setiastrosuitepro 1.6.1__py3-none-any.whl

setiastro/saspro/remove_stars.py

@@ -0,0 +1,1625 @@
+# pro/remove_stars.py
+from __future__ import annotations
+import os
+import platform
+import shutil
+import stat
+import tempfile
+import numpy as np
+
+from PyQt6.QtCore import QThread, pyqtSignal
+from PyQt6.QtWidgets import (
+    QInputDialog, QMessageBox, QFileDialog,
+    QDialog, QVBoxLayout, QTextEdit, QPushButton,
+    QLabel, QComboBox, QCheckBox, QSpinBox, QFormLayout, QDialogButtonBox, QWidget, QHBoxLayout
+)
+
+# use your legacy I/O functions (as requested)
+from setiastro.saspro.legacy.image_manager import save_image, load_image
+import glob
+try:
+    import cv2
+except Exception:
+    cv2 = None
+
+# Shared utilities
+from setiastro.saspro.widgets.image_utils import extract_mask_from_document as _active_mask_array_from_doc
+
+_MAD_NORM = 1.4826
+
+# --------- deterministic, invertible stretch used for StarNet ----------
+# ---------- Siril-like MTF (linked) pre-stretch for StarNet ----------
+def _robust_peak_sigma(gray: np.ndarray) -> tuple[float, float]:
+    gray = gray.astype(np.float32, copy=False)
+    med = float(np.median(gray))
+    mad = float(np.median(np.abs(gray - med)))
+    sigma = 1.4826 * mad if mad > 0 else float(gray.std())
+    # optional: refine "peak" as histogram mode around median
+    try:
+        hist, edges = np.histogram(gray, bins=2048, range=(gray.min(), gray.max()))
+        peak = float(0.5 * (edges[np.argmax(hist)] + edges[np.argmax(hist)+1]))
+    except Exception:
+        peak = med
+    return peak, max(sigma, 1e-8)
+
+def _mtf_apply(x: np.ndarray, shadows: float, midtones: float, highlights: float) -> np.ndarray:
+    # x in [0, +], returns [0..1]ish given s,h
+    s, m, h = float(shadows), float(midtones), float(highlights)
+    denom = max(h - s, 1e-8)
+    xp = (x - s) / denom
+    # clamp xp to avoid crazy values
+    xp = np.clip(xp, 0.0, 1.0)
+    num = (m - 1.0) * xp
+    den = ((2.0 * m - 1.0) * xp) - m
+    y = np.divide(num, den, out=np.zeros_like(xp, dtype=np.float32), where=np.abs(den) > 1e-12)
+    return np.clip(y, 0.0, 1.0).astype(np.float32, copy=False)
+
+def _mtf_inverse(y: np.ndarray, shadows: float, midtones: float, highlights: float) -> np.ndarray:
+    """
+    Pseudoinverse of MTF, matching Siril's MTF_pseudoinverse() implementation.
+
+    C reference:
+
+    float MTF_pseudoinverse(float y, struct mtf_params params) {
+        return ((((params.shadows + params.highlights) * params.midtones
+                - params.shadows) * y - params.shadows * params.midtones
+                + params.shadows)
+                / ((2 * params.midtones - 1) * y - params.midtones + 1));
+    }
+    """
+    s = float(shadows)
+    m = float(midtones)
+    h = float(highlights)
+
+    yp = np.clip(y.astype(np.float32, copy=False), 0.0, 1.0)
+
+    num = (((s + h) * m - s) * yp - s * m + s)
+    den = (2.0 * m - 1.0) * yp - m + 1.0
+
+    x = np.divide(
+        num,
+        den,
+        out=np.full_like(yp, s, dtype=np.float32),  # fallback ~shadows if denom≈0
+        where=np.abs(den) > 1e-12
+    )
+
+    # Clamp back into [s, h] and then [0,1] for safety
+    x = np.clip(x, s, h)
+    return np.clip(x, 0.0, 1.0).astype(np.float32, copy=False)
+
+def _mtf_params_linked(img_rgb01: np.ndarray, shadowclip_sigma: float = -2.8, targetbg: float = 0.25):
+    """
+    Compute linked (single) MTF parameters for RGB image in [0..1].
+    Returns dict(s=..., m=..., h=...).
+    """
+    # luminance proxy for stats
+    if img_rgb01.ndim == 2:
+        gray = img_rgb01
+    else:
+        gray = img_rgb01.mean(axis=2)
+    peak, sigma = _robust_peak_sigma(gray)
+    s = peak + shadowclip_sigma * sigma
+    # keep [0..1) with margin
+    s = float(np.clip(s, gray.min(), max(gray.max() - 1e-6, 0.0)))
+    h = 1.0  # Siril effectively normalizes to <=1 before 16-bit TIFF
+    # solve for midtones m so that mtf(xp(peak)) = targetbg
+    x = (peak - s) / max(h - s, 1e-8)
+    x = float(np.clip(x, 1e-6, 1.0 - 1e-6))
+    y = float(np.clip(targetbg, 1e-6, 1.0 - 1e-6))
+    denom = (2.0 * y * x) - y - x
+    m = (x * (y - 1.0)) / denom if abs(denom) > 1e-12 else 0.5
+    m = float(np.clip(m, 1e-4, 1.0 - 1e-4))
+    return {"s": s, "m": m, "h": h}
+
+def _apply_mtf_linked_rgb(img_rgb01: np.ndarray, p: dict) -> np.ndarray:
+    if img_rgb01.ndim == 2:
+        img_rgb01 = np.stack([img_rgb01]*3, axis=-1)
+    y = np.empty_like(img_rgb01, dtype=np.float32)
+    for c in range(3):
+        y[..., c] = _mtf_apply(img_rgb01[..., c], p["s"], p["m"], p["h"])
+    return np.clip(y, 0.0, 1.0)
+
+def _invert_mtf_linked_rgb(img_rgb01: np.ndarray, p: dict) -> np.ndarray:
+    y = np.empty_like(img_rgb01, dtype=np.float32)
+    for c in range(3):
+        y[..., c] = _mtf_inverse(img_rgb01[..., c], p["s"], p["m"], p["h"])
+    return y
+
+
+def _mtf_params_unlinked(img_rgb01: np.ndarray,
+                         shadows_clipping: float = -2.8,
+                         targetbg: float = 0.25) -> dict:
+    """
+    Siril-style per-channel MTF parameter estimation, matching
+    find_unlinked_midtones_balance_default() / find_unlinked_midtones_balance().
+
+    Works on float32 data assumed in [0,1].
+    Returns dict with arrays: {'s': (C,), 'm': (C,), 'h': (C,)}.
+    """
+    x = np.asarray(img_rgb01, dtype=np.float32)
+    # Force 3 channels internally (Siril expects 1 or 3; we always give it 3 here)
+    if x.ndim == 2:
+        x = np.stack([x] * 3, axis=-1)
+    elif x.ndim == 3 and x.shape[2] == 1:
+        x = np.repeat(x, 3, axis=2)
+
+    C = x.shape[2]
+    s = np.zeros(C, dtype=np.float32)
+    m = np.zeros(C, dtype=np.float32)
+    h = np.zeros(C, dtype=np.float32)
+
+    med = np.zeros(C, dtype=np.float32)
+    mad = np.zeros(C, dtype=np.float32)
+    inverted_flags = np.zeros(C, dtype=bool)
+
+    # --- stats per channel (Siril: median / normValue, mad / normValue * MAD_NORM) ---
+    # Here normValue == 1.0 because we're already in [0,1]
+    for c in range(C):
+        ch = x[..., c].astype(np.float32, copy=False)
+        med_c = float(np.median(ch))
+        mad_raw = float(np.median(np.abs(ch - med_c)))
+        mad_c = mad_raw * _MAD_NORM
+        if mad_c == 0.0:
+            mad_c = 0.001
+
+        med[c] = med_c
+        mad[c] = mad_c
+        if med_c > 0.5:
+            inverted_flags[c] = True
+
+    inverted_channels = int(inverted_flags.sum())
+
+    # --- Main branch (non-inverted dominant) ---
+    if inverted_channels < C:
+        for c in range(C):
+            median = float(med[c])
+            mad_c = float(mad[c])
+
+            c0 = median + shadows_clipping * mad_c
+            if c0 < 0.0:
+                c0 = 0.0
+            # Siril: m2 = median - c0; midtones = MTF(m2, target_bg, 0,1)
+            m2 = median - c0
+            mid = float(_mtf_scalar(m2, targetbg, 0.0, 1.0))
+
+            s[c] = c0
+            m[c] = mid
+            h[c] = 1.0
+
+    # --- Inverted channel branch ---
+    else:
+        for c in range(C):
+            median = float(med[c])
+            mad_c = float(mad[c])
+
+            c1 = median - shadows_clipping * mad_c
+            if c1 > 1.0:
+                c1 = 1.0
+            m2 = c1 - median
+            mid = 1.0 - float(_mtf_scalar(m2, targetbg, 0.0, 1.0))
+
+            s[c] = 0.0
+            m[c] = mid
+            h[c] = c1
+
+    return {"s": s, "m": m, "h": h}
+
+
+def _apply_mtf_unlinked_rgb(img_rgb01: np.ndarray, p: dict) -> np.ndarray:
+    """
+    Apply per-channel MTF exactly. p from _mtf_params_unlinked.
+    """
+    x = np.asarray(img_rgb01, dtype=np.float32)
+    if x.ndim == 2:
+        x = np.stack([x]*3, axis=-1)
+    elif x.ndim == 3 and x.shape[2] == 1:
+        x = np.repeat(x, 3, axis=2)
+
+    out = np.empty_like(x, dtype=np.float32)
+    for c in range(x.shape[2]):
+        out[..., c] = _mtf_apply(x[..., c], float(p["s"][c]), float(p["m"][c]), float(p["h"][c]))
+    return np.clip(out, 0.0, 1.0)
+
+
+def _invert_mtf_unlinked_rgb(img_rgb01: np.ndarray, p: dict) -> np.ndarray:
+    """
+    Exact analytic inverse per channel (uses same s/m/h arrays).
+    """
+    y = np.asarray(img_rgb01, dtype=np.float32)
+    if y.ndim == 2:
+        y = np.stack([y]*3, axis=-1)
+    elif y.ndim == 3 and y.shape[2] == 1:
+        y = np.repeat(y, 3, axis=2)
+
+    out = np.empty_like(y, dtype=np.float32)
+    for c in range(y.shape[2]):
+        out[..., c] = _mtf_inverse(y[..., c], float(p["s"][c]), float(p["m"][c]), float(p["h"][c]))
+    return np.clip(out, 0.0, 1.0)
+
+def _stat_stretch_rgb(img: np.ndarray,
+                      lo_pct: float = 0.25,
+                      hi_pct: float = 99.75) -> tuple[np.ndarray, dict]:
+    """
+    Make sure img is RGB float32 in [0,1], stretch each channel to [0,1]
+    using percentiles. Returns (stretched_img, params) where params can be
+    fed to _stat_unstretch_rgb() to invert exactly.
+    """
+    was_single = (img.ndim == 2) or (img.ndim == 3 and img.shape[2] == 1)
+    if was_single:
+        img = np.stack([img] * 3, axis=-1)
+
+    x = img.astype(np.float32, copy=False)
+    out = np.empty_like(x, dtype=np.float32)
+    lo_vals, hi_vals = [], []
+
+    for c in range(3):
+        ch = x[..., c]
+        lo = float(np.percentile(ch, lo_pct))
+        hi = float(np.percentile(ch, hi_pct))
+        if not np.isfinite(lo): lo = 0.0
+        if not np.isfinite(hi): hi = 1.0
+        if hi - lo < 1e-6:
+            hi = lo + 1e-6
+        lo_vals.append(lo); hi_vals.append(hi)
+        out[..., c] = (ch - lo) / (hi - lo)
+
+    out = np.clip(out, 0.0, 1.0)
+    params = {"lo": lo_vals, "hi": hi_vals, "was_single": was_single}
+    return out, params
+
+
+def _stat_unstretch_rgb(img: np.ndarray, params: dict) -> np.ndarray:
+    """
+    Inverse of _stat_stretch_rgb. Expects img RGB float32 [0,1].
+    """
+    lo = np.asarray(params["lo"], dtype=np.float32)
+    hi = np.asarray(params["hi"], dtype=np.float32)
+    out = img.astype(np.float32, copy=True)
+    for c in range(3):
+        out[..., c] = out[..., c] * (hi[c] - lo[c]) + lo[c]
+    out = np.clip(out, 0.0, 1.0)
+    if params.get("was_single", False):
+        out = out.mean(axis=2, keepdims=False)  # back to single channel if needed
+        # StarNet needs RGB during processing; we keep RGB after removal for consistency.
+        # If you want to return mono to the doc when the source was mono, do it at the very end.
+        out = np.stack([out] * 3, axis=-1)
+    return out
+
+def _mtf_scalar(x: float, m: float, lo: float = 0.0, hi: float = 1.0) -> float:
+    """
+    Scalar midtones transfer function matching the PixInsight / Siril spec.
+
+    For x in [lo, hi], rescale to [0,1] and apply:
+
+        M(x; m) = (m - 1) * xp / ((2*m - 1)*xp - m)
+
+    with the special cases x<=lo -> 0, x>=hi -> 1.
+    """
+    # clamp to the input domain
+    if x <= lo:
+        return 0.0
+    if x >= hi:
+        return 1.0
+
+    denom_range = hi - lo
+    if abs(denom_range) < 1e-12:
+        return 0.0
+
+    xp = (x - lo) / denom_range  # normalized x in [0,1]
+
+    num = (m - 1.0) * xp
+    den = (2.0 * m - 1.0) * xp - m
+
+    if abs(den) < 1e-12:
+        # the spec says M(m; m) = 0.5, but if we ever hit this numerically
+        # just return 0.5 as a safe fallback
+        return 0.5
+
+    y = num / den
+    # clamp to [0,1] as PI/Siril do
+    if y < 0.0:
+        y = 0.0
+    elif y > 1.0:
+        y = 1.0
+    return float(y)
+
+
+# ------------------------------------------------------------
+# Settings helper
+# ------------------------------------------------------------
+def _get_setting_any(settings, keys: tuple[str, ...], default: str = "") -> str:
+    if not settings:
+        return default
+    for k in keys:
+        try:
+            v = settings.value(k, "", type=str)
+        except Exception:
+            v = settings.value(k, "")
+        if isinstance(v, str) and v.strip():
+            return v.strip()
+    return default
+
+
+# ================== HEADLESS, ARRAY-IN → STARLESS-ARRAY-OUT ==================
+
+def starnet_starless_from_array(arr_rgb01: np.ndarray, settings, *, tmp_prefix="comet") -> np.ndarray:
+    """
+    Siril-style MTF round-trip for 32-bit data:
+
+      1) Normalize to [0,1] (preserving overall scale separately)
+      2) Compute unlinked MTF params per channel (Siril auto-stretch)
+      3) Apply unlinked MTF -> 16-bit TIFF for StarNet
+      4) StarNet -> read starless 16-bit TIFF
+      5) Apply per-channel MTF pseudoinverse with SAME params
+      6) Restore original scale if >1.0
+    """
+    import os
+    import platform
+    import subprocess
+    import numpy as np
+
+    # save_image / load_image / _get_setting_any assumed available
+    arr = np.asarray(arr_rgb01, dtype=np.float32)
+    was_single = (arr.ndim == 2) or (arr.ndim == 3 and arr.shape[2] == 1)
+
+    exe = _get_setting_any(settings, ("starnet/exe_path", "paths/starnet"), "")
+    if not exe or not os.path.exists(exe):
+        raise RuntimeError("StarNet executable not configured (settings 'paths/starnet').")
+
+    workdir = os.path.dirname(exe) or os.getcwd()
+    in_path  = os.path.join(workdir, f"{tmp_prefix}_in.tif")
+    out_path = os.path.join(workdir, f"{tmp_prefix}_out.tif")
+
+    # --- Normalize input shape and safe values ---
+    x = arr
+    if x.ndim == 2:
+        x = np.stack([x] * 3, axis=-1)
+    elif x.ndim == 3 and x.shape[2] == 1:
+        x = np.repeat(x, 3, axis=2)
+    x = np.nan_to_num(x, nan=0.0, posinf=0.0, neginf=0.0).astype(np.float32, copy=False)
+
+    # Preserve original numeric scale if users pass >1.0
+    xmax = float(np.max(x)) if x.size else 1.0
+    scale_factor = xmax if xmax > 1.01 else 1.0
+    xin = (x / scale_factor) if scale_factor > 1.0 else x
+    xin = np.clip(xin, 0.0, 1.0)
+
+    # --- Siril-style unlinked MTF params + pre-stretch ---
+    mtf_params = _mtf_params_unlinked(xin, shadows_clipping=-2.8, targetbg=0.25)
+    x_for_starnet = _apply_mtf_unlinked_rgb(xin, mtf_params).astype(np.float32, copy=False)
+
+    # --- Write 16-bit TIFF for StarNet ---
+    save_image(
+        x_for_starnet, in_path,
+        original_format="tif", bit_depth="16-bit",
+        original_header=None, is_mono=False, image_meta=None, file_meta=None
+    )
+
+    # --- Run StarNet ---
+    exe_name = os.path.basename(exe).lower()
+    if platform.system() in ("Windows", "Linux"):
+        cmd = [exe, in_path, out_path, "256"]
+    else:
+        cmd = [exe, "--input", in_path, "--output", out_path] if "starnet2" in exe_name else [exe, in_path, out_path]
+
+    rc = subprocess.call(cmd, cwd=workdir)
+    if rc != 0 or not os.path.exists(out_path):
+        _safe_rm(in_path); _safe_rm(out_path)
+        raise RuntimeError(f"StarNet failed rc={rc}")
+
+    starless_s, _, _, _ = load_image(out_path)
+    _safe_rm(in_path); _safe_rm(out_path)
+
+    if starless_s.ndim == 2:
+        starless_s = np.stack([starless_s] * 3, axis=-1)
+    elif starless_s.ndim == 3 and starless_s.shape[2] == 1:
+        starless_s = np.repeat(starless_s, 3, axis=2)
+    starless_s = np.clip(starless_s.astype(np.float32, copy=False), 0.0, 1.0)
+
+    # --- Apply Siril-style pseudoinverse MTF with SAME params ---
+    starless_lin01 = _invert_mtf_unlinked_rgb(starless_s, mtf_params)
+
+    # Restore original scale if we normalized earlier
+    if scale_factor > 1.0:
+        starless_lin01 *= scale_factor
+
+    result = np.clip(starless_lin01, 0.0, 1.0).astype(np.float32, copy=False)
+
+    # If the source was mono, return mono
+    if was_single and result.ndim == 3:
+        result = result.mean(axis=2)
+
+    return result
+
+
+def darkstar_starless_from_array(arr_rgb01: np.ndarray, settings, *, tmp_prefix="comet",
+                                 disable_gpu=False, mode="unscreen", stride=512) -> np.ndarray:
+    """
+    Save arr -> run DarkStar -> load starless -> return starless RGB float32 [0..1].
+    """
+    exe, base = _resolve_darkstar_exe(type("dummy", (), {"settings": settings}) )
+    if not exe or not base:
+        raise RuntimeError("Cosmic Clarity DarkStar executable not configured.")
+    arr = np.asarray(arr_rgb01, dtype=np.float32)
+    was_single = (arr.ndim == 2) or (arr.ndim == 3 and arr.shape[2] == 1)
+    input_dir  = os.path.join(base, "input")
+    output_dir = os.path.join(base, "output")
+    os.makedirs(input_dir, exist_ok=True)
+    os.makedirs(output_dir, exist_ok=True)
+    _purge_darkstar_io(base, prefix=None, clear_input=True, clear_output=True)
+
+    in_path = os.path.join(input_dir, f"{tmp_prefix}_in.tif")
+    save_image(
+        arr, in_path,
+        original_format="tif", bit_depth="32-bit floating point",
+        original_header=None, is_mono=was_single, image_meta=None, file_meta=None
+    )
+
+    args = []
+    if disable_gpu: args.append("--disable_gpu")
+    args += ["--star_removal_mode", mode, "--chunk_size", str(int(stride))]
+    import subprocess
+    rc = subprocess.call([exe] + args, cwd=output_dir)
+    if rc != 0:
+        _safe_rm(in_path); raise RuntimeError(f"DarkStar failed rc={rc}")
+
+    starless_path = os.path.join(output_dir, "imagetoremovestars_starless.tif")
+    starless, _, _, _ = load_image(starless_path)
+    if starless is None:
+        _safe_rm(in_path); raise RuntimeError("DarkStar produced no starless image.")
+    if starless.ndim == 2 or (starless.ndim == 3 and starless.shape[2] == 1):
+        starless = np.stack([starless] * 3, axis=-1)
+    starless = np.clip(starless.astype(np.float32, copy=False), 0.0, 1.0)
+
+    # If the source was mono, collapse back to single channel
+    if was_single and starless.ndim == 3:
+        starless = starless.mean(axis=2)
+
+    # cleanup typical outputs
+    _purge_darkstar_io(base, prefix="imagetoremovestars", clear_input=True, clear_output=True)
+    return starless
+
+
+# ------------------------------------------------------------
+# Public entry
+# ------------------------------------------------------------
+def remove_stars(main, target_doc=None):
+    # block interactive UI during/just-after a headless preset run
+    if getattr(main, "_remove_stars_headless_running", False):
+        return
+    if getattr(main, "_remove_stars_guard", False):
+        return    
+
+    tool, ok = QInputDialog.getItem(
+        main, "Select Star Removal Tool", "Choose a tool:",
+        ["StarNet", "CosmicClarityDarkStar"], 0, False
+    )
+    if not ok:
+        return
+
+    # explicit doc wins; otherwise fall back to _active_doc
+    doc = target_doc
+    if doc is None:
+        doc = getattr(main, "_active_doc", None)
+        if callable(doc):
+            doc = doc()
+
+    if doc is None or getattr(doc, "image", None) is None:
+        QMessageBox.warning(main, "No Image", "Please load an image before removing stars.")
+        return
+
+    if tool == "CosmicClarityDarkStar":
+        _run_darkstar(main, doc)
+    else:
+        _run_starnet(main, doc)
+
+
+
+
+def _first_nonzero_bp_per_channel(img3: np.ndarray) -> np.ndarray:
+    """Per-channel minimum positive sample (0 if none)."""
+    bps = np.zeros(3, dtype=np.float32)
+    for c in range(3):
+        ch = img3[..., c].reshape(-1)
+        pos = ch[ch > 0.0]
+        bps[c] = float(pos.min()) if pos.size else 0.0
+    return bps
+
+
+def _prepare_statstretch_input_for_starnet(img_rgb01: np.ndarray) -> tuple[np.ndarray, dict]:
+    """
+    Build the input to StarNet using your statistical stretch flow:
+      • record per-channel first-nonzero blackpoints
+      • subtract pedestals
+      • record per-channel medians
+      • unlinked statistical stretch to target 0.25
+    Returns: (stretched_for_starnet_01, meta_dict)
+    """
+    import numpy as np
+    from setiastro.saspro.imageops.stretch import stretch_color_image
+
+    x = np.asarray(img_rgb01, dtype=np.float32)
+    if x.ndim == 2:
+        x = np.stack([x]*3, axis=-1)
+    elif x.ndim == 3 and x.shape[2] == 1:
+        x = np.repeat(x, 3, axis=2)
+    x = np.nan_to_num(x, nan=0.0, posinf=0.0, neginf=0.0)
+    x = np.clip(x, 0.0, 1.0)
+
+    # per-channel pedestal
+    bp = _first_nonzero_bp_per_channel(x)
+    xin_ped = np.clip(x - bp.reshape((1, 1, 3)), 0.0, 1.0)
+
+    # per-channel medians (after pedestal removal)
+    m0 = np.array([float(np.median(xin_ped[..., c])) for c in range(3)], dtype=np.float32)
+
+    # unlinked stat-stretch to 0.25
+    x_for_starnet = stretch_color_image(
+        xin_ped, target_median=0.25, linked=False,
+        normalize=False, apply_curves=False, curves_boost=0.0
+    ).astype(np.float32, copy=False)
+
+    meta = {
+        "statstretch": True,
+        "bp": bp,              # pedestals we subtracted (in 0..1 domain)
+        "m0": m0,              # per-channel original medians (post-pedestal)
+    }
+    return x_for_starnet, meta
+
+
+def _inverse_statstretch_from_starless(starless_s01: np.ndarray, meta: dict) -> np.ndarray:
+    """
+    Inverse of the stat-stretch prep:
+      • per-channel stretch back to each original median m0[c]
+      • add back the saved pedestal bp[c]
+    Returns starless in 0..1 domain (float32).
+    """
+    import numpy as np
+    from setiastro.saspro.imageops.stretch import stretch_mono_image
+
+    s = np.asarray(starless_s01, dtype=np.float32)
+    if s.ndim == 2:
+        s = np.stack([s]*3, axis=-1)
+    elif s.ndim == 3 and s.shape[2] == 1:
+        s = np.repeat(s, 3, axis=2)
+    s = np.clip(s, 0.0, 1.0)
+
+    bp = np.asarray(meta.get("bp"), dtype=np.float32).reshape((1, 1, 3))
+    m0 = np.asarray(meta.get("m0"), dtype=np.float32)
+
+    out = np.empty_like(s, dtype=np.float32)
+    for c in range(3):
+        out[..., c] = stretch_mono_image(
+            s[..., c], target_median=float(m0[c]),
+            normalize=False, apply_curves=False, curves_boost=0.0
+        )
+
+    out = out + bp
+    return np.clip(out, 0.0, 1.0).astype(np.float32, copy=False)
+
+
+# ------------------------------------------------------------
+# StarNet (SASv2-like: 16-bit TIFF in StarNet folder)
+# ------------------------------------------------------------
+def _run_starnet(main, doc):
+    import os
+    import platform
+    import numpy as np
+    from PyQt6.QtWidgets import QFileDialog, QMessageBox
+
+    # --- Resolve StarNet exe, persist in settings
+    exe = _get_setting_any(getattr(main, "settings", None),
+                           ("starnet/exe_path", "paths/starnet"), "")
+    if not exe or not os.path.exists(exe):
+        exe_path, _ = QFileDialog.getOpenFileName(main, "Select StarNet Executable", "", "Executable Files (*)")
+        if not exe_path:
+            return
+        exe = exe_path
+        s = getattr(main, "settings", None)
+        if s:
+            s.setValue("starnet/exe_path", exe)
+            s.setValue("paths/starnet", exe)
+
+    if platform.system() in ("Darwin", "Linux"):
+        _ensure_exec_bit(exe)
+
+    sysname = platform.system()
+    if sysname not in ("Windows", "Darwin", "Linux"):
+        QMessageBox.critical(main, "Unsupported OS",
+                             f"The current operating system '{sysname}' is not supported.")
+        return
+
+    # --- Ask linearity (SASv2 behavior)
+    reply = QMessageBox.question(
+        main, "Image Linearity", "Is the current image linear?",
+        QMessageBox.StandardButton.Yes | QMessageBox.StandardButton.No,
+        QMessageBox.StandardButton.No
+    )
+    is_linear = (reply == QMessageBox.StandardButton.Yes)
+    did_stretch = is_linear 
+    try:
+        main._last_remove_stars_params = {
+            "engine": "StarNet",
+            "is_linear": bool(is_linear),
+            "did_stretch": bool(did_stretch),
+            "label": "Remove Stars (StarNet)",
+        }
+    except Exception:
+        pass
+    # 🔁 Record headless command for Replay Last
+    try:
+        main._last_headless_command = {
+            "command_id": "remove_stars",
+            "preset": {
+                "tool": "starnet",
+                "linear": bool(is_linear),
+            },
+        }
+        if hasattr(main, "_log"):
+            main._log(
+                f"[Replay] Recorded remove_stars (StarNet, linear="
+                f"{'yes' if is_linear else 'no'})"
+            )
+    except Exception:
+        pass    
+    # --- Ensure RGB float32 in safe range
+    src = np.asarray(doc.image)
+    if src.ndim == 2:
+        processing_image = np.stack([src]*3, axis=-1)
+    elif src.ndim == 3 and src.shape[2] == 1:
+        processing_image = np.repeat(src, 3, axis=2)
+    else:
+        processing_image = src
+    processing_image = np.nan_to_num(processing_image.astype(np.float32, copy=False),
+                                     nan=0.0, posinf=0.0, neginf=0.0)
+
+    # --- Scale normalization if >1.0 (same reason as before: 16-bit export safety)
+    scale_factor = float(np.max(processing_image))
+    if scale_factor > 1.0:
+        processing_norm = processing_image / scale_factor
+    else:
+        processing_norm = processing_image
+
+    # --- Build input/output paths
+    starnet_dir = os.path.dirname(exe) or os.getcwd()
+    input_image_path  = os.path.join(starnet_dir, "imagetoremovestars.tif")
+    output_image_path = os.path.join(starnet_dir, "starless.tif")
+
+    # --- Prepare input for StarNet (Siril-style MTF pre-stretch for linear data) ---
+    img_for_starnet = processing_norm
+    if is_linear:
+        # Siril-style unlinked MTF params from linear normalized image
+        mtf_params = _mtf_params_unlinked(processing_norm, shadows_clipping=-2.8, targetbg=0.25)
+        img_for_starnet = _apply_mtf_unlinked_rgb(processing_norm, mtf_params)
+
+        # 🔐 Stash EXACT params for inverse step later
+        try:
+            setattr(main, "_starnet_stat_meta", {
+                "scheme": "siril_mtf",
+                "s": np.asarray(mtf_params["s"], dtype=np.float32),
+                "m": np.asarray(mtf_params["m"], dtype=np.float32),
+                "h": np.asarray(mtf_params["h"], dtype=np.float32),
+                "scale": float(scale_factor),
+            })
+        except Exception:
+            pass
+    else:
+        # non-linear: do not try to invert any pre-stretch later
+        if hasattr(main, "_starnet_stat_meta"):
+            delattr(main, "_starnet_stat_meta")
+
+
+    # --- Write TIFF for StarNet
+    try:
+        save_image(img_for_starnet, input_image_path,
+                   original_format="tif", bit_depth="16-bit",
+                   original_header=None, is_mono=False, image_meta=None, file_meta=None)
+    except Exception as e:
+        QMessageBox.critical(main, "StarNet", f"Failed to write input TIFF:\n{e}")
+        return
+
+    # --- Launch StarNet in a worker (keeps your progress dialog)
+    exe_name = os.path.basename(exe).lower()
+    if sysname in ("Windows", "Linux"):
+        command = [exe, input_image_path, output_image_path, "256"]
+    else:  # macOS
+        if "starnet2" in exe_name:
+            command = [exe, "--input", input_image_path, "--output", output_image_path]
+        else:
+            command = [exe, input_image_path, output_image_path]
+
+    dlg = _ProcDialog(main, title="StarNet Progress")
+    thr = _ProcThread(command, cwd=starnet_dir)
+    thr.output_signal.connect(dlg.append_text)
+
+    # Capture everything we need in the closure for finish handler
+    thr.finished_signal.connect(
+        lambda rc, ds=did_stretch: _on_starnet_finished(
+            main, doc, rc, dlg, input_image_path, output_image_path, ds
+        )
+    )
+    dlg.cancel_button.clicked.connect(thr.cancel)
+
+    dlg.show()
+    thr.start()
+    dlg.exec()
+
+
+def _on_starnet_finished(main, doc, return_code, dialog, input_path, output_path, did_stretch):
+    import os
+    import numpy as np
+    from PyQt6.QtWidgets import QMessageBox
+    from setiastro.saspro.imageops.stretch import stretch_mono_image  # used for statistical inverse
+
+    def _first_nonzero_bp_per_channel(img3: np.ndarray) -> np.ndarray:
+        bps = np.zeros(3, dtype=np.float32)
+        for c in range(3):
+            ch = img3[..., c].reshape(-1)
+            pos = ch[ch > 0.0]
+            bps[c] = float(pos.min()) if pos.size else 0.0
+        return bps
+
+    dialog.append_text(f"\nProcess finished with return code {return_code}.\n")
+    if return_code != 0:
+        QMessageBox.critical(main, "StarNet Error", f"StarNet failed with return code {return_code}.")
+        _safe_rm(input_path); _safe_rm(output_path)
+        dialog.close()
+        return
+
+    if not os.path.exists(output_path):
+        QMessageBox.critical(main, "StarNet Error", "Starless image was not created.")
+        _safe_rm(input_path)
+        dialog.close()
+        return
+
+    dialog.append_text(f"Starless image found at {output_path}. Loading image...\n")
+    starless_rgb, _, _, _ = load_image(output_path)
+    _safe_rm(input_path); _safe_rm(output_path)
+
+    if starless_rgb is None:
+        QMessageBox.critical(main, "StarNet Error", "Failed to load starless image.")
+        dialog.close()
+        return
+
+    # ensure 3ch float32 in [0..1]
+    if starless_rgb.ndim == 2:
+        starless_rgb = np.stack([starless_rgb] * 3, axis=-1)
+    elif starless_rgb.ndim == 3 and starless_rgb.shape[2] == 1:
+        starless_rgb = np.repeat(starless_rgb, 3, axis=2)
+    starless_rgb = np.clip(starless_rgb.astype(np.float32, copy=False), 0.0, 1.0)
+
+    # original image (from the doc) as 3ch float32, track if it was mono
+    orig = np.asarray(doc.image)
+    if orig.ndim == 2:
+        original_rgb = np.stack([orig] * 3, axis=-1)
+        orig_was_mono = True
+    elif orig.ndim == 3 and orig.shape[2] == 1:
+        original_rgb = np.repeat(orig, 3, axis=2)
+        orig_was_mono = True
+    else:
+        original_rgb = orig
+        orig_was_mono = False
+    original_rgb = original_rgb.astype(np.float32, copy=False)
+
+
+    # ---- Inversion back to the document’s domain ----
+    if did_stretch:
+        # Prefer the new Siril-style MTF meta if present
+        meta = getattr(main, "_starnet_stat_meta", None)
+        mtf_params_legacy = getattr(main, "_starnet_last_mtf_params", None)
+
+        if isinstance(meta, dict) and meta.get("scheme") == "siril_mtf":
+            dialog.append_text("Unstretching (Siril-style MTF pseudoinverse)...\n")
+            try:
+                s_vec = np.asarray(meta.get("s"), dtype=np.float32)
+                m_vec = np.asarray(meta.get("m"), dtype=np.float32)
+                h_vec = np.asarray(meta.get("h"), dtype=np.float32)
+                scale_factor = float(meta.get("scale", 1.0))
+
+                p = {"s": s_vec, "m": m_vec, "h": h_vec}
+                inv = _invert_mtf_unlinked_rgb(starless_rgb, p)
+
+                if scale_factor > 1.0:
+                    inv = inv * scale_factor
+
+                starless_rgb = np.clip(inv, 0.0, 1.0)
+            except Exception as e:
+                dialog.append_text(f"⚠️ Siril-style MTF inverse failed: {e}\n")
+
+        elif isinstance(meta, dict) and meta.get("scheme") == "statstretch":
+            # Back-compat: statistical round-trip with bp/m0
+            dialog.append_text("Unstretching (statistical inverse w/ original BP/M0)...\n")
+
+            bp_vec = np.asarray(meta.get("bp"), dtype=np.float32)
+            m0_vec = np.asarray(meta.get("m0"), dtype=np.float32)
+            scale_factor = float(meta.get("scale", 1.0))
+
+            inv = np.empty_like(starless_rgb, dtype=np.float32)
+            for c in range(3):
+                inv[..., c] = stretch_mono_image(
+                    starless_rgb[..., c],
+                    target_median=float(m0_vec[c]),
+                    normalize=False, apply_curves=False, curves_boost=0.0
+                )
+
+            inv += bp_vec.reshape((1, 1, 3))
+            inv = np.clip(inv, 0.0, 1.0)
+            if scale_factor > 1.0:
+                inv *= scale_factor
+            starless_rgb = np.clip(inv, 0.0, 1.0)
+
+        elif mtf_params_legacy:
+            # Very old MTF path (linked, single triple) – keep for safety
+            dialog.append_text("Unstretching (legacy MTF inverse)...\n")
+            try:
+                starless_rgb = _invert_mtf_linked_rgb(starless_rgb, mtf_params_legacy)
+                sc = float(mtf_params_legacy.get("scale", 1.0))
+                if sc > 1.0:
+                    starless_rgb = starless_rgb * sc
+            except Exception as e:
+                dialog.append_text(f"⚠️ Legacy MTF inverse failed: {e}\n")
+            starless_rgb = np.clip(starless_rgb, 0.0, 1.0)
+
+        # Clean up stashed meta so it can't leak to future ops
+        try:
+            if hasattr(main, "_starnet_stat_meta"):
+                delattr(main, "_starnet_stat_meta")
+        except Exception:
+            pass
+
+
+
+    # ---- Stars-Only = original − starless (linear-domain diff) ----
+    dialog.append_text("Generating stars-only image...\n")
+    stars_only = np.clip(original_rgb - starless_rgb, 0.0, 1.0)
+
+    # apply active mask (doc-based)
+    m3 = _active_mask3_from_doc(doc, stars_only.shape[1], stars_only.shape[0])
+    if m3 is not None:
+        stars_only *= m3
+        dialog.append_text("✅ Applied active mask to the stars-only image.\n")
+    else:
+        dialog.append_text("ℹ️ No active mask for stars-only; skipping.\n")
+
+    # If the original doc was mono, return a mono stars-only image
+    if orig_was_mono:
+        stars_to_push = stars_only.mean(axis=2).astype(np.float32, copy=False)
+    else:
+        stars_to_push = stars_only
+
+    # push Stars-Only as new document with suffix _stars
+    _push_as_new_doc(main, doc, stars_to_push, title_suffix="_stars", source="Stars-Only (StarNet)")
+    dialog.append_text("Stars-only image pushed.\n")
+
+    # mask-blend starless with original using active mask, then overwrite current view
+    dialog.append_text("Preparing to update current view with starless (mask-blend)...\n")
+    final_starless = _mask_blend_with_doc_mask(doc, starless_rgb, original_rgb)
+
+    # If the original doc was mono, collapse back to single-channel
+    if orig_was_mono:
+        final_to_apply = final_starless.mean(axis=2).astype(np.float32, copy=False)
+    else:
+        final_to_apply = final_starless.astype(np.float32, copy=False)
+
+    try:
+        meta = {
+            "step_name": "Stars Removed",
+            "bit_depth": "32-bit floating point",
+            "is_mono": bool(orig_was_mono),
+        }
+
+        # 🔹 Attach replay-last metadata
+        rp = getattr(main, "_last_remove_stars_params", None)
+        if isinstance(rp, dict):
+            replay_params = dict(rp)  # shallow copy so we don't mutate the stored one
+        else:
+            replay_params = {
+                "engine": "StarNet",
+                "is_linear": bool(did_stretch),
+                "did_stretch": bool(did_stretch),
+                "label": "Remove Stars (StarNet)",
+            }
+
+        replay_params.setdefault("engine", "StarNet")
+        replay_params.setdefault("label", "Remove Stars (StarNet)")
+
+        meta["replay_last"] = {
+            "op": "remove_stars",
+            "params": replay_params,
+        }
+
+        # Clean up the stash so it can't leak to the next unrelated op
+        try:
+            if hasattr(main, "_last_remove_stars_params"):
+                delattr(main, "_last_remove_stars_params")
+        except Exception:
+            pass
+
+        doc.apply_edit(
+            final_to_apply,
+            metadata=meta,
+            step_name="Stars Removed"
+        )
+        if hasattr(main, "_log"):
+            main._log("Stars Removed (StarNet)")
+    except Exception as e:
+        QMessageBox.critical(main, "StarNet Error", f"Failed to apply starless result:\n{e}")
+
+    dialog.append_text("Temporary files cleaned up.\n")
+    dialog.close()
+
+
+
+# ------------------------------------------------------------
+# CosmicClarityDarkStar
+# ------------------------------------------------------------
+def _run_darkstar(main, doc):
+    exe, base = _resolve_darkstar_exe(main)
+    if not exe or not base:
+        QMessageBox.critical(main, "Cosmic Clarity Folder Error",
+                             "Cosmic Clarity Dark Star executable not set.")
+        return
+
+    # --- Input/output folders per SASv2 ---
+    input_dir  = os.path.join(base, "input")
+    output_dir = os.path.join(base, "output")
+    os.makedirs(input_dir, exist_ok=True)
+    os.makedirs(output_dir, exist_ok=True)
+    _purge_darkstar_io(base, prefix=None, clear_input=True, clear_output=True)
+
+    # --- Config dialog (same as before) ---
+    cfg = DarkStarConfigDialog(main)
+    if not cfg.exec():
+        return
+    params = cfg.get_values()
+    disable_gpu = params["disable_gpu"]
+    mode = params["mode"]                         # "unscreen" or "additive"
+    show_extracted_stars = params["show_extracted_stars"]
+    stride = params["stride"]                     # 64..1024, default 512
+
+    # 🔹 Ask if image is linear (so we know whether to MTF-prestretch)
+    reply = QMessageBox.question(
+        main, "Image Linearity", "Is the current image linear?",
+        QMessageBox.StandardButton.Yes | QMessageBox.StandardButton.No,
+        QMessageBox.StandardButton.Yes
+    )
+    is_linear = (reply == QMessageBox.StandardButton.Yes)
+    did_prestretch = is_linear
+
+    # 🔹 Stash parameters for replay-last
+    try:
+        main._last_remove_stars_params = {
+            "engine": "CosmicClarityDarkStar",
+            "disable_gpu": bool(disable_gpu),
+            "mode": mode,
+            "show_extracted_stars": bool(show_extracted_stars),
+            "stride": int(stride),
+            "is_linear": bool(is_linear),
+            "did_prestretch": bool(did_prestretch),
+            "label": "Remove Stars (DarkStar)",
+        }
+    except Exception:
+        pass
+
+    # 🔁 Record headless command for Replay Last
+    try:
+        main._last_headless_command = {
+            "command_id": "remove_stars",
+            "preset": {
+                "tool": "darkstar",
+                "disable_gpu": bool(disable_gpu),
+                "mode": mode,
+                "show_extracted_stars": bool(show_extracted_stars),
+                "stride": int(stride),
+                "is_linear": bool(is_linear),
+                "did_prestretch": bool(did_prestretch),
+            },
+        }
+        if hasattr(main, "_log"):
+            main._log(
+                "[Replay] Recorded remove_stars (DarkStar, "
+                f"mode={mode}, stride={int(stride)}, "
+                f"gpu={'off' if disable_gpu else 'on'}, "
+                f"stars={'on' if show_extracted_stars else 'off'}, "
+                f"linear={'yes' if is_linear else 'no'})"
+            )
+    except Exception:
+        pass
+
+    # --- Build processing image (RGB float32, normalized) ---
+    src = np.asarray(doc.image)
+    if src.ndim == 2:
+        processing_image = np.stack([src] * 3, axis=-1)
+    elif src.ndim == 3 and src.shape[2] == 1:
+        processing_image = np.repeat(src, 3, axis=2)
+    else:
+        processing_image = src
+
+    processing_image = np.nan_to_num(
+        processing_image.astype(np.float32, copy=False),
+        nan=0.0, posinf=0.0, neginf=0.0
+    )
+
+    scale_factor = float(np.max(processing_image)) if processing_image.size else 1.0
+    if scale_factor > 1.0:
+        processing_norm = processing_image / scale_factor
+    else:
+        processing_norm = processing_image
+    processing_norm = np.clip(processing_norm, 0.0, 1.0)
+
+    # --- Optional Siril-style MTF pre-stretch for linear data ---
+    img_for_darkstar = processing_norm
+    if is_linear:
+        try:
+            mtf_params = _mtf_params_unlinked(
+                processing_norm,
+                shadows_clipping=-2.8,
+                targetbg=0.25
+            )
+            img_for_darkstar = _apply_mtf_unlinked_rgb(processing_norm, mtf_params)
+
+            # 🔐 Stash EXACT params for inverse step later
+            setattr(main, "_darkstar_mtf_meta", {
+                "s": np.asarray(mtf_params["s"], dtype=np.float32),
+                "m": np.asarray(mtf_params["m"], dtype=np.float32),
+                "h": np.asarray(mtf_params["h"], dtype=np.float32),
+                "scale": float(scale_factor),
+            })
+            if hasattr(main, "_log"):
+                main._log("[DarkStar] Applying Siril-style MTF pre-stretch for linear image.")
+        except Exception as e:
+            # If anything goes wrong, fall back to un-stretched normalized image
+            img_for_darkstar = processing_norm
+            try:
+                if hasattr(main, "_darkstar_mtf_meta"):
+                    delattr(main, "_darkstar_mtf_meta")
+            except Exception:
+                pass
+            if hasattr(main, "_log"):
+                main._log(f"[DarkStar] MTF pre-stretch failed, using normalized image only: {e}")
+    else:
+        # Non-linear: don't store any pre-stretch meta
+        try:
+            if hasattr(main, "_darkstar_mtf_meta"):
+                delattr(main, "_darkstar_mtf_meta")
+        except Exception:
+            pass
+
+    # --- Save pre-stretched image as 32-bit float TIFF for DarkStar ---
+    in_path = os.path.join(input_dir, "imagetoremovestars.tif")
+    try:
+        save_image(
+            img_for_darkstar,
+            in_path,
+            original_format="tif",
+            bit_depth="32-bit floating point",
+            original_header=None,
+            is_mono=False,  # we always send RGB to DarkStar
+            image_meta=None,
+            file_meta=None
+        )
+    except Exception as e:
+        QMessageBox.critical(main, "Cosmic Clarity", f"Failed to write input TIFF:\n{e}")
+        return
+
+    # --- Build CLI exactly like SASv2 (using --chunk_size, not chunk_size) ---
+    args = []
+    if disable_gpu:
+        args.append("--disable_gpu")
+    args += ["--star_removal_mode", mode]
+    if show_extracted_stars:
+        args.append("--show_extracted_stars")
+    args += ["--chunk_size", str(stride)]
+
+    command = [exe] + args
+
+    dlg = _ProcDialog(main, title="CosmicClarityDarkStar Progress")
+    thr = _ProcThread(command, cwd=output_dir)
+    thr.output_signal.connect(dlg.append_text)
+    thr.finished_signal.connect(
+        lambda rc, base=base, ds=did_prestretch: _on_darkstar_finished(
+            main, doc, rc, dlg, in_path, output_dir, base, ds
+        )
+    )
+    dlg.cancel_button.clicked.connect(thr.cancel)
+
+    dlg.show()
+    thr.start()
+    dlg.exec()
+
+
+
+
+def _resolve_darkstar_exe(main):
+    """
+    Return (exe_path, base_folder) or (None, None) on cancel/error.
+    Accepts either a folder (stored) or a direct executable path.
+    Saves the folder back to QSettings under 'paths/cosmic_clarity'.
+    """
+    settings = getattr(main, "settings", None)
+    raw = _get_setting_any(settings, ("paths/cosmic_clarity", "cosmic_clarity_folder"), "")
+
+    def _platform_exe_name():
+        return "setiastrocosmicclarity_darkstar.exe" if platform.system() == "Windows" \
+               else "setiastrocosmicclarity_darkstar"
+
+    exe_name = _platform_exe_name()
+
+    exe_path = None
+    base_folder = None
+
+    if raw:
+        if os.path.isfile(raw):
+            # user stored the executable path directly
+            exe_path = raw
+            base_folder = os.path.dirname(raw)
+        elif os.path.isdir(raw):
+            # user stored the parent folder
+            base_folder = raw
+            exe_path = os.path.join(base_folder, exe_name)
+
+    # if missing or invalid, let user pick the executable directly
+    if not exe_path or not os.path.exists(exe_path):
+        picked, _ = QFileDialog.getOpenFileName(main, "Select CosmicClarityDarkStar Executable", "", "Executable Files (*)")
+        if not picked:
+            return None, None
+        exe_path = picked
+        base_folder = os.path.dirname(picked)
+
+    # ensure exec bit on POSIX
+    if platform.system() in ("Darwin", "Linux"):
+        _ensure_exec_bit(exe_path)
+
+    # persist folder (not the exe) to the canonical key
+    if settings:
+        settings.setValue("paths/cosmic_clarity", base_folder)
+        settings.sync()
+
+    return exe_path, base_folder
+
+
+def _on_darkstar_finished(main, doc, return_code, dialog, in_path, output_dir, base_folder, did_prestretch):
+    dialog.append_text(f"\nProcess finished with return code {return_code}.\n")
+    if return_code != 0:
+        QMessageBox.critical(main, "CosmicClarityDarkStar Error",
+                             f"CosmicClarityDarkStar failed with return code {return_code}.")
+        _safe_rm(in_path); dialog.close(); return
+
+    starless_path = os.path.join(output_dir, "imagetoremovestars_starless.tif")
+    if not os.path.exists(starless_path):
+        QMessageBox.critical(main, "CosmicClarityDarkStar Error", "Starless image was not created.")
+        _safe_rm(in_path); dialog.close(); return
+
+    dialog.append_text(f"Loading starless image from {starless_path}...\n")
+    starless, _, _, _ = load_image(starless_path)
+    if starless is None:
+        QMessageBox.critical(main, "CosmicClarityDarkStar Error", "Failed to load starless image.")
+        _safe_rm(in_path); dialog.close(); return
+
+    if starless.ndim == 2 or (starless.ndim == 3 and starless.shape[2] == 1):
+        starless_rgb = np.stack([starless] * 3, axis=-1)
+    else:
+        starless_rgb = starless
+    starless_rgb = starless_rgb.astype(np.float32, copy=False)
+
+    src = np.asarray(doc.image)
+    if src.ndim == 2:
+        original_rgb = np.stack([src] * 3, axis=-1)
+        orig_was_mono = True
+    elif src.ndim == 3 and src.shape[2] == 1:
+        original_rgb = np.repeat(src, 3, axis=2)
+        orig_was_mono = True
+    else:
+        original_rgb = src
+        orig_was_mono = False
+    original_rgb = original_rgb.astype(np.float32, copy=False)
+
+    # --- Undo the MTF pre-stretch (if we did one) ---
+    if did_prestretch:
+        meta = getattr(main, "_darkstar_mtf_meta", None)
+        if isinstance(meta, dict):
+            dialog.append_text("Unstretching starless result (DarkStar MTF inverse)...\n")
+            try:
+                s_vec = np.asarray(meta.get("s"), dtype=np.float32)
+                m_vec = np.asarray(meta.get("m"), dtype=np.float32)
+                h_vec = np.asarray(meta.get("h"), dtype=np.float32)
+                scale = float(meta.get("scale", 1.0))
+
+                p = {"s": s_vec, "m": m_vec, "h": h_vec}
+                inv = _invert_mtf_unlinked_rgb(starless_rgb, p)
+
+                if scale > 1.0:
+                    inv *= scale
+
+                starless_rgb = np.clip(inv, 0.0, 1.0)
+            except Exception as e:
+                dialog.append_text(f"⚠️ DarkStar MTF inverse failed: {e}\n")
+
+        # Clean up pre-stretch meta so it can't leak into another op
+        try:
+            if hasattr(main, "_darkstar_mtf_meta"):
+                delattr(main, "_darkstar_mtf_meta")
+        except Exception:
+            pass
+
+    # --- stars-only optional push (as before) ---
+    stars_path = os.path.join(output_dir, "imagetoremovestars_stars_only.tif")
+    if os.path.exists(stars_path):
+        dialog.append_text(f"Loading stars-only image from {stars_path}...\n")
+        stars_only, _, _, _ = load_image(stars_path)
+        if stars_only is not None:
+            if stars_only.ndim == 2 or (stars_only.ndim == 3 and stars_only.shape[2] == 1):
+                stars_only = np.stack([stars_only] * 3, axis=-1)
+            stars_only = stars_only.astype(np.float32, copy=False)
+            m3 = _active_mask3_from_doc(doc, stars_only.shape[1], stars_only.shape[0])
+            if m3 is not None:
+                stars_only *= m3
+                dialog.append_text("✅ Applied active mask to stars-only image.\n")
+            else:
+                dialog.append_text("ℹ️ Mask not active for stars-only; skipping.\n")
+
+            # If the original doc was mono, collapse stars-only back to single channel
+            if orig_was_mono:
+                stars_to_push = stars_only.mean(axis=2).astype(np.float32, copy=False)
+            else:
+                stars_to_push = stars_only
+
+            _push_as_new_doc(main, doc, stars_to_push, title_suffix="_stars", source="Stars-Only (DarkStar)")
+        else:
+            dialog.append_text("Failed to load stars-only image.\n")
+    else:
+        dialog.append_text("No stars-only image generated.\n")
+
+    # --- Mask-blend starless → overwrite current doc (in original domain) ---
+    dialog.append_text("Mask-blending starless image before update...\n")
+    final_starless = _mask_blend_with_doc_mask(doc, starless_rgb, original_rgb)
+
+    # If the original doc was mono, collapse back to single-channel
+    if orig_was_mono:
+        final_to_apply = final_starless.mean(axis=2).astype(np.float32, copy=False)
+    else:
+        final_to_apply = final_starless.astype(np.float32, copy=False)
+
+    try:
+        meta = {
+            "step_name": "Stars Removed",
+            "bit_depth": "32-bit floating point",
+            "is_mono": bool(orig_was_mono),
+        }
+
+        # 🔹 Attach replay-last metadata
+        rp = getattr(main, "_last_remove_stars_params", None)
+        if isinstance(rp, dict):
+            replay_params = dict(rp)
+        else:
+            replay_params = {
+                "engine": "CosmicClarityDarkStar",
+                "label": "Remove Stars (DarkStar)",
+            }
+
+        replay_params.setdefault("engine", "CosmicClarityDarkStar")
+        replay_params.setdefault("label", "Remove Stars (DarkStar)")
+
+        meta["replay_last"] = {
+            "op": "remove_stars",
+            "params": replay_params,
+        }
+
+        # Clean up stash
+        try:
+            if hasattr(main, "_last_remove_stars_params"):
+                delattr(main, "_last_remove_stars_params")
+        except Exception:
+            pass
+
+        doc.apply_edit(
+            final_to_apply,
+            metadata=meta,
+            step_name="Stars Removed"
+        )
+        if hasattr(main, "_log"):
+            main._log("Stars Removed (DarkStar)")
+    except Exception as e:
+        QMessageBox.critical(main, "CosmicClarityDarkStar", f"Failed to apply result:\n{e}")
+
+    # --- cleanup ---
+    try:
+        _safe_rm(in_path)
+        _safe_rm(starless_path)
+        _safe_rm(os.path.join(output_dir, "imagetoremovestars_stars_only.tif"))
+
+        # 🔸 Final sweep: nuke any imagetoremovestars* leftovers in both dirs
+        base_folder = os.path.dirname(output_dir)  # <-- derive CC base from output_dir
+        _purge_darkstar_io(base_folder, prefix="imagetoremovestars", clear_input=True, clear_output=True)
+
+        dialog.append_text("Temporary files cleaned up.\n")
+    except Exception as e:
+        dialog.append_text(f"Cleanup error: {e}\n")
+
+    dialog.close()
+
+
+# ------------------------------------------------------------
+# Mask helpers (doc-centric)
+# ------------------------------------------------------------
+# _active_mask_array_from_doc is now imported from setiastro.saspro.widgets.image_utils
+
+
+def _active_mask3_from_doc(doc, w, h) -> np.ndarray | None:
+    """Return 3-channel mask resized to (h,w) if a doc-level mask exists; else None."""
+    m = _active_mask_array_from_doc(doc)
+    if m is None:
+        return None
+    if m.shape != (h, w):
+        if cv2 is not None:
+            m = cv2.resize(m, (w, h), interpolation=cv2.INTER_NEAREST)
+        else:
+            yi = (np.linspace(0, m.shape[0] - 1, h)).astype(np.int32)
+            xi = (np.linspace(0, m.shape[1] - 1, w)).astype(np.int32)
+            m = m[yi][:, xi]
+    return np.repeat(m[:, :, None], 3, axis=2).astype(np.float32, copy=False)
+
+
+def _mask_blend_with_doc_mask(doc, starless_rgb: np.ndarray, original_rgb: np.ndarray) -> np.ndarray:
+    """Blend using mask from doc if present: result = starless*m + original*(1-m)."""
+    m = _active_mask_array_from_doc(doc)
+    if m is None:
+        return starless_rgb
+    h, w = starless_rgb.shape[:2]
+    if m.shape != (h, w):
+        if cv2 is not None:
+            m = cv2.resize(m, (w, h), interpolation=cv2.INTER_NEAREST)
+        else:
+            yi = (np.linspace(0, m.shape[0] - 1, h)).astype(np.int32)
+            xi = (np.linspace(0, m.shape[1] - 1, w)).astype(np.int32)
+            m = m[yi][:, xi]
+    m3 = np.repeat(m[:, :, None], 3, axis=2)
+    return np.clip(starless_rgb * m3 + original_rgb * (1.0 - m3), 0.0, 1.0).astype(np.float32, copy=False)
+
+
+def _derive_view_base_title(main, doc) -> str:
+    """
+    Prefer the active view's title (respecting per-view rename/override),
+    fallback to the document display name, then to doc.name, and finally 'Image'.
+    Also strips any decorations (mask glyph, 'Active View:' prefix) if available.
+    """
+    # 1) Ask main for a subwindow for this document, if it exposes a helper
+    try:
+        if hasattr(main, "_subwindow_for_document"):
+            sw = main._subwindow_for_document(doc)
+            if sw:
+                w = sw.widget() if hasattr(sw, "widget") else sw
+                # Preferred: view's effective title (includes per-view override)
+                if hasattr(w, "_effective_title"):
+                    t = w._effective_title() or ""
+                else:
+                    t = sw.windowTitle() if hasattr(sw, "windowTitle") else ""
+                if hasattr(w, "_strip_decorations"):
+                    t, _ = w._strip_decorations(t)
+                if t.strip():
+                    return t.strip()
+    except Exception:
+        pass
+
+    # 2) Try scanning MDI for a subwindow whose widget holds this document
+    try:
+        mdi = (getattr(main, "mdi_area", None)
+               or getattr(main, "mdiArea", None)
+               or getattr(main, "mdi", None))
+        if mdi and hasattr(mdi, "subWindowList"):
+            for sw in mdi.subWindowList():
+                w = sw.widget()
+                if getattr(w, "document", None) is doc:
+                    t = sw.windowTitle() if hasattr(sw, "windowTitle") else ""
+                    if hasattr(w, "_strip_decorations"):
+                        t, _ = w._strip_decorations(t)
+                    if t.strip():
+                        return t.strip()
+    except Exception:
+        pass
+
+    # 3) Fallback to document's display name (then name, then generic)
+    try:
+        if hasattr(doc, "display_name"):
+            t = doc.display_name()
+            if t and t.strip():
+                return t.strip()
+    except Exception:
+        pass
+    return (getattr(doc, "name", "") or "Image").strip()
+
+
+# ------------------------------------------------------------
+# New document helper
+# ------------------------------------------------------------
+def _push_as_new_doc(main, doc, arr: np.ndarray, title_suffix="_stars", source="Stars-Only"):
+    dm = getattr(main, "docman", None)
+    if not dm or not hasattr(dm, "open_array"):
+        return
+    try:
+        # Use the current view's title if available (respects per-view rename)
+        base = _derive_view_base_title(main, doc)
+
+        # avoid double-suffix if user already named it with the suffix
+        if title_suffix and base.endswith(title_suffix):
+            title = base
+        else:
+            title = f"{base}{title_suffix}"
+
+        meta = {
+            "bit_depth": "32-bit floating point",
+            "is_mono": (arr.ndim == 2),
+            "source": source,
+        }
+        newdoc = dm.open_array(arr.astype(np.float32, copy=False), metadata=meta, title=title)
+        if hasattr(main, "_spawn_subwindow_for"):
+            main._spawn_subwindow_for(newdoc)
+    except Exception:
+        pass
+
+
+
+# ------------------------------------------------------------
+# Utilities
+# ------------------------------------------------------------
+def _ensure_exec_bit(path: str):
+    if platform.system() == "Windows":
+        return
+    try:
+        st = os.stat(path)
+        os.chmod(path, st.st_mode | stat.S_IXUSR | stat.S_IXGRP | stat.S_IXOTH)
+    except Exception:
+        pass
+
+
+def _safe_rm(p):
+    try:
+        if p and os.path.exists(p):
+            os.remove(p)
+    except Exception:
+        pass
+
+def _safe_rm_globs(patterns: list[str]):
+    for pat in patterns:
+        try:
+            for fp in glob.glob(pat):
+                _safe_rm(fp)
+        except Exception:
+            pass
+
+def _purge_darkstar_io(base_folder: str, *, prefix: str | None = None, clear_input=True, clear_output=True):
+    """Delete old image-like files from CC DarkStar input/output."""
+    try:
+        inp = os.path.join(base_folder, "input")
+        out = os.path.join(base_folder, "output")
+        if clear_input and os.path.isdir(inp):
+            for fn in os.listdir(inp):
+                fp = os.path.join(inp, fn)
+                if os.path.isfile(fp) and (prefix is None or fn.startswith(prefix)):
+                    _safe_rm(fp)
+        if clear_output and os.path.isdir(out):
+            for fn in os.listdir(out):
+                fp = os.path.join(out, fn)
+                if os.path.isfile(fp) and (prefix is None or fn.startswith(prefix)):
+                    _safe_rm(fp)
+    except Exception:
+        pass
+
+
+# ------------------------------------------------------------
+# Proc runner & dialog (merged stdout/stderr)
+# ------------------------------------------------------------
+class _ProcThread(QThread):
+    output_signal = pyqtSignal(str)
+    finished_signal = pyqtSignal(int)
+
+    def __init__(self, command: list[str], cwd: str | None = None, parent=None):
+        super().__init__(parent)
+        self.command = command
+        self.cwd = cwd
+        self.process = None
+
+    def cancel(self):
+        """Request the subprocess to stop."""
+        if self.process:
+            try:
+                self.process.kill()
+            except Exception:
+                pass
+
+
+    def run(self):
+        import subprocess
+        import os
+        env = os.environ.copy()
+        for k in ("PYTHONHOME","PYTHONPATH","DYLD_LIBRARY_PATH","DYLD_FALLBACK_LIBRARY_PATH","PYTHONEXECUTABLE"):
+            env.pop(k, None)
+        rc = -1
+        try:
+            self.process = subprocess.Popen(
+                self.command, cwd=self.cwd,
+                stdout=subprocess.PIPE, stderr=subprocess.STDOUT,
+                universal_newlines=True, text=True, start_new_session=True, env=env
+            )
+            for line in iter(self.process.stdout.readline, ""):
+                if not line: break
+                self.output_signal.emit(line.rstrip())
+            try:
+                self.process.stdout.close()
+            except Exception:
+                pass
+            rc = self.process.wait()
+        except Exception as e:
+            self.output_signal.emit(str(e))
+            rc = -1
+        finally:
+            self.process = None
+        self.finished_signal.emit(rc)
+
+
+class _ProcDialog(QDialog):
+    def __init__(self, parent, title="Process"):
+        super().__init__(parent)
+        self.setWindowTitle(title)
+        self.setMinimumSize(600, 420)
+        lay = QVBoxLayout(self)
+        self.text = QTextEdit(self); self.text.setReadOnly(True)
+        lay.addWidget(self.text)
+        self.cancel_button = QPushButton("Cancel", self)
+        lay.addWidget(self.cancel_button)
+
+    def append_text(self, s: str):
+        try:
+            self.text.append(s)
+        except Exception:
+            pass
+
+
+class DarkStarConfigDialog(QDialog):
+    """
+    SASv2-style config UI:
+      - Disable GPU: Yes/No (default No)
+      - Star Removal Mode: unscreen | additive (default unscreen)
+      - Show Extracted Stars: Yes/No (default No)
+      - Stride (powers of 2): 64,128,256,512,1024 (default 512)
+    """
+    def __init__(self, parent=None):
+        super().__init__(parent)
+        self.setWindowTitle("CosmicClarity Dark Star Settings")
+
+        self.chk_disable_gpu = QCheckBox("Disable GPU")
+        self.chk_disable_gpu.setChecked(False)  # default No (unchecked)
+
+        self.cmb_mode = QComboBox()
+        self.cmb_mode.addItems(["unscreen", "additive"])
+        self.cmb_mode.setCurrentText("unscreen")
+
+        self.chk_show_stars = QCheckBox("Show Extracted Stars")
+        self.chk_show_stars.setChecked(True)
+
+        self.cmb_stride = QComboBox()
+        for v in (64, 128, 256, 512, 1024):
+            self.cmb_stride.addItem(str(v), v)
+        self.cmb_stride.setCurrentText("512")  # default 512
+
+        form = QFormLayout()
+        form.addRow("Star Removal Mode:", self.cmb_mode)
+        form.addRow("Stride (power of two):", self.cmb_stride)
+        form.addRow("", self.chk_disable_gpu)
+        form.addRow("", self.chk_show_stars)
+
+        btns = QDialogButtonBox(QDialogButtonBox.StandardButton.Ok | QDialogButtonBox.StandardButton.Cancel, parent=self)
+        btns.accepted.connect(self.accept)
+        btns.rejected.connect(self.reject)
+
+        layout = QVBoxLayout(self)
+        layout.addLayout(form)
+        layout.addWidget(btns)
+
+    def get_values(self):
+        return {
+            "disable_gpu": self.chk_disable_gpu.isChecked(),
+            "mode": self.cmb_mode.currentText(),
+            "show_extracted_stars": self.chk_show_stars.isChecked(),
+            "stride": int(self.cmb_stride.currentData()),
+        }