OpenFill: Free Generative Fill for GIMP 3.2 (Photoshop Alternative With Local ComfyUI)

Photoshop’s Generative Fill is one of the most useful AI features in any image editor, but it needs an Adobe subscription and sends your images to the cloud. OpenFill brings the same idea to free, open-source editors. It adds Generative Fill, Remove Object, Generative Expand and Create to GIMP 3.2 and to the Rust editor PhotoCraft, and runs everything on your own GPU through a hidden ComfyUI backend. This guide covers what OpenFill can do, how to install it on Windows, Linux and macOS, which models to download and exactly where they go, which custom nodes it needs, and how to get the best results on 4 GB to 16 GB graphics cards.

What Is OpenFill?

OpenFill is a plug-in that puts an AI panel inside your image editor. You select an area, type what you want, and the result appears on a new layer. Under the hood it builds a ComfyUI workflow, sends it to a local ComfyUI server, and brings the image back. You never see ComfyUI. It starts when the editor opens, stays running while you work, and stops when you close the editor.

It is a GIMP 3 port of Torben Giesselmann’s older gimp-stable-boy plug-in, and its inpainting graphs are built on Acly’s comfyui-inpaint-nodes, the same work behind the popular Krita AI Diffusion plug-in. Both builds shipped on October 26, 2026 and are still marked beta.

Privacy is a big part of the appeal. The ComfyUI backend only listens on 127.0.0.1, so nothing outside your PC can reach it. The only network traffic is model search and the downloads you ask for, and every download is checked by size and SHA-256.

The Four Tools

Every result lands on a new layer in one history step, so your original pixels are never painted over. Undo works as normal. In GIMP the tools are under Filters > OpenFill. In PhotoCraft they are in a panel next to the canvas and under Filter > Plug-ins > OpenFill.

What You Need

• Editor: GIMP 3.2 or newer (Windows, Linux or macOS), or PhotoCraft 0.12 to 0.16.
• Graphics card: 8 GB or more for the SDXL and FLUX models, from NVIDIA or AMD Radeon RX 7000 or newer. LaMa and the SD 1.5 models run on less. Apple Silicon works. Without a usable GPU it runs on the CPU, which takes minutes per fill.
• ComfyUI: on Windows, OpenFill can download it for you (about 2 GB, 7 GB unpacked). On Linux and macOS you set it up once, or point OpenFill at the one you have.
• Disk space for models: from 0.2 GB (LaMa) to about 13 GB (FLUX.1 Fill Q6).
• PhotoCraft only: Python 3.10 or newer with Pillow. ComfyUI’s own Python works.

Install OpenFill for GIMP 3.2

Windows

1. Install GIMP 3.2, open it once, then close it so your profile folder exists.
2. Download openfill-2.0.0-beta.6.zip from the OpenFill releases page, unzip it, and double-click Install OpenFill.cmd. Or drag the gimp-openfill folder into %APPDATA%\GIMP\3.2\plug-ins.
3. Restart GIMP and open Filters > OpenFill > Generative Fill...
4. If ComfyUI is not found, click Download ComfyUI. It checks your GPU, disk space and port 8188, downloads the official portable build into %LOCALAPPDATA%\OpenFill\ComfyUI, verifies its SHA-256, and adds the inpaint nodes and OpenCV. Git is not needed. If you already have ComfyUI, click I already have ComfyUI instead.
5. In the model manager (the Models tab), download LaMa (0.2 GB) and DreamShaper 8 Inpainting (2 GB).

The first request takes about a minute while the backend starts. After that it is much faster. Keep ComfyUI near the drive root, such as C:\AI\ComfyUI. Deep folder paths can cause missing-file errors at startup on Windows.

Linux (including the GIMP Flatpak)

Start GIMP once so the profile exists, set up ComfyUI if you do not have it (see the macOS steps below for the commands), then from the unzipped OpenFill folder run:

sh install.sh --comfy-dir ~/ComfyUI

The script finds your GIMP 3 profile, including the Flatpak one at ~/.config/GIMP/3.2. It was tested on Fedora with the GIMP 3.2.6 Flatpak, where ComfyUI starts from inside the sandbox and stops when GIMP closes.

macOS (Apple Silicon)

The Download ComfyUI button is Windows-only, so set up ComfyUI once in Terminal. The macOS system Python is too old, so install Python 3.13 from python.org or with brew install python@3.13:

git clone https://github.com/comfyanonymous/ComfyUI.git ~/ComfyUI
cd ~/ComfyUI
python3.13 -m venv venv
venv/bin/pip install torch torchvision torchaudio
venv/bin/pip install -r requirements.txt opencv-python-headless

Then run sh install.sh in the unzipped OpenFill folder (it finds ~/ComfyUI on its own), restart GIMP, and open Filters > OpenFill > Generative Fill.... On an M3 MacBook Air, a 512 px fill with DreamShaper 8 took 50 to 90 seconds once warm, and Remove with LaMa took about 5 seconds.

Install OpenFill for PhotoCraft

PhotoCraft is a newer open-source image editor written in Rust. OpenFill is not affiliated with it.

1. Install PhotoCraft (the .msi on Windows, or the DMG on macOS).
2. Unzip openfill-photocraft-0.2.4.zip. On Windows, double-click Install OpenFill for PhotoCraft.cmd. On Linux and macOS, run sh install.sh.
3. Open PhotoCraft with OpenFill from the Start menu, desktop or Applications.
4. When the panel asks, press Download ComfyUI (Windows) or Choose folder and pick the folder that has main.py in it (or your portable folder).
5. Press Open model manager and get LaMa and DreamShaper 8 Inpainting.

If you do not want a second Python install, point OpenFill at ComfyUI’s Python before running the installer. On macOS or Linux, for example:

OPENFILL_PYTHON="$HOME/ComfyUI/.venv/bin/python" sh install.sh

On Windows portable ComfyUI, set OPENFILL_PYTHON to C:\ComfyUI_windows_portable\python_embeded\python.exe in Command Prompt, then run the installer from the same window.

Use Your Existing ComfyUI

If you already use ComfyUI, you can share its models instead of downloading them twice:

1. Find the folder you start ComfyUI from. A portable install has both python_embeded and ComfyUI. A manual install has main.py and a venv or .venv folder.
2. In GIMP choose I already have ComfyUI (or change it later in Generative Fill Settings). In PhotoCraft use Choose folder, or later the gear icon, Engine, Browse.
3. Open Models. OpenFill lists the supported files it finds: SD 1.5 inpainting, SDXL and FLUX.1 Fill checkpoints in .safetensors or .gguf.

Three warnings from the docs:

• Do not point it at Krita AI Diffusion’s managed server. OpenFill may install nodes or packages that affect Krita. Use a separate ComfyUI.
• ComfyUI Desktop (the older Electron layout) is not supported by Choose folder. Use a portable or manual ComfyUI.
• Models stored only through extra_model_paths.yaml are not all detected yet.

Which Models to Download (and Where They Go)

OpenFill ships no models. Its model manager recommends weights that fit your GPU and disk and downloads them into the right ComfyUI folders. If you prefer to download by hand, or want to know what is on your disk, these are the exact files and folders from OpenFill’s model catalogue:

The SD models also use tiny preview decoders in ComfyUI/models/vae_approx/ (taesd_decoder.pth, taesdxl_decoder.pth, taef1_decoder.pth). The model manager adds these for you. If you use the Windows automatic install, ComfyUI/ means %LOCALAPPDATA%\OpenFill\ComfyUI\ComfyUI inside the portable build.

The default speeds on the author’s 8 GB RTX 4070 Laptop GPU:

Custom Nodes OpenFill Needs

The model manager installs the right custom nodes at pinned versions when a model needs them, so most people never touch this. For reference, and for manual installs:

To install the two public packs by hand (on a portable build, use python_embeded/python.exe -m pip instead of pip):

cd ComfyUI/custom_nodes
git clone https://github.com/Acly/comfyui-inpaint-nodes.git
git clone https://github.com/city96/ComfyUI-GGUF.git
pip install -r ComfyUI-GGUF/requirements.txt
pip install opencv-python-headless

OpenCV matters: without it, Generative Expand paints the new border from a flat colour. OpenFill installs it in the background the first time a fill needs it, and tells you.

Best Setup for Your GPU

• 4 to 6 GB (GTX 1660, RTX 3050, older laptops): LaMa for removal and DreamShaper 8 Inpainting or Realistic Vision V6 Inpainting for fills. These are SD 1.5 models, so keep selections small to medium.
• 8 GB (RTX 3060 Ti, 4060, 4070 Laptop): add RealVisXL V5 with the Fooocus patch for much better photo fills. FLUX.1 Fill Q4 works but is slow at this size.
• 12 GB (RTX 3060 12 GB, 4070): FLUX.1 Fill Q6 or Qwen-Image-Edit-2511 Q4 for the most prompt-accurate fills and expands.
• 16 GB (RTX 4080, 5080): everything above runs comfortably, including the full FLUX.1 Fill safetensors loaded as fp8.
• AMD: RX 7000 or newer. On Windows the Download ComfyUI button fetches ComfyUI’s AMD build automatically. Support is new, with one RX 7800 XT report so far.
• Mac: M-series with 16 GB works, but larger fills are slow. Fill big areas in a few smaller pieces.

Tips for Better Results

• Use Remove, not Fill, to delete things. LaMa needs no prompt and works at the image’s full resolution, so it stays sharp.
• Use SDXL or FLUX for small or flat areas. SD 1.5 models often ignore the prompt on flat areas and small selections.
• Select a little extra. Including some surrounding pixels helps the model blend lighting and texture.
• Generate several variations. Keep the best of up to five.
• Expand in steps. Expanding by a moderate amount, then again, looks more natural than one huge expand.
• Watch resolution. FLUX, Qwen, Expand and Create work at about 1024 px and are scaled up afterwards, so big images can come back softer in the filled area. The author says higher-resolution support is in progress.

OpenFill vs Photoshop Generative Fill

Photoshop is still smoother and more polished, especially at high resolution. OpenFill wins on cost, privacy and choice, and for object removal LaMa is already excellent.

Limitations and Licences

• Beta software. Expect rough edges. The author says most of them are in the model output rather than the plug-in.
• Soft results on large images with FLUX, Qwen, Expand and Create, which work at about 1024 px. One Reddit user saw a blurred patch after removing a car using the RealVisXL demo at that size.
• FLUX is slow on 8 GB cards.
• Licences: OpenFill itself is Apache-2.0. Each model keeps its own licence. FLUX.1 Fill and MAT are non-commercial, so do not use them for paid work. LaMa, Qwen-Image-Edit-2511 and RORem are Apache-2.0. The SD 1.5 and SDXL models use CreativeML OpenRAIL licences.

FAQ

Is OpenFill free?

Yes. OpenFill is Apache-2.0 and all its models are free downloads. It runs on your own hardware, so there are no credits or subscriptions.

Does OpenFill work offline?

Yes, once the models are downloaded. ComfyUI runs locally on 127.0.0.1 and only goes online when you search for or download a model.

Which GIMP version do I need?

GIMP 3.2 or newer. The plug-in uses GIMP 3’s Python and needs no extra packages. It does not support GIMP 2.10.

Where does OpenFill put the models?

In your ComfyUI’s model folders, for example ComfyUI/models/inpaint/big-lama.pt for LaMa and ComfyUI/models/checkpoints/ for DreamShaper and RealVisXL. The full list is in the table above.

Can I use it with an AMD GPU?

Yes, Radeon RX 7000 or newer. The Windows installer fetches ComfyUI’s AMD build automatically.

Which model is best for removing objects?

LaMa. It is tiny (0.2 GB), needs no prompt, runs in about 5 seconds and keeps the image at full resolution.

Final Thoughts

OpenFill finally gives GIMP users a real Generative Fill, and it does it the right way: local, private and free. The two starter models cover most everyday edits on almost any GPU, and you can step up to RealVisXL, FLUX or Qwen when your card allows. It is still beta and soft at very high resolutions, but for removing objects, fixing backgrounds and extending photos, it is already worth installing today.

More guides:

• ComfyUI-RMBG 3.3: Canvas Studio and Video Background Removal
• Best Open Source Text-to-Image Models in 2026
• How Much VRAM Do You Need for ComfyUI?

Sources

• GitHub: beckettharriman/openfill (README, docs/install.md and models/catalog.json, read October 31, 2026)
• ComfyUI Wiki: OpenFill Brings ComfyUI Generative Fill to PhotoCraft and GIMP
• GitHub: Acly/comfyui-inpaint-nodes
• GitHub: PhotoCraft