diff --git a/docs/plans/2026-04-06-mask-generation-design.md b/docs/plans/2026-04-06-mask-generation-design.md
new file mode 100644
index 0000000..90ab0a4
--- /dev/null
+++ b/docs/plans/2026-04-06-mask-generation-design.md
@@ -0,0 +1,78 @@
+# Mask Generation Design
+
+## Overview
+
+Add per-frame PNG mask generation to the 8-cut export pipeline for SELVA dataset creation. Two methods are supported: Depth Anything V2 (fast, depth-based foreground) and SAM2 (accurate, propagated segmentation). Both run as isolated subprocesses inside a dedicated ML venv.
+
+## UI Changes
+
+### Settings dialog
+
+A "Settings…" button added to the main window top bar opens a `SettingsDialog` (QDialog). It contains a "ML Tools" section with:
+- Status label: "Not installed" / "Installed"
+- **Install** button (becomes **Reinstall** once venv exists)
+- Read-only `QPlainTextEdit` streaming pip install output line-by-line
+
+### Mask generation row
+
+A new row added to the main window below the export row:
+- `QComboBox`: "Depth Anything" / "SAM"
+- **Generate Masks** button — disabled until venv is installed
+- Operates on `self._last_export_path` (set after each successful export)
+- Output folder: `<clip_stem>_masks/` next to the video
+- Frames named `frame_0000.png`, `frame_0001.png`, …
+- Progress streamed to status bar; button disabled during run
+
+## Venv
+
+Path: `~/.8cut/venv/`
+
+Install command (run as subprocess, output streamed to dialog):
+```
+~/.8cut/venv/bin/pip install torch torchvision transformers opencv-python Pillow segment-anything-2
+```
+
+Setup steps:
+1. `python -m venv ~/.8cut/venv`
+2. `~/.8cut/venv/bin/pip install --upgrade pip`
+3. `~/.8cut/venv/bin/pip install torch torchvision transformers opencv-python Pillow segment-anything-2`
+
+Venv existence check: `Path("~/.8cut/venv/bin/python").expanduser().exists()`.
+
+## tools/depth_masks.py
+
+CLI: `python tools/depth_masks.py --input <video.mp4> --output <dir>`
+
+1. Extract frames with OpenCV (`cv2.VideoCapture`)
+2. Load `transformers` depth-estimation pipeline: `depth-anything/Depth-Anything-V2-Large-hf`, device=`cuda` if available else `cpu`
+3. For each frame:
+   - Run depth estimation → float depth array
+   - Normalise to 0–255
+   - Apply Otsu threshold (`cv2.threshold(..., cv2.THRESH_OTSU)`) → binary mask
+   - Save as `frame_NNNN.png`
+   - Print `frame N/total` to stdout
+4. Exit 0 on success, non-zero on error
+
+## tools/sam_masks.py
+
+CLI: `python tools/sam_masks.py --input <video.mp4> --output <dir>`
+
+1. Extract all frames to a temp directory with OpenCV
+2. Load SAM2 video predictor (`sam2.build_sam.build_sam2_video_predictor`), checkpoint auto-downloaded from HuggingFace
+3. Init predictor on the frame directory
+4. Add point prompt: center pixel of frame 0 as positive point `[[w//2, h//2]]`, label `[1]`
+5. Propagate masks across all frames (`propagate_in_video`)
+6. For each frame: extract binary mask, save as `frame_NNNN.png`
+7. Print `frame N/total` to stdout
+
+## MaskWorker
+
+`MaskWorker(QThread)` mirrors `ExportWorker`:
+- `__init__(script: str, input_path: str, output_dir: str, venv_python: str)`
+- `run()`: calls `subprocess.Popen([venv_python, script, "--input", input_path, "--output", output_dir])`, reads stdout line-by-line, emits `progress = pyqtSignal(str)`, emits `finished = pyqtSignal()` or `error = pyqtSignal(str)`
+
+## MainWindow wiring
+
+- `self._last_export_path: str = ""` — set in `_on_export_done`
+- `_on_generate_masks()`: builds output dir path, creates it, instantiates `MaskWorker` with the selected script
+- Venv installed check on startup: if venv exists, enable the Generate Masks button