Your own image studio.
Free, local, honest.
Text to image, editing, sketch to image, upscaling, background removal and image to text, with permissively licensed models, on the machine in front of you. Nothing is uploaded. Nothing is paid for.
Quality is a setting, never a consequence
Most local tools quietly give you less when your machine is small. This one does not. The rules are written down and the software follows them.
The default is the best your machine can do
Full resolution, full steps. The machine decides how long it takes, not how good it looks. Before you commit, the studio tells you the expected time, measured on your own hardware.
Faster only with your consent
balanced and fast presets exist, and each states its trade-off in words. Nothing is downgraded silently, ever. If a setting will not fit in memory, the studio says so instead of pretending.
Permissive licences only
FLUX.1 schnell, Qwen-Image, Qwen-Image-Edit, Real-ESRGAN, IS-Net, Qwen2.5-VL. Apache-2.0, MIT or BSD throughout. Models with non-commercial or revenue-gated terms are excluded by policy and listed as such.
Your machine, your files
Downloads are pinned by SHA-256 and asked for first. After setup nothing leaves the computer. Every finished PNG carries its full recipe, so any image can be reproduced exactly.
One studio, every step
A local web app, a command line, and an MCP server for Hivemind or any other agent that speaks MCP. Same engine, same policy.
Create
Prompt parameters Midjourney-style: --seed 7 --size 1024x768 --preset fast --n 4 --lora inkstyle:0.8. Live preview while it renders. A load meter that shows the machine's limit and an explicit unlock past it.
Edit and Variations
Change a picture by instruction with Qwen-Image-Edit, repaint only a painted mask, remove an object, or re-roll at a chosen distance. Strength defaults sit below the measured cliff on fast models.
Sketch to image
Brush, eraser, shapes, fill, eyedropper, a faded reference layer, pan and zoom, pen pressure. Draw the layout, describe it, let the model paint.
Upscale and cut out
Real-ESRGAN 4× and 16×. Background removal with IS-Net or U²-Net to a transparent PNG in seconds.
Image to text
Captions, alt text, tags and "write the prompt that would make this", with a local vision model. Optional auto-captions make the gallery searchable.
Settings that tell the truth
Benchmark once and the estimates calibrate. CPU or Vulkan engine, LoRA folder, your own models labelled as yours, every component with its licence, size and an Install, Show and Remove.
Measured, not promised
A 16 GB laptop with no discrete GPU: Intel i5-1235U, Intel UHD integrated graphics, Windows 11, the same FLUX.1 schnell job while the owner kept working. These are the numbers the studio itself showed before each run.
"a lighthouse on a rocky coast at dusk", 512 px, 4 steps, seed 7. Identical output from both engine builds below.
| Engine build | Wall time | To first step | Peak memory |
|---|---|---|---|
| CPU | 30 min 5 s | 27 min 57 s | 6.99 GiB |
| Vulkan on the integrated GPU | 12 min 43 s | 7 min 57 s | 5.25 GiB |
2.4× faster with the Vulkan build, even on integrated graphics. Full-resolution best renders on this machine take about two hours on the CPU build; the studio says so before you press Generate. A desktop with a discrete GPU is minutes, not hours.
Install
A Windows app, or Python 3.11 or newer on any OS. Everything else is downloaded by the studio itself, with your consent, hash-verified.
Windows app
ImageStudio-x.y.z-windows-x64.zip → unzip → ImageStudio.exe
From the latest release. A real window of its own; close it during a render and it offers to keep rendering in the tray. Not code-signed yet: SmartScreen asks once per download ("More info" → "Run anyway"). The SHA-256 is next to the download.
With pipx or pip
pipx install hivemind-imagestudio imagestudio ui
Opens the studio in your browser. Everything still runs on your machine. Background removal needs one more package: pipx install "hivemind-imagestudio[bg]", or install it later from Settings.
First run
imagestudio doctor imagestudio install-engine --yes imagestudio download-model flux-schnell --yes
Or do the same from Settings in the app: engine, models and tools each have an Install button and a getting-started checklist on Home.
As an MCP server
{ "mcpServers": { "imagestudio":
{ "command": "imagestudio", "args": ["serve"] } } }Tools: generate, plan, job_status, job_result, jobs, capabilities, download_model, variations, edit, upscale, remove_background, describe.
Official builds are the GitHub Releases of the repository and the hivemind-imagestudio package on PyPI. On corporate Windows machines, endpoint protection may ask you to allow the downloaded engine the first time it runs; it is the pinned stable-diffusion.cpp release and its hash is in the source.
Licences and the name
The code is free. The name is not. Both are deliberate.
Code: Apache-2.0
Use it, fork it, ship it, in products and at work. Keep the NOTICE file and the author lines, as the licence asks.
Models and engines
stable-diffusion.cpp and llama.cpp are MIT. FLUX.1 schnell, Qwen-Image, Qwen-Image-Edit, their encoders and the vision models are Apache-2.0. Real-ESRGAN is BSD-3-Clause; IS-Net and U²-Net are Apache-2.0.
Name: a trademark
"Hivemind", "Image Studio by Hivemind" and the logo identify the official project. Forks are welcome under their own name and may say "based on Hivemind Image Studio". The policy is in TRADEMARK.md.