BLS 4.0 (Bundeslebensmittelschlüssel) Icon Set

In my project ACP (Adaptive Calorie Tracker) I am using the BLS Dataset for the nutritional information of generic German products. Unfortunatley I was missing clean, same-styled Icons for each of the 7140 individual entries. Thats why I used AI to genrate my own. If you need something like this too: here you go.

10 random samples per BLS Hauptgruppe, one row per category:

Same items with backgrounds removed (the checker is just to show the alpha — the actual files are transparent):

Mirror notice. This repo lives canonically at git.moritz.run/moritz/bls-icons with all icons stored via Git LFS. The GitHub mirror carries the code and metadata only — clone from the canonical source if you want the binary images.

Use it in your app

git clone ssh://git@git.moritz.run:2222/moritz/bls-icons.git
cd bls-icons
git lfs pull        # download all 7140 PNGs (~6.4 GB working tree)

icon_path = f"icons/{bls_code}.png"           # transparent
icon_path = f"icons_raw/{bls_code}.png"       # white background

Every BLS 4.0 code resolves to a file directly — no alias lookup, no preprocessing. Without git lfs pull you only get the metadata (~23 MB, clones in seconds).

Dataset

BLS 4.0 codes covered	7140
Files in icons/ and icons_raw/	7140 each
Distinct images	4987
Resolution	1024×1024 PNG
icons_raw/	source images, white background
icons/	transparent (after background removal)
items.csv	4987 canonical entries: code, name_de, name_en
aliases.csv	7140 bls_code → icon_code map (informational; lookup is no longer required)

The 7140 → 4987 reduction collapses items that differ only by preparation method (raw, cooked, fried, …) onto the same image. v3 materializes each duplicate as its own file, so callers can icons/{bls_code}.png directly. Server-side LFS storage stays at ~4.5 GB because identical blobs are content-addressed; only the working tree grows to ~6.4 GB.

How it was made

1. Source. BLS 4.0 Excel parsed via openpyxl, 7140 items.
2. Deduplication. Regex-strip preparation suffixes from item names → 4987 canonical icons. aliases.csv keeps the full mapping.
3. Prompt generation. Per item, gpt-5-mini reads the German + English name and the style spec (comic_v4.md in act-img-gen) and returns an image prompt. Sync API call, ~3 cents per 100 items.
4. Image generation. gpt-image-2 at quality low, 1024×1024, via the OpenAI Batch API (50 % discount, async with 24 h window). Output is a PNG with white background. ~$22 for the full 4987-item run.
5. Background removal. BiRefNet-massive via the rembg library, run on Modal serverless A10G GPUs. ~$0.25 and ~10–12 min wall time for 4987 icons.
6. Alias materialization. tools/sync_icons.py expands aliases.csv into the 7140-file working tree.

The orchestrator run_pipeline.py chains all steps with retries, idempotent resume, full logging to pipeline.log, and synchronous regeneration for items that hit OpenAI's image moderation (raw animal products occasionally trigger false-positive safety_violations=[sexual]).

After the initial generation, every icon was hand-reviewed and ~22 % were refined: 489 alpha-mask swaps where naïve PIL flood-fill beat BiRefNet, 612 prompt rewrites + image regens informed by per-item reviewer feedback (across two rounds), 16 white-powder items (flour, starch, milk powder) collapsed onto a single canonical image to avoid the BiRefNet-on-white mask issue, plus a handful of hand-picked swaps and reverts. Total refinement cost: ~$2.66.

Earlier dataset shapes are preserved as tags: v1.0 (pre-refinement, 4987 files), v2.0 (post-review, 4987 files with alias lookup), v3.0 (this version, 7140 files with aliases pre-resolved).

Regenerate

# companion repo: prompter + image-gen + style spec
git clone ssh://git@git.moritz.run:2222/moritz/act-img-gen.git ../act-img-gen

# local deps
pip install -r requirements.txt
pip install -r ../act-img-gen/requirements.txt

# OpenAI key for prompter + image gen
echo "OPENAI_API_KEY=sk-..." > ../act-img-gen/.env

# end-to-end run (~$22 + ~$0.25 Modal, completes in <24 h)
python run_pipeline.py

Models

Step	Model	Mode	Approx. cost (full 4987-item run)
Prompter	gpt-5-mini (reasoning_effort=minimal)	sync	~$1
Image gen	gpt-image-2 quality low	OpenAI Batch	~$22
Background removal	BiRefNet-massive (rembg)	Modal A10G GPU	~$0.25

Repo layout

.
├── items.csv               4987 canonical icons (code, name_de, name_en)
├── aliases.csv             7140 bls_code → icon_code mapping (informational)
├── run_pipeline.py         end-to-end orchestrator (entry point)
├── modal_postprocess.py    Modal entry point for background removal
├── grid.png                README preview (regenerable via tools/make_grid.py)
├── grid_alpha.png          transparent variant of the same grid
├── icons_raw/              white-bg PNGs, one per BLS code (LFS, 7140 files)
├── icons/                  transparent PNGs, one per BLS code (LFS, 7140 files)
└── tools/
    ├── sync_icons.py       copies act-img-gen output → icons_raw/, expands aliases
    ├── fix_missing.py      sync regen for items missing from icons_raw/
    └── make_grid.py        regenerates grid.png / grid_alpha.png

Storage

Icons are stored via Git LFS (*.png in icons/ and icons_raw/). A plain git clone only fetches the small text/CSV files; binaries arrive on first checkout (or git lfs pull). The repo itself stays small enough to clone in seconds.

LFS deduplicates identical blobs by SHA-256 on the server, so the 7140 files materialize from only ~4987 unique objects (~4.5 GB). The working tree on disk after git lfs pull is ~6.4 GB because each file is checked out as a real copy.

License

Released into the public domain under CC0 1.0. Use, modify, and redistribute the icons, code, and metadata for any purpose without attribution.