feat: auto-generate scene descriptions from image/video assets (#689)

scripts/auto-scene-descriptions.py

@@ -0,0 +1,160 @@
+#!/usr/bin/env python3
+"""
+auto-scene-descriptions.py — Generate scene descriptions from image/video assets.
+
+Scans an assets directory, uses vision model to describe each asset,
+outputs training pairs in timmy-config format.
+
+Usage:
+    python3 scripts/auto-scene-descriptions.py --scan ~/assets/
+    python3 scripts/auto-scene-descriptions.py --scan ~/assets/ --output training-data/scene-from-media.jsonl
+    python3 scripts/auto-scene-descriptions.py --scan ~/assets/ --dry-run
+"""
+
+import argparse
+import json
+import os
+import subprocess
+import sys
+from pathlib import Path
+
+IMAGE_EXTS = {".jpg", ".jpeg", ".png", ".gif", ".webp", ".bmp"}
+VIDEO_EXTS = {".mp4", ".webm", ".mov", ".avi", ".mkv"}
+SUPPORTED_EXTS = IMAGE_EXTS | VIDEO_EXTS
+
+
+def scan_assets(directory: str) -> list[Path]:
+    """Find all image/video assets in a directory."""
+    assets = []
+    for root, dirs, files in os.walk(directory):
+        dirs[:] = [d for d in dirs if d not in {".git", "node_modules", "__pycache__"}]
+        for f in sorted(files):
+            ext = Path(f).suffix.lower()
+            if ext in SUPPORTED_EXTS:
+                assets.append(Path(root) / f)
+    return assets
+
+
+def extract_video_frame(video_path: Path) -> Path | None:
+    """Extract a representative frame from a video using ffmpeg."""
+    frame_path = video_path.with_suffix(".frame.jpg")
+    try:
+        subprocess.run(
+            ["ffmpeg", "-i", str(video_path), "-vframes", "1", "-ss", "5", "-y", str(frame_path)],
+            capture_output=True, timeout=30
+        )
+        if frame_path.exists():
+            return frame_path
+    except (subprocess.TimeoutExpired, FileNotFoundError):
+        pass
+    return None
+
+
+def describe_with_ollama(image_path: Path) -> str:
+    """Generate a scene description using local Ollama vision model."""
+    try:
+        result = subprocess.run(
+            ["ollama", "run", "llava", f"Describe this image as a visual scene for a film. Include mood, colors, composition, and camera angle. Be specific and vivid in 2-3 sentences. Image: {image_path}"],
+            capture_output=True, text=True, timeout=60
+        )
+        if result.returncode == 0:
+            return result.stdout.strip()
+    except (subprocess.TimeoutExpired, FileNotFoundError):
+        pass
+    return ""
+
+
+def describe_with_fallback(image_path: Path) -> str:
+    """Generate a basic scene description from filename/path."""
+    name = image_path.stem.replace("_", " ").replace("-", " ")
+    parent = image_path.parent.name.replace("_", " ").replace("-", " ")
+    return f"A scene depicting {name} in a {parent} setting. Visual composition inferred from asset location and naming conventions."
+
+
+def build_training_pair(asset_path: Path, description: str, asset_index: int) -> dict:
+    """Build a training pair in timmy-config scene description format."""
+    name = asset_path.stem
+    parent = asset_path.parent.name
+
+    # Infer mood from path/name keywords
+    mood_keywords = {
+        "dark": "melancholic", "light": "hopeful", "warm": "nostalgic",
+        "cold": "isolated", "bright": "energetic", "sunset": "bittersweet",
+        "night": "mysterious", "morning": "refreshing", "rain": "contemplative",
+    }
+    mood = "neutral"
+    name_lower = name.lower() + parent.lower()
+    for keyword, m in mood_keywords.items():
+        if keyword in name_lower:
+            mood = m
+            break
+
+    return {
+        "song": f"asset-{asset_index:04d}",
+        "beat": 1,
+        "lyric_line": f"[Visual asset: {asset_path.name}]",
+        "scene": {
+            "mood": mood,
+            "colors": ["inferred"],
+            "composition": "frame",
+            "camera": "static",
+            "description": description,
+            "source": "auto-generated",
+            "asset_path": str(asset_path),
+        }
+    }
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Auto-generate scene descriptions from assets")
+    parser.add_argument("--scan", required=True, help="Assets directory to scan")
+    parser.add_argument("--output", default=None, help="Output JSONL path")
+    parser.add_argument("--dry-run", action="store_true", help="Show what would be generated")
+    parser.add_argument("--vision", action="store_true", help="Use Ollama vision model (slow)")
+    args = parser.parse_args()
+
+    assets = scan_assets(args.scan)
+    print(f"Found {len(assets)} assets in {args.scan}", file=sys.stderr)
+
+    if not assets:
+        print("No assets found.", file=sys.stderr)
+        sys.exit(0)
+
+    pairs = []
+    for i, asset in enumerate(assets):
+        if args.vision:
+            # For videos, extract frame first
+            image_for_vision = asset
+            if asset.suffix.lower() in VIDEO_EXTS:
+                frame = extract_video_frame(asset)
+                if frame:
+                    image_for_vision = frame
+
+            description = describe_with_ollama(image_for_vision)
+            if not description:
+                description = describe_with_fallback(asset)
+        else:
+            description = describe_with_fallback(asset)
+
+        pair = build_training_pair(asset, description, i)
+        pairs.append(pair)
+
+        if args.dry_run and i < 5:
+            print(f"  {asset.name}: {description[:80]}")
+
+    if args.dry_run:
+        print(f"\nWould generate {len(pairs)} training pairs")
+        return
+
+    # Write output
+    output_path = args.output or os.path.expanduser("~/.hermes/training-data/scene-from-media.jsonl")
+    os.makedirs(os.path.dirname(output_path), exist_ok=True)
+    with open(output_path, "w") as f:
+        for p in pairs:
+            f.write(json.dumps(p, ensure_ascii=False) + "\n")
+
+    print(f"Generated {len(pairs)} scene descriptions → {output_path}", file=sys.stderr)
+
+
+if __name__ == "__main__":
+    main()