Self-hosted service
Woolverine94/biniou avatar
Woolverine94/biniou

biniou pins one git dependency to a commit and floats the other

a self-hosted webui for 30+ generative ai

1,154 stars135 forksPythonGPL-3.0

At a glance

What is it?
Woolverine94/biniou is a GPL-3.0 self-hosted webui that covers image, audio, video and chat generation from 8GB of RAM with no dedicated GPU. Its single release is a 0.0.1 pre-release from May 2024, its Dockerfile clones whatever is on main at build time, and its requirements file pins two packages that install into the same namespace while leaving fifteen unpinned.
Who is it for?
biniou fits someone who wants one local interface covering many generative tasks, is willing to manage model files and disk, and treats a breaking dependency change as something to expect rather than avoid. Four things to check before you install it.
Can I use it commercially?
Yes, with conditions. GPL-3.0 is a copyleft licence: if you distribute software that includes it, you must release that software's source code under the same licence. Running it internally without distributing it does not trigger that obligation.
Is it still maintained?
Yes. The repository last received commits 1 day ago.
What is it written in?
Mainly Python, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 21, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • Woolverine94/biniou1,154 ★ · +2
Star history
Repository2026-09-112026-11-21
Woolverine94/biniou1.2k1.2k

2 Stars in the last 71 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

Weekly updates, and one release tagged 0.0.1 in May 2024

The project ships on a weekly cadence that is visible in the README itself. The five most recent entries are dated 2026-09-26, 2026-09-19, 2026-09-12, 2026-09-05, and 2026-08-29, each marked as a weekly update, and each listing specific model additions and module changes. The 2026-09-12 entry added automated support for any Bartowski, mradermacher, and Unsloth GGUF models to the chatbot module, which is the kind of change that turns a per-model code edit into a general rule.

Against that, the repository has exactly one release: v0.0.1, marked as a pre-release, dated 2024-05-29.

So the release history and the update history describe different things. The last push is 2026-09-30, more than two years after the only tag, and the version file at the repository root is a separate matter again from the release number.

The practical consequence is that there is no version to pin. Anyone deploying biniou in a repeatable way is tracking the default branch and reading the weekly entries to find out what changed, which is why the older entries are archived to a wiki page rather than accumulating in the README.

The documentation follows the same pattern. The README is an index that points at a wiki for documentation and a showroom, and three third-party video walkthroughs are linked, including one Windows install tutorial.

The Dockerfile clones main, so the image is not reproducible from a tag

The container build starts from debian:bookworm-slim, installs a long apt line covering bash, git, python3, python3-venv, ffmpeg, libavformat-dev, gcc, perl, make, and libtcmalloc-minimal4, then creates a user called biniou with a disabled password and switches to it.

The repository is then fetched like this:

`git clone --branch main https://github.com/Woolverine94/biniou.git`

followed by running ./install.sh and making webui.sh executable. The service is exposed on port 7860, the Gradio default, and the entrypoint is the webui.sh script.

The branch name is the problem. A clone of main inside a build means two builds on different days produce different trees, and neither matches the only tag, which is a 0.0.1 pre-release from May 2024. The base image is a floating tag as well, so the layer beneath is unpinned too.

That is a normal choice for a project that wants the latest model support on every rebuild, and it is exactly the choice that makes an incident hard to trace. If you need a repeatable image, build it once, tag that image yourself, and record which README update entry it was built from.

The default image is CPU-only, and the CUDA conversion is commented out

The Dockerfile contains a pair of lines that are present in the file and inactive in the build. The comment reads Replace pyTorch cpu-only version by CUDA-enabled one, and the two commands beneath it would create a flag file, activate the environment, uninstall torch, torchvision, torchaudio, and llama-cpp-python, and run update_cuda.sh.

So the published image installs the CPU build of PyTorch. A second badge in the README points at a separate CUDA-docker-image workflow, which is where GPU support actually gets built.

That arrangement is consistent with what the README promises, since it says you can run this on your own computer even without a dedicated GPU and starting from 8GB RAM. The CPU path is the default path, and the CUDA path is an extra.

The repository carries the same split in its scripts. There is update_cuda.sh for NVIDIA and update_rocm.sh for AMD, plus a CUDA/ directory at the root and a Dockerfile at the root. The ROCm script has no corresponding badge, no workflow link, and no entry in the platform list, so the AMD path exists as a script without a documented route to reaching it.

Two opencv packages install into the same namespace

The requirements file pins opencv-contrib-python at 4.8.1.78 and, seventeen lines later, lists opencv-python with no version at all.

Both packages install the cv2 module. They are alternatives, not complements, and installing both into one environment is the classic way to end up with an import that resolves to whichever one pip placed last. The contrib build is the one with the extra modules, so a user who wanted contrib and got the plain build would lose features silently rather than get an error.

The rest of the file mixes pinning styles in a way that makes this harder to see. Roughly half the entries carry exact versions: numba 0.60.0, onnx 1.15.0, pyarrow 14.0.0, datasets 2.14.6, huggingface-hub 0.34.4, gradio 3.50.2, transformers 4.51.3, accelerate 1.9.0, diffusers 0.34.0, numpy 1.25.2, audiocraft 1.2.0, onnxruntime 1.20.0, mediapipe 0.10.21, peft 0.17.1, timm 1.0.19, torchmetrics 1.3.2, and starlette 0.52.1 among them.

The rest do not, and the unpinned set includes packages that change shape often: sentencepiece, scipy, opencv-python, imageio, imageio-ffmpeg, gfpgan, optimum, insightface, controlnet-aux, tomesd, ffmpeg-python, trimesh, exiv2, music-tag, and gguf.

One entry is older than the update cadence it sits in: tb-nightly is pinned to a 2023 build, 2.16.0a20231101. And gradio is held at 3.50.2, a major version behind where the web framework now sits, which is worth knowing before you look for a Gradio-specific fix.

One git dependency has a commit pinned, the other has nothing

Two requirements are installed straight from GitHub rather than from an index, and they are handled differently.

PhotoMaker is pinned to a specific commit: git+https://github.com/TencentARC/PhotoMaker.git@ec35a8716ea790a404cb7e1ba282a320454906d1. That is a full commit hash, so a rebuild gets exactly that tree.

Real-ESRGAN is installed with no reference at all: git+https://github.com/ai-forever/Real-ESRGAN.git. With no tag or branch named, pip installs whatever the default branch points at that day.

So one vendored dependency is reproducible and the other floats, inside the same requirements file, for the same kind of reason.

That matters more than usual here, because install.sh is run inside the Docker build as well as on a workstation. A floating default branch means the image can gain or lose a package between builds with nothing in the project's own versioning to record the change. If you need a repeatable environment, resolve both to commits yourself before you build.

Five distro families, four Windows scripts, and one AMD path with no route

Platform support is organised as five Linux families, one Windows path, one macOS path, and Docker. The Linux families are OpenSUSE, the RHEL family covering Rocky, Alma, CentOS Stream, and Fedora, Arch by way of CachyOS, Mandriva, and the Debian family covering Debian, two Ubuntu releases, two Linux Mint releases, and Pop OS.

Each of the five has a matching OCI install script at the repository root: oci-arch.sh, oci-debian.sh, oci-mandriva.sh, oci-opensuse.sh, and oci-rhel.sh.

Windows is handled by a different naming convention entirely. There is install_win.cmd, update_win.cmd, update_win_cuda.cmd, venv.cmd, and webui.cmd, plus a win_installer/ directory, against three extensionless shell scripts on the other platforms, install.sh, update.sh, and webui.sh. Six entry points for Windows against three for everything else.

macOS is offered as Intel and marked experimental, with a Homebrew install, which means there is no documented Apple silicon path.

The update scripts split three ways: plain, CUDA, and ROCm. Only two of those three have a documented route.

One more detail of the layout: __init__.py sits at the repository root next to webui.py, so the repository itself is shaped as a Python package, and version is a file rather than a variable in it.

Editorial conclusion

biniou fits someone who wants one local interface covering many generative tasks, is willing to manage model files and disk, and treats a breaking dependency change as something to expect rather than avoid. Four things to check before you install it. Which platform you are on, since macOS is labelled experimental for Intel only, and the AMD path lives in an update script with no documented install route. How you pin your build, because the Docker image clones the main branch at build time and cannot be reproduced from a tag. Whether your environment already has an opencv package, since the requirements list installs both a contrib build and the plain one into the same namespace. And how much disk the models will take, since offline operation depends on downloading every model you intend to use.

Frequently asked questions

What is biniou and what does it run on?

It is a self-hosted webui for generative AI that lets you generate multimedia content and use a chatbot on your own computer, even without a dedicated GPU and starting from 8GB RAM. Once deployed and the required models are downloaded, it can work offline.

Which platforms does biniou support?

Five Linux families with a matching OCI script each, namely OpenSUSE, the RHEL family, Arch through CachyOS, Mandriva, and the Debian family, plus Windows 10 and 11, macOS on Intel which is marked experimental, and Docker. The container is built from debian:bookworm-slim and exposes port 7860.

Does biniou publish releases?

There is one, v0.0.1, marked as a pre-release and dated 2024-05-29, while the last push to the repository is dated 2026-09-30. Updates are recorded as dated weekly entries in the README, with older entries archived to a wiki page.

Is the biniou Docker image pinned to a version?

No. The Dockerfile clones the main branch from GitHub during the build and then runs ./install.sh, so the tree is whatever is on that branch at build time. It installs the CPU build of PyTorch, with the CUDA replacement steps present but commented out.

Which GPU backends does biniou have scripts for?

update_cuda.sh for NVIDIA and update_rocm.sh for AMD, alongside the plain update.sh, plus a CUDA/ directory. The CUDA path also has a dedicated image workflow, while the ROCm script has no documented install route in the README.

Where is the biniou documentation and the installers?

Documentation and a showroom live in the project wiki rather than in the repository. Installers are install.sh for Linux and macOS, install_win.cmd plus a win_installer/ directory for Windows, and the Dockerfile for containers.

Official sources

  1. Issues
  2. License: GPL-3.0
  3. README
  4. Releases
  5. Woolverine94/biniou on GitHub
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/woolverine94-biniou.svg)](https://hysenlabs.com/projects/woolverine94-biniou)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/woolverine94-biniou.svg)](https://hysenlabs.com/projects/woolverine94-biniou)

Was this useful?