------------------------------------
+        with gr.Row(elem_id="hero"):
+            gr.Markdown(
+                "## 🎬 BackgroundFX Pro (CSP-Safe)\n"
+                "Replace your video background with cinema-quality AI matting. "
+                "Built for Hugging Face Spaces CSP.\n\n"
+                "_Tip: press **Load Models** once after the Space spins up._"
             )
 
-        except Exception as e:
-            logger.error(f"Stage 1 error: {e}\n{traceback.format_exc()}")
-            return None, f"Stage 1 failed: {e}"
-
-    # ---------------------------------------------------------------------
-    # Stage 2 – keyed video → final composite  (hybrid matte)
-    # ---------------------------------------------------------------------
-    def stage2_greenscreen_to_final(
-        self,
-        gs_path: str,
-        background: np.ndarray | str,
-        output_path: str,
-        *,
-        chroma_settings: Optional[Dict[str, Any]] = None,
-        progress_callback: Optional[Callable[[float, str], None]] = None,
-        stop_event: Optional["threading.Event"] = None,
-        key_bgr: Optional[np.ndarray] = None,   # pass chosen key color
-    ) -> Tuple[Optional[str], str]:
-
-        def _prog(p, d):
-            if progress_callback:
-                try:
-                    progress_callback(float(p), str(d))
-                except Exception:
-                    pass
-
-        try:
-            _prog(0.0, "Stage 2: opening keyed video…")
-            cap = cv2.VideoCapture(gs_path)
-            if not cap.isOpened():
-                return None, "Could not open keyed video"
-
-            fps   = cap.get(cv2.CAP_PROP_FPS) or 25.0
-            total = int(cap.get(cv2.CAP_PROP_FRAME_COUNT)) or 0
-            w     = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
-            h     = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
-
-            # Load or prepare background
-            if isinstance(background, str):
-                bg = cv2.imread(background)
-                if bg is None:
-                    cap.release()
-                    return None, "Could not load background image"
-                bg = cv2.resize(bg, (w, h))
-            else:
-                bg = cv2.resize(background, (w, h))
-
-            # Optional tiny BG blur per profile to hide seams on flat BGs
-            sigma = float(self.q.get("bg_sigma", 0.0))
-            if sigma > 0:
-                bg = cv2.GaussianBlur(bg, (0, 0), sigmaX=sigma, sigmaY=sigma)
-
-            writer, out_path = create_video_writer(output_path, fps, w, h)
-            if writer is None:
-                cap.release()
-                return None, "Could not create output writer"
-
-            # Load cached masks if available
-            masks = None
-            try:
-                cache_file = self.mask_cache_dir / (Path(gs_path).stem + "_masks.pkl")
-                if cache_file.exists():
-                    with open(cache_file, "rb") as f:
-                        masks = pickle.load(f)
-                    logger.info(f"Loaded {len(masks)} cached masks")
-            except Exception as e:
-                logger.warning(f"Could not load mask cache: {e}")
-
-            # Get chroma settings and override with profile
-            settings = chroma_settings or CHROMA_PRESETS.get('standard', {})
-            tolerance = int(settings.get('tolerance', 38))  # keep user tolerance
-            edge_softness = int(self.q.get('edge_softness', settings.get('edge_softness', 2)))
-            spill_suppression = float(self.q.get('spill', settings.get('spill_suppression', 0.35)))
-
-            # If caller didn't pass key_bgr, try preset or default green
-            if key_bgr is None:
-                key_bgr = np.array(settings.get('key_color', [0,255,0]), dtype=np.uint8)
-
-            self._alpha_prev = None  # reset temporal smoothing per render
-
-            frame_idx = 0
-            while True:
-                if stop_event and stop_event.is_set():
-                    _prog(1.0, "Stage 2: cancelled")
-                    break
-
-                ok, frame = cap.read()
-                if not ok:
-                    break
-
-                # Apply chroma keying with optional mask assistance
-                if masks and frame_idx < len(masks):
-                    mask = masks[frame_idx]
-                    final_frame = self._hybrid_composite(
-                        frame, bg, mask,
-                        tolerance=tolerance,
-                        edge_softness=edge_softness,
-                        spill_suppression=spill_suppression,
-                        key_bgr=key_bgr
+        # ------------------------------------------------------------------
+        # TAB – Quick Start
+        # ------------------------------------------------------------------
+        with gr.Tab("🏁 Quick Start"):
+
+            with gr.Row():
+                # ── Left column ────────────────────────────────────────────
+                with gr.Column(scale=1):
+                    video = gr.Video(label="Upload Video", interactive=True)
+
+                    # Hidden: effective preset key (still used by callbacks / defaults)
+                    bg_style = gr.Dropdown(
+                        label="Background Style (hidden)",
+                        choices=_BG_CHOICES,
+                        value="minimalist",
+                        visible=False,
                     )
-                else:
-                    # Pure chroma key
-                    final_frame = self._chroma_key_composite(
-                        frame, bg,
-                        tolerance=tolerance,
-                        edge_softness=edge_softness,
-                        spill_suppression=spill_suppression,
-                        key_bgr=key_bgr
+
+                    # =======================
+                    # Background Source block
+                    # =======================
+                    gr.Markdown("### 🖼️ Background Source")
+
+                    bg_method = gr.Radio(
+                        label="Choose method",
+                        choices=[
+                            "Upload file",
+                            "Pre-loaded professional images",
+                            "Pre-loaded Gradients / Colors",
+                            "AI generated background",
+                        ],
+                        value="Upload file",
                     )
 
-                writer.write(final_frame)
-                frame_idx += 1
-                pct = 0.05 + 0.9 * (frame_idx / total) if total else min(0.95, 0.05 + frame_idx * 0.002)
-                _prog(pct, f"Stage 2: {frame_idx}/{total or '?'}")
-
-            cap.release()
-            writer.release()
-
-            _prog(1.0, "Stage 2: complete")
-            return out_path, f"Final composite created ({frame_idx} frames)"
-
-        except Exception as e:
-            logger.error(f"Stage 2 error: {e}\n{traceback.format_exc()}")
-            return None, f"Stage 2 failed: {e}"
-
-    # ---------------- chroma + hybrid compositors (polished) ----------------
-    def _chroma_key_composite(self, frame, bg, *, tolerance=38, edge_softness=2, spill_suppression=0.35, key_bgr: Optional[np.ndarray] = None):
-        """Apply chroma key compositing with soft color distance + edge refinement."""
-        # 1) spill first
-        if spill_suppression > 0:
-            frame = self._suppress_green_spill(frame, spill_suppression)
-
-        # 2) build alpha
-        if key_bgr is not None:
-            alpha = self._soft_key_mask(frame, key_bgr, tol=int(tolerance))
-        else:
-            # Fallback: HSV green range
-            hsv = cv2.cvtColor(frame, cv2.COLOR_BGR2HSV)
-            lower_green = np.array([40, 40, 40])
-            upper_green = np.array([80, 255, 255])
-            alpha = cv2.bitwise_not(cv2.inRange(hsv, lower_green, upper_green))
-
-        # 3) soft edges + refinement
-        if edge_softness > 0:
-            k = edge_softness * 2 + 1
-            alpha = cv2.GaussianBlur(alpha, (k, k), 0)
-        alpha = self._refine_alpha_edges(frame, alpha, radius=max(1, edge_softness), iters=1)
-
-        # 4) temporal smoothing
-        if self._alpha_prev is not None and self._alpha_prev.shape == alpha.shape:
-            alpha = cv2.addWeighted(alpha, 0.75, self._alpha_prev, 0.25, 0)
-        self._alpha_prev = alpha
-
-        # 5) composite
-        mask_3ch = cv2.cvtColor(alpha, cv2.COLOR_GRAY2BGR).astype(np.float32) / 255.0
-        out = frame.astype(np.float32) * mask_3ch + bg.astype(np.float32) * (1.0 - mask_3ch)
-        return np.clip(out, 0, 255).astype(np.uint8)
-
-    def _hybrid_composite(self, frame, bg, mask, *, tolerance=38, edge_softness=2, spill_suppression=0.35, key_bgr: Optional[np.ndarray] = None):
-        """Apply hybrid compositing using both chroma key and cached mask, with profile controls."""
-        chroma_result = self._chroma_key_composite(
-            frame, bg,
-            tolerance=tolerance,
-            edge_softness=edge_softness,
-            spill_suppression=spill_suppression,
-            key_bgr=key_bgr
-        )
-        if mask is None:
-            return chroma_result
-
-        # profile-driven dilate/feather on cached mask to close pinholes + soften edges
-        m = mask
-        d = int(self.q.get("dilate", 0))
-        if d > 0:
-            k = 2*d + 1
-            se = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (k, k))
-            m = cv2.dilate(m, se, iterations=1)
-        b = int(self.q.get("blur", 0))
-        if b > 0:
-            m = cv2.GaussianBlur(m, (2*b+1, 2*b+1), 0)
-
-        m3 = cv2.cvtColor(m, cv2.COLOR_GRAY2BGR) if m.ndim == 2 else m
-        m3f = (m3.astype(np.float32) / 255.0)
-
-        seg_comp = frame.astype(np.float32) * m3f + bg.astype(np.float32) * (1.0 - m3f)
-
-        mix = float(self.q.get("mix", 0.7))  # weight towards segmentation on "max"
-        out = chroma_result.astype(np.float32) * (1.0 - mix) + seg_comp * mix
-        return np.clip(out, 0, 255).astype(np.uint8)
-
-    # ---------------------------------------------------------------------
-    # Combined pipeline
-    # ---------------------------------------------------------------------
-    def process_full_pipeline(
-        self,
-        video_path: str,
-        background: np.ndarray | str,
-        output_path: str,
-        *,
-        key_color_mode: str = "auto",
-        chroma_settings: Optional[Dict[str, Any]] = None,
-        progress_callback: Optional[Callable[[float, str], None]] = None,
-        stop_event: Optional["threading.Event"] = None,
-    ) -> Tuple[Optional[str], str]:
-        """Run both stages in sequence."""
-
-        def _combined_progress(pct, desc):
-            # Scale progress: Stage 1 is 0-50%, Stage 2 is 50-100%
-            if "Stage 1" in desc:
-                actual_pct = pct * 0.5
-            else:  # Stage 2
-                actual_pct = 0.5 + pct * 0.5
-
-            if progress_callback:
-                try:
-                    progress_callback(actual_pct, desc)
-                except Exception:
-                    pass
-
-        try:
-            # Reset per-video state
-            self._mat_bootstrapped = False
-            self._alpha_prev = None
-            if self.matanyone is not None and hasattr(self.matanyone, "reset"):
-                try:
-                    self.matanyone.reset()
-                except Exception:
-                    pass
-
-            # Stage 1
-            temp_gs_path = tempfile.mktemp(suffix="_greenscreen.mp4")
-            stage1_result, stage1_msg = self.stage1_extract_to_greenscreen(
-                video_path, temp_gs_path,
-                key_color_mode=key_color_mode,
-                progress_callback=_combined_progress,
-                stop_event=stop_event
+                    # a) Upload file option
+                    with gr.Group(visible=True) as grp_upload:
+                        custom_bg = gr.Image(
+                            label="Upload Background Image",
+                            interactive=True,
+                            type="filepath",  # returns file path
+                            elem_classes=["preview-img"]
+                        )
+
+                    # b) Pre-loaded professional images
+                    with gr.Group(visible=False) as grp_pro_images:
+                        pro_image_dd = gr.Dropdown(
+                            label="Professional Images",
+                            choices=PRO_IMAGE_CHOICES,
+                            value=PRO_IMAGE_CHOICES[0],
+                            info="Pre-defined photo-like backgrounds",
+                        )
+                        gr.Markdown(
+                            "<span class='sm'>Selecting a preset updates the preview below.</span>"
+                        )
+
+                    # c) Pre-loaded gradients & full colors
+                    with gr.Group(visible=False) as grp_gradients:
+                        gradient_dd = gr.Dropdown(
+                            label="Gradients & Full Colors",
+                            choices=GRADIENT_COLOR_CHOICES,
+                            value=GRADIENT_COLOR_CHOICES[0],
+                            info="Clean gradients or solid color styles",
+                        )
+                        gr.Markdown(
+                            "<span class='sm'>Selecting a preset updates the preview below.</span>"
+                        )
+
+                    # d) AI-generated background (inline, lightweight)
+                    with gr.Group(visible=False) as grp_ai:
+                        prompt = gr.Textbox(
+                            label="Describe vibe",
+                            value="modern office",
+                            info="e.g. 'soft sunset studio', 'cool tech dark', 'forest ambience'"
+                        )
+                        with gr.Row():
+                            gen_width  = gr.Slider(640, 1920, 1280, step=10, label="Width")
+                            gen_height = gr.Slider(360, 1080, 720,  step=10, label="Height")
+                        with gr.Row():
+                            bokeh    = gr.Slider(0, 30, 8,    step=1,    label="Bokeh Blur")
+                            vignette = gr.Slider(0, 0.6, 0.15, step=0.01, label="Vignette")
+                            contrast = gr.Slider(0.8, 1.4, 1.05, step=0.01, label="Contrast")
+                        with gr.Row():
+                            btn_gen_bg_inline = gr.Button("✨ Generate Background", variant="primary")
+                            use_gen_as_custom_inline = gr.Button("📌 Use as Custom Background", variant="secondary")
+                        gen_preview = gr.Image(
+                            label="Generated Background",
+                            interactive=False,
+                            elem_classes=["preview-img"]
+                        )
+                        gen_path = gr.Textbox(label="Saved Path", interactive=False)
+
+                    # ── Advanced options accordion ───────────────────────
+                    with gr.Accordion("Advanced", open=False):
+                        chroma_preset = gr.Dropdown(
+                            label="Chroma Preset",
+                            choices=["standard"],    # can add 'studio', 'outdoor' later
+                            value="standard"
+                        )
+                        key_color_mode = gr.Dropdown(
+                            label="Key-Colour Mode",
+                            choices=["auto", "green", "blue", "cyan", "magenta"],
+                            value="auto",
+                            info="Auto picks a colour far from your clothes; override if needed."
+                        )
+                        preview_mask = gr.Checkbox(
+                            label="Preview Mask only (mute audio)",
+                            value=False
+                        )
+                        preview_greenscreen = gr.Checkbox(
+                            label="Preview Green-screen only (mute audio)",
+                            value=False
+                        )
+
+                    # ── Controls row: Load → Quality → Process / Cancel ──
+                    with gr.Row():
+                        btn_load = gr.Button("🔄 Load Models", variant="secondary")
+
+                        quality = gr.Dropdown(
+                            label="Quality",
+                            choices=["speed", "balanced", "max"],
+                            value=os.getenv("BFX_QUALITY", "balanced"),
+                            info="Speed = fastest; Max = best edges & spill control.",
+                            elem_classes=["inline-quality"],
+                        )
+
+                        btn_run = gr.Button("🎬 Process Video", variant="primary")
+                        btn_cancel = gr.Button("⏹️ Cancel", variant="secondary")
+
+                # ── Right column ──────────────────────────────────────────
+                with gr.Column(scale=1):
+                    out_video = gr.Video(label="Processed Output", interactive=False)
+                    statusbox = gr.Textbox(label="Status", lines=8, elem_id="statusbox")
+                    with gr.Row():
+                        btn_refresh = gr.Button("🔍 Refresh Status", variant="secondary")
+                        btn_clear   = gr.Button("🧹 Clear",           variant="secondary")
+
+        # ------------------------------------------------------------------
+        # TAB – Status & settings
+        # ------------------------------------------------------------------
+        with gr.Tab("📈 Status & Settings"):
+            with gr.Row():
+                with gr.Column(scale=1, elem_classes=["card"]):
+                    model_status = gr.JSON(label="Model Status")
+                with gr.Column(scale=1, elem_classes=["card"]):
+                    cache_status = gr.JSON(label="Cache / System Status")
+
+            gr.Markdown(
+                "<div class='footer-note'>If models fail to load, fallbacks keep the UI responsive. "
+                "Check the runtime log for details.</div>"
             )
-            if stage1_result is None:
-                return None, stage1_msg
-
-            # Stage 2 (pass through chosen key color)
-            key_bgr = np.array(stage1_result.get("key_bgr", [0,255,0]), dtype=np.uint8)
-            final_path, stage2_msg = self.stage2_greenscreen_to_final(
-                stage1_result["path"], background, output_path,
-                chroma_settings=chroma_settings,
-                progress_callback=_combined_progress,
-                stop_event=stop_event,
-                key_bgr=key_bgr,
+
+        # ------------------------------------------------------------------
+        # Callback wiring
+        # ------------------------------------------------------------------
+
+        # Toggle which background sub-section is visible
+        def _toggle_bg_sections(choice: str):
+            return (
+                gr.update(visible=(choice == "Upload file")),
+                gr.update(visible=(choice == "Pre-loaded professional images")),
+                gr.update(visible=(choice == "Pre-loaded Gradients / Colors")),
+                gr.update(visible=(choice == "AI generated background")),
             )
 
-            # Clean up temp file
-            try:
-                os.remove(temp_gs_path)
-            except Exception:
-                pass
+        bg_method.change(
+            _toggle_bg_sections,
+            inputs=[bg_method],
+            outputs=[grp_upload, grp_pro_images, grp_gradients, grp_ai],
+        )
+
+        # Load models
+        btn_load.click(cb_load_models, outputs=statusbox)
+
+        # Tiny wrapper to set env for quality before calling the existing callback
+        def _run_with_quality(video_pth, bg_style_val, custom_bg_pth,
+                              use_two_stage_state, chroma_p, key_mode,
+                              prev_mask, prev_gs, quality_val):
+            os.environ["BFX_QUALITY"] = (quality_val or "balanced")
+            return cb_process_video(
+                video_pth, bg_style_val, custom_bg_pth,
+                use_two_stage_state, chroma_p, key_mode, prev_mask, prev_gs
+            )
+
+        # Always two-stage: pass use_two_stage=True to callback via State
+        btn_run.click(
+            _run_with_quality,
+            inputs=[
+                video,
+                bg_style,
+                custom_bg,
+                gr.State(value=True),  # Always two-stage
+                chroma_preset,
+                key_color_mode,
+                preview_mask,
+                preview_greenscreen,
+                quality,               # <-- the new control
+            ],
+            outputs=[out_video, statusbox],
+        )
+
+        # Cancel / Status / Clear
+        btn_cancel.click(cb_cancel, outputs=statusbox)
+        btn_refresh.click(cb_status, outputs=[model_status, cache_status])
+
+        btn_clear.click(
+            cb_clear,
+            outputs=[out_video, statusbox, gen_preview, gen_path, custom_bg]
+        )
 
-            return final_path, stage2_msg
+        # Preloaded presets → update preview (write into custom_bg)
+        pro_image_dd.change(
+            cb_preset_bg_preview,
+            inputs=[pro_image_dd],
+            outputs=[custom_bg],
+        )
+        gradient_dd.change(
+            cb_preset_bg_preview,
+            inputs=[gradient_dd],
+            outputs=[custom_bg],
+        )
+
+        # AI background generation (inline)
+        btn_gen_bg_inline.click(
+            cb_generate_bg,
+            inputs=[prompt, gen_width, gen_height, bokeh, vignette, contrast],
+            outputs=[gen_preview, gen_path],
+        )
+        use_gen_as_custom_inline.click(
+            cb_use_gen_bg,
+            inputs=[gen_path],
+            outputs=[custom_bg],
+        )
+
+        # Initialize with a default preset preview on load
+        demo.load(
+            cb_preset_bg_preview,
+            inputs=[bg_style],
+            outputs=[custom_bg]
+        )
 
-        except Exception as e:
-            logger.error(f"Full pipeline error: {e}\n{traceback.format_exc()}")
-            return None, f"Pipeline failed: {e}"
+    return demo