pixeldump 0.1.2

Gen-Z-flavored AI photo organizer CLI/TUI

  ____  _____  ___  _  _____  ___   _   _  __  __  ____
 |  _ \|_ _\ \/ / |/ \|  _  || _ \ | | | ||  \/  ||  _ \
 | |_) || | \  /| | _ | |_| || | || | | | || |\/| || |_) |
 |  __/ | | /  \| ||_||  _  || |_||_|_|_|||_|  |_||  __/
 |_|   |___/_/\_\_|   |_| |_||___/         |_|     |_|

because your camera roll is a cry for help 🫠

PixelDump is an agentic AI photo organizer that uses vision LLMs to sort your camera roll into intelligently named folders. It deduplicates near-identical shots, groups photos into events, and writes folder names with the energy of a chronically-online friend.

---

✨ what it does

• 📸 Sorts photos into events — uses EXIF dates + GPS clustering, not just "by year"
• 🧠 Vision-LLM classification — Claude Code (no API key), Claude API, or local Ollama (Gemma 3 / LLaVA)
• 💀 Catches duplicates — exact (SHA-256) and near-duplicate (perceptual hash)
• 🏷️ Three naming vibes — corporate, chaotic, or unhinged (your call)
• 📂 Granular document routing — screenshots, bills, boarding passes, prescriptions, and more each land in their own typed folder automatically
• ↩️ Undo support — every move is logged in a manifest, fully reversible
• 🛟 Dry-run by default — preview before anything moves
• 🔁 Resumable — interrupt a run, pick up where you left off
• 🧾 Receipt at the end — stats, easter eggs, and a roast

🚀 quick start

pip install pixeldump
pixeldump run ~/Pictures/CameraRoll

The first run drops you into an interactive wizard. After that, flags are your friends:

# show me what you'd do, don't actually move anything (default)
pixeldump run ~/Photos --dry-run

# yes, do it
pixeldump run ~/Photos --apply

# use claude-code provider (no API key needed)
pixeldump run ~/Photos --provider claude-code --apply

# corporate mode, sass off, batch hard
pixeldump run ~/Photos --mode corporate --sass 0 --apply

# how much will Claude cost me?
pixeldump run ~/Photos --estimate

# undo the last run
pixeldump undo ~/Photos/.pixeldump/manifests/20260509_143022.json

🏷️ naming modes (with real examples)

The vision model looks at the photos and writes the folder name based on what it actually sees. Pick your aesthetic:

Mode	Example output
--mode corporate 🏢	2024_03/tokyo_client_summit/ · 2024_06/sarah_alex_wedding_tuscany/ · 2024_07/team_offsite_q2_2024/
--mode chaotic 🔥	2024_03/ate_in_tokyo_fr/ · 2024_06/main_character_bali_era/ · 2024_07/family_chaos_thanksgiving/
--mode unhinged 🫠	2024_03/proof_i_touched_grass/ · 2024_06/the_yassification_of_brunch/ · 2024_07/my_villain_era_in_milan/

🧠 provider setup

PixelDump supports three vision providers. Run pixeldump setup to configure whichever you want.

Claude Code (recommended — no API key required):

If you already have Claude Code installed and logged in, PixelDump reuses your existing session. Nothing extra to configure.

# confirm claude is on your PATH
claude --version
pixeldump setup   # pick "claude-code"

Claude API (best accuracy, pay-per-use):

export ANTHROPIC_API_KEY=sk-ant-...
pixeldump setup   # pick "claude"

Ollama (fully local, free, slower):

# install ollama from https://ollama.com
ollama pull gemma3
pixeldump setup   # pick "ollama"

Auto-detection order when --provider auto (the default): claude-code → claude API → ollama. Override with --provider claude-code, --provider claude, or --provider ollama.

📁 output structure

After a run, your photo directory looks like this:

Photos/
├── 2024/
│   ├── 2024_03/                        # month folder groups all events from that month
│   │   ├── ate_good_in_tokyo/          # LLM-named event subfolders live inside
│   │   └── cherry_blossom_picnic/
│   ├── 2024_06/
│   │   ├── sarah_alex_wedding_tuscany/
│   │   └── main_character_bali_era/
│   │
│   ├── screenshots/                    # document-type folders sit flat under the year
│   ├── screenshots_apps/
│   ├── screenshots_conversations/
│   ├── screenshots_social/
│   ├── screenshots_memes/
│   ├── screenshots_web/
│   ├── screenshots_maps/
│   │
│   ├── documents_receipts/
│   ├── documents_bills/
│   ├── documents_financial/            # bank statements, credit cards, tax docs, paychecks
│   ├── documents_ids/                  # passports, licences, insurance cards, visas
│   ├── documents_travel/               # boarding passes, hotel confirmations, event tickets
│   ├── documents_medical/              # prescriptions, lab results, vaccine records
│   ├── documents_legal/                # contracts, leases, certificates
│   └── documents_misc/                 # menus, business cards, whiteboards, QR codes
│
├── 2023/
│   └── ...
├── _review/
│   ├── duplicates/        # near-identical shots — highest-res copy kept
│   ├── no_date/           # photos with no EXIF date
│   └── low_confidence/    # photos the AI wasn't sure about (confidence < 0.5)
└── .pixeldump/
    ├── manifests/         # one JSON per run, used for undo
    ├── runs/              # per-run logs
    └── thumbnails/        # cached thumbnails

Events land in YYYY/YYYY_MM/event_name/ so each month is its own browsable folder. Documents and screenshots sit flat under the year (YYYY/documents_receipts/) since they accumulate all year and don't need month grouping.

🧰 all flags

pixeldump run TARGET [OPTIONS]

  --provider [claude-code|claude|ollama|auto]  vision LLM provider (default: auto)
  --mode [corporate|chaotic|unhinged]  folder name vibe (default: chaotic)
  --sass [0-3]                         personality intensity (default: 2)
  --burst-hours INT                    event clustering gap (default: 72h)
  --batch-size INT                     photos per LLM batch (default: 5)
  --concurrency INT                    parallel hashing workers (default: 4)
  --include-videos                     also sort .mp4/.mov by date
  --skip-duplicates                    don't run dup detection
  --skip-clustering                    classify each photo individually
  --output PATH                        sort into a different dir (copy mode)
  --dry-run / --apply                  preview vs. execute (default: dry-run)
  --quiet                              minimal output, no live dashboard
  --no-wizard                          skip the interactive wizard
  --estimate                           print cost estimate and exit

pixeldump undo MANIFEST [--yes]        reverse a previous run
pixeldump status TARGET                show the most recent manifest
pixeldump config [--reset]             show or reset config
pixeldump setup                        re-run provider setup

💸 cost

Claude Code: free — runs inside your existing Claude Code subscription, no extra charges.

Claude API: rough rule of thumb ~$0.001 per photo, since PixelDump batches photos by event and sends one classification call per cluster (not per photo). A 1,500-photo library typically costs under $2. Use --estimate to see the exact number for your library before running.

Ollama: free, just slower. Bring your GPU energy.

🔒 your photos stay yours

• Claude Code reuses your local CLI session — 512px thumbnails are passed to the local claude process, same data boundary as using Claude Code interactively
• Local Ollama sees nothing leave your machine
• Claude API receives 512px thumbnails (not full-resolution photos), and Anthropic does not train on API inputs
• Nothing is ever deleted. Files are moved, never removed. Every operation is logged and reversible.
• Config files containing API keys are written with owner-only permissions (mode 0600)

🛠️ development

git clone https://github.com/your-handle/pixeldump
cd pixeldump
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pre-commit install

make test           # pytest
make lint           # ruff
make typecheck      # mypy --strict

Architecture: see docs/ARCHITECTURE.md. Contributing: see .github/CONTRIBUTING.md.

🤝 contributing

Issues and PRs welcome. Look for good first issue labels. The sass engine especially welcomes new lines — see src/pixeldump/assets/sass_lines.json.

📜 license

MIT. See LICENSE.

---

built by me because my photos folder was a crime scene