img2nl 0.1.5

Image → natural language summary with heuristics, thumbnails, and LLM transport hints

img2nl

AI Cost Tracking

• 🤖 LLM usage: $1.1910 (4 commits)
• 👤 Human dev: ~$350 (3.5h @ $100/h, 30min dedup)

Generated on 2026-06-09 using openrouter/qwen/qwen3-coder-next

---

Heuristic image → natural language summary for transport to LLM and other services.

No vision LLM required for the core path — uses layered heuristics (Pillow, optional OpenCV, perceptual hash, conditional OCR/QR/YOLO).

Features

Layer	Module	Extra	What it detects
0	colors, dynamics, noise, objects, patterns	[analyze]	palette, contrast, flat regions, UI blocks
1	edges	[opencv]	blur, edge density, text likelihood
2	fingerprint, similarity	[similarity]	pHash/dHash/wHash, screen match
3	special_hits	[scan] / [ocr]	QR/barcode, OCR (conditional)
4	semantic_hits	[detect]	YOLO labels (opt-in)

Output includes features.scene.scene_class, llm_hint, and optional similarity when reference_fingerprint is passed.

Full architecture: docs/detection-pipeline.md

Install

pip install -e ".[analyze]"              # core (Pillow + NumPy)
pip install -e ".[full]"                 # analyze + opencv + similarity + scan
pip install -e ".[analyze,translate]"    # + argostranslate offline
bash install-dev.sh                        # full *2img2nl stack

Optional extras: opencv, similarity, scan, ocr, detect — see pyproject.toml.

CLI

img2nl analyze photo.png --json
img2nl analyze photo.png --locale de --translate-mode offline
dsl2img2nl -c "ANALYZE photo.png" --json
uri2img2nl query "img2nl://analyze?path=photo.png&locale=pl"
python -c "from img2nl.i18n import supported_locales; print(supported_locales())"

Offline translation (argostranslate)

Static catalog covers 38 European langs; for scalable updates use neural offline translation:

pip install img2nl[analyze,translate]
img2nl translate-install en pl
img2nl analyze photo.png --locale de --translate-mode offline

Modes: auto (catalog pl/en, else argos), offline (require argos), catalog (JSON only).

Python API

from img2nl import analyze_image

result = analyze_image("screen.png", skip_thumbnail=True)
print(result.text, result.features["scene"]["scene_class"], result.llm_hint)

# compare with previous capture
prev = analyze_image("screen_a.png", skip_thumbnail=True)
cur = analyze_image(
    "screen_b.png",
    skip_thumbnail=True,
    reference_fingerprint=prev.features["fingerprint"],
    enable_detect=False,  # True → YOLO (heavy)
)
print(cur.features.get("similarity", {}))

Packages

Package	Role
img2nl	Core heuristics + describe + thumbnail + layered detection
uri2img2nl	img2nl:// URI layer
dsl2img2nl	DSL bus (ANALYZE, QUERY, LLM_HINT)
cli2img2nl	Shell adapter

See packages/README.md.

VQL integration

Layered pipeline with img2vql in oqlos/vql:

pip install -e ".[analyze,similarity,opencv,scan]"
pip install -e ~/github/oqlos/vql/packages/img2vql

# adopt → metadata (fingerprint, special_hits, scene_class)
uri2vql analyze-window --image capture.png --out app.vql.json

# smart skip when screen unchanged
uri2vql analyze-window --image capture.png --out app.vql.json

uri2vql refresh-window --vql-program app.vql.json --image capture.png
uri2vql compare-window --vql-program app.vql.json --image capture.png
img2vql diagnose capture.png --vql-program app.vql.json --save
uri2vql resolve "odśwież metadata vql" --file app.vql.json --image capture.png

# end-to-end demo
bash ~/github/oqlos/vql/examples/img2nl-vql-flow.sh capture.png app.vql.json

Docs

Doc	Content
docs/detection-pipeline.md	Warstwy 0–4, schema JSON, VQL cache
CHANGELOG.md	Historia zmian
TODO.md	Backlog

License

Licensed under Apache-2.0.