Images Dataset Studio: build, clean and export image datasets locally

TL;DR

• It is local-first tooling for building, cleaning, reviewing, annotating, versioning and exporting image datasets — and it does not train models.
• The original files stay safe, the automated checks stay explainable, and the final dataset decisions stay with a human reviewer.
• Import covers folders, ZIP, ImageFolder, YOLO, COCO, Pascal VOC, ImageNet and URL manifests, preserving source paths, original bytes, SHA-256 and perceptual hashes, dimensions and EXIF.
• Export covers ImageFolder, COCO, YOLO, JSONL, CSV and Parquet, with immutable versions and diffs.

What it is and who it is for

Images Dataset Studio is a local application for the unglamorous part of computer vision: turning a folder of images into a dataset an annotation team or a trainer can actually use. The README states the boundary clearly — it does not train models, and it does not hide destructive actions inside an opaque pipeline.

It is for the person who has three thousand images and no idea which ones are blurry, duplicated, screenshots, watermarked or corrupt; for the team that needs COCO or YOLO out of images scattered across folders; and for anyone who has lost track of which images were already reviewed.

Import and what is preserved

Import takes folders, ZIP files, ImageFolder, YOLO, COCO, Pascal VOC, ImageNet and URL manifests. Every import preserves source paths, original bytes, SHA-256 hashes, perceptual hashes, dimensions, thumbnails and EXIF metadata. Preserving the perceptual hash is what makes near-duplicate detection possible later without re-reading every file.

Review, duplicates and decisions

The quality review covers blur, exposure, contrast, entropy, resolution, aspect ratio, grayscale, screenshots, watermarks and corrupt files. Findings arrive as issues, not deletions: exact and near duplicates are found without anything being deleted automatically, and every decision is reversible — keep, review, quarantine, reject and restore.

Each image can be inspected with a color RGB histogram, general file facts, adaptive EXIF metadata, labels, quality issues and decisions. That inspector is the answer to "why did this one get flagged", which is the question most dataset tools leave you guessing.

Annotate and generate

Annotation covers classifications, bounding boxes, polygons, masks and keypoints. Generated content — captions, label suggestions, OCR results, detections, masks and PII issues — comes through local plugins or opt-in cloud providers, and text or reference-image search works when the optional embedding provider is installed.

The provider layer is honest by design: Settings → Functions & Models lets an operator choose a local model or a cloud provider independently for pre-annotation, captioning and PII scanning, each function with its own model ID. API keys are write-only in the UI and never returned by the API, and the local server stores them in data/provider-settings.json with owner-only permissions where the platform supports them. Environment variables still work for headless deployments.

Provider status is exposed through GET /api/providers, and an unavailable provider fails the job with an actionable explanation instead of returning fake results.

Versions and export

Versions are immutable, with manifests, checksums, diffs and restore support — so an experiment on a subset never overwrites what you had. Export covers ImageFolder, COCO, YOLO, JSONL, CSV and Parquet.

How to install it

Python 3.11+, uv, Node 20+ or Bun, and macOS, Linux or Windows. SQLite is the default local database, and GPU support is optional, needed only for selected local ML providers.

git clone https://github.com/artificialguybr/images-dataset-studio.git
cd images-dataset-studio
python tools/setup.py
cp .env.example .env       # macOS/Linux; copy manually on Windows
python tools/start.py

The UI opens at http://127.0.0.1:5199 and the API at http://127.0.0.1:8766. On Windows PowerShell: py tools\setup.py, Copy-Item .env.example .env, py tools\start.py. setup.py installs backend dependencies with uv and frontend dependencies with Bun or npm; start.py starts both services and handles Ctrl+C on every OS.

The ids-studio.sh launcher remains available when a single-process FastAPI preview is preferred — set IDS_PORT to change its port. The cross-platform Python launcher is the recommended default because it keeps the API and the Vite development server separate.

There is also an MCP stdio server: start the API, configure your MCP client with mcp-config.example.json replacing its cwd with the repository path, and run uv run python -m backend.mcp_server.

When to use it, when not to

Use it when an image folder needs to become a COCO or YOLO dataset with a paper trail, when a crawl left you with near-duplicates and screenshots mixed in, or when two annotators have been working on the same set and the labels no longer agree.

Do not use it to train or evaluate a model, and do not expect it to make annotation decisions for you — boxes, polygons and masks are yours, and generated captions are suggestions. Do not use it as a labeling platform for a team with per-person quotas either; it manages one dataset and its versions, not a workforce.

Sources

1. Images Dataset Studio — repository — features, requirements, install commands, provider table, MCP setup quoted above.
2. README.md — install and run commands, provider configuration, license (MIT).

Project page: Images Dataset Studio

For the same workflow on the language-model side, see LLM Dataset Studio, and for what an image dataset needs before training a LoRA, automated dataset creation.

The repository is MIT-licensed, and the original files stay untouched, so a bad decision is always reversible.

FAQ

Does it train a model? No. The README says the Studio does not train models; it builds, cleans, reviews, annotates, versions and exports the dataset.

Does it delete duplicates automatically? No. Exact and near duplicates are found and reported, and the decisions keep, review, quarantine, reject and restore are reversible.

Can I use a cloud vision model for captions? Yes, opt-in: Settings → Functions & Models lets each function choose a local model or a cloud provider with its own model ID, and API keys are write-only.

Which export formats are supported? ImageFolder, COCO, YOLO, JSONL, CSV and Parquet.