Image Kit
nateober/image-kit
Inspeção, upscale, restauro e composição de imagens locais para agentes
Ver no GitHubToolkit local de manipulação de imagens para agentes de IA em Mac, sem cloud ou subscrições. Oferece inspeção (EXIF, faces via Apple Vision), upscaling 2×/4× com Real-ESRGAN, restauração de faces com GFPGAN, remoção de fundo, crops inteligentes e preparação para compartilhamento com strip de metadados.
Instalação
npx convez-ai nateober/image-kit1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Upscaling de frames de vídeo e gerações de IA para alta resolução
- Restauração de faces desfocadas em fotos reais e imagens geradas
- Remoção de background e criação de recortes inteligentes centrados em faces
- Preparação de imagens para compartilhamento com remoção de metadados EXIF/GPS
Veredito

Excelente skill especializado que resolve um nicho real: processamento de imagens local e inteligente para agentes Claude. Bem integrado, com zero setup para maioria das funções e pipeline robusto testado em produção. Recomendado para quem precisa de manipulação de imagens programática sem cloud.
README
A tradução para o português deste README ainda está na fila.
image-kit
A local image-manipulation toolkit for AI agents (and humans) on the Mac — inspect → transform → enhance → verify, no cloud, no subscriptions, no uploads:
- Inspect — one command returns JSON: dimensions, EXIF/GPS, sharpness, brightness, and faces (boxes, quality, yaw) via Apple Vision
- Upscale 2×/4× (Real-ESRGAN) — video frames to print size, AI generations to wallpaper size; ~5s per 1080p frame on Apple Silicon. Doubles as a denoiser (2× up, scale back down)
- Restore faces (GFPGAN) — fixes blurry faces in real photos and janky faces in AI images, with an identity-preserving blend weight
- Face-centered crop — portrait crops to any aspect ratio, centered on a chosen face (Apple Vision)
- Remove background (rembg/u2net) —
transparent-PNG cutouts, zero install via
uvx - Captions & title cards — dependable text compositing (homebrew ffmpeg usually lacks drawtext; this is the workaround made permanent)
- Share prep — strip EXIF/GPS/device metadata, auto-orient, HEIC→JPEG, resize. Stop texting your GPS coordinates to group chats.
Built as an Agent Skill for
Claude Code: ask for "upscale this", "fix
the faces", "crop this to a portrait of the kid on the left", "remove the
background", "are there GPS coordinates in this photo?" — and the agent drives
the right tool, then looks at the output to catch ML artifacts before
handing it back. SKILL.md doubles as a human-readable runbook; every step is
a plain command.
Field-tested on graduation footage: a ~70-pixel face from a 1080p ceremony video → upscale 2× → face restore → clean small-print quality. (Companion skill for the video side: ceremony-clip.)
Requirements
- Apple Silicon Mac (CPU fallback works, just slower)
uv(brew install uv)- ImageMagick for the transform/share recipes (
brew install imagemagick)
Inspection, upscaling, cropping, captions, and background removal need zero
setup — uv resolves dependencies on first run, model weights auto-download
to ~/.cache/image-kit/ (~67 MB per upscale model), and the Vision helper
compiles itself. Face restoration is the one heavyweight: run
scripts/setup_faces.sh once (pinned venv, two upstream gfpgan dependency
bugs patched, 349 MB weights — all documented in the script). Verified
end-to-end on a clean machine: ~2 minutes.
Use it
Install as a skill:
mkdir -p ~/.claude/skills && git clone https://github.com/nateober/image-kit ~/.claude/skills/image-kit
Then, in Claude Code:
Upscale
~/Pictures/gen-042.pngto 4K and fix the faces.
Make a 4:5 portrait of my daughter from this group shot — she's the one in glasses — and clean it up for printing.
Or drive it by hand — SKILL.md has every command, the recommended chains,
and the traps (python -m rembg.cli silently does nothing; -strip before
-auto-orient ships photos sideways; GFPGAN at full strength makes people
look like their own siblings; a face-quality score finds a good face, not
the right face).
License
MIT (the skill and scripts). Model weights retain their own licenses: Real-ESRGAN (BSD-3-Clause), GFPGAN (Apache-2.0), u2net (Apache-2.0).