GIMP 3 AI Plugins

A collection of Python-based AI plugins for GIMP 3 (GNU Image Manipulation Program), using the GIMP 3 Python API via PyGObject introspection.

Plugins

Plugin	What it does	How it works	Dependencies
Background Remove (bgremove/)	Removes backgrounds from images using AI	Exports layer → runs backgroundremover CLI → imports result as new transparent layer	backgroundremover (PyPI)
AI Upscale (upscale/)	Upscales images 2x/4× using AI upscalers	Exports layer → runs PyTorch upscaler (3 backends) → imports result as new upscaled layer	torch, pillow, image_gen_aux / diffusers
Depth Map (depthmap/)	Creates a depth map layer or Stereo 3D photos (L-R/Cyan-Magenta/Facebook 3D) from the active layer	Exports layer → runs Depth Anything V2 in-process (Small / Base / Large) → imports result as new depth layer; optional Facebook 3D photo export	depth-anything-v2, torch, torchvision, opencv-python, matplotlib
AI Edit (aiedit/)	Edits images using text prompts via a diffusion model and vision LLM	Exports layer → runs sd-cli with a diffusion GGUF model + vision LLM → imports edited result as new layer	sd-cli (stable-diffusion.cpp), diffusion GGUF, vision LLM GGUF, VAE
SD Server (sd-server/)	Generates/edits images via stable-diffusion.cpp server API	Connects to sd-server HTTP API; supports txt2img, img2img, and Kontext/Boogu/Qwen-style reference editing	sd-server (stable-diffusion.cpp), requests

Security

Arbitrary command execution. Plugins for most applications can run any commands. This is not new. Open source relies on the community to find and fix bugs and exploits. You should always review new source code or test in a secure environment. Read our repo notes on why a sandbox is not secure.

Don't download from untrusted places. The official location for these plugins is https://github.com/themanyone/gimp-plugins

Requirements

• GIMP 3.0+ — these plugins use the GIMP 3 Python API (gi.require_version('Gimp', '3.0'))
• Python 3 — GIMP 3 ships with its own Python environment
• GPU Acceleration — e.g. CUDA strongly recommended for the upscale, image edit plugins.

Installation

Installer. The install.py installer (experimental) will attempt to install missing dependencies and put things in the right places. If something is not right, check Manual Install steps below & report issues.

3. Restart GIMP

After restart, plugins appear in the GIMP menus:

• AI Image: File → Create → AI Image...
• Background Remove: Filters → AI → Remove Background...
• AI Upscale: Filters → AI → Upscale...
• AI Edit: Filters → AI → AI Edit...
• Depth Map: Filters → AI → Depth Map...
• SD Server: Filters → AI → SD Server...

Keyboard Shortcuts. Go to Edit -> Keyboard Shortcuts and type the name of the plugin.

Manual Install

Developers & power users may want fine-grained control & knowledge of install locations here:

AI Edit

The AI Edit plugin uses sd-cli from stable-diffusion.cpp (not a Python package). See aiedit/README.md for model downloads and setup.

Others. Other plugins have their own Python dependencies. The installer will attempt to install these dependencies with pip.

# For Background Remove
pip install backgroundremover

# For AI Upscale
pip install image_gen_aux diffusers torch pillow

# For SD Server
pip install requests

# For Depth Map
pip install --no-deps depth-anything-v2 torch pillow
pip install opencv-python matplotlib  # only if not already installed

2. Place plugins in GIMP's plug-ins directory

The installer looks for the highest-numbered GIMP/xx.x/plug-ins directory. Install whichever plugins you need. If you have sd-server running, you might be able to get by with just bgremove, depthmap, and sd-server plugins. With Qwen Image 2.1 (default), and careful prompting, sd-server can technically do it all, removing backgrounds, color separations, halftones, sharpening photos that have been scaled-up the old way & making depth maps/3D photos, but it's slower & experimental. Not likely to be supported.

git clone https://github.com/themanyone/gimp-plugins.git
cd gimp-plugins
ln -srf bgremove ~/.config/GIMP/3.2/plug-ins/
ln -srf sd-server ~/.config/GIMP/3.2/plug-ins/
# install additional plugins the same way

The plugins import shared modules from the repo root: uihelpers.py for all of them, hfhub.py where models are resolved, and depthmap's dibr.py (already inside its own directory). Symlink the repo-root ones next to the plugins too, otherwise Python cannot find them and GIMP silently disables the plugin:

ln -sfr uihelpers.py hfhub.py ~/.config/GIMP/3.2/plug-ins/

python install.py does this for you, and it repairs links left pointing at a moved checkout. If plugins stop appearing after moving the repo, re-run the installer; if a plugin that failed to load earlier keeps not showing up, remove ~/.config/GIMP/3.2/pluginrc so GIMP re-queries its plug-ins.

Plugin architecture

Every plugin follows the same pattern:

export layer as temp PNG → process (CLI or ML) → load result → new layer

Shared hfhub.py (repo root): reusable huggingface_hub helpers for any plugin that needs to locate/download model checkpoints — resolve_model() checks the HF cache, adopts local disk models into it, then downloads. See AGENTS.md for the full architecture guide and development notes.

Shared uihelpers.py (repo root): keyboard/focus behavior for the custom dialogs — wire_text_dialog() makes Ctrl+Enter submit from any field, Tab and Shift+Tab move between input fields in visual order, and the prompt is focused when the dialog opens. submit_on_ctrl_enter() adds Ctrl+Enter submit to dialogs without a prompt.

SD Server

The SD Server plugin connects to a running sd-server instance from stable-diffusion.cpp via its HTTP API. It supports:

• txt2img — generate new images from a text prompt
• img2img — edit the current layer using a diffusion model
• Batch — generate several candidates per run (set the Batch field next to the sampler); extra results become layers of the result so you can compare and keep the best
• Reference-based editing — Qwen Image 2.1, Kontext, Boogu, Z-Image models (uses extra_images for reference conditioning)

Start the server first read their docs for more info, then use these plugins from GIMP:

# Qwen Image 2.1 Turbo with the Qwen3-VL 8B vision LLM (same files as AI Edit below)
sd-server \
  --diffusion-model ~/Downloads/src/ComfyUI/models/diffusion_models/qwen-image-2.1-turbo-Q4_K_M.gguf \
  --vae ~/Downloads/src/ComfyUI/models/vae/qwen_image_2.1_vae_bf16.safetensors \
  --llm ~/models/Qwen3-VL-8B-Instruct-UD-IQ3_XXS/Qwen3-VL-8B-Instruct-UD-IQ3_XXS.gguf \
  --llm_vision ~/models/Qwen3-VL-8B-Instruct-UD-IQ3_XXS/mmproj-F16.gguf \
  --diffusion-fa --offload-to-cpu -v

Swap in any other diffusion model, VAE and text encoder.

The plugin auto-detects the served model and available features via /sdcpp/v1/capabilities. The model name is shown in the dialog.

API Compatibility. Works with many servers that implement the stable-diffusion API:

Endpoint	Purpose
/sdapi/v1/txt2img	Text-to-image generation
/sdapi/v1/img2img	Image-to-image / reference editing
/sdcpp/v1/capabilities	Model & feature detection
/sdapi/v1/progress	Sampling progress for the GIMP progress bar (optional; servers without it just show activity)

OpenAI-compatible servers. The plugin can also connect to any OpenAI-compatible /v1/images/generations endpoint (not just sd-server) — just point the Server URL field to any compatible API.

Multi-layer edit (object & style transfer). Operates on layers of an image. First, load reference images as layers. To open an image as a layer in GIMP, go to File > Open as layers… from the top menu or press Ctrl + Alt + O on your keyboard. Or paste selections from other tabs. For best results, scale layers so they are the same size on the long dimension, and a multiple of 32. Either 1024 or 512 for fast processing. Depending on the system, larger images produce better quality but may take a long time to edit, or run out of memory. Experiment.

With a reference-capable model (Qwen Image 2.1, Kontext, Boogu, Z-Image), tick Multi-layer edit in the dialog to send layers to the server as separate reference images. Each checked layer is labeled positionally (image 1, image 2, ...) in layer-stack order (topmost = image 1); the layer being edited is always included. Rename layers (F2) to identify them in the list, then refer to them by number in the prompt:

• Object transfer: check the couch layer and the room layer, then prompt Put the couch in image 1 into the room in image 2.
• Style transfer: check the style source and the target, then prompt Render image 2 in the style of image 1.

Placing objects needs near-full Init Strength and Guidance above 1.0, so enabling multi-layer edit raises them (to 0.9 and 3.0) automatically — adjust as needed.

Background Remove

• This works with existing tools.
• It assumes you have installed backgroundremover.

Don't want to use backgroundremover?

• Install any old command-line AI tool to remove backgrounds from images.
• You can use pip install rembg for example.
• Edit bgremove/bgremove.py and modify the command list to use your favorite CLI tool (e.g., rembg instead of backgroundremover).

AI Upscale

Try small images first. Then try 1024x1024 pixels to upscale to 4K resolution. Larger images will take a lot longer. And without enough VRAM upscale will fail.

The Upscale plugin prompts you to choose an AI model when you use the plugin in Gimp. And it allows you to save preferences. Models are downloaded automatically, so it might take some time to upscale your first image.

It is not necessary to edit upscale/upscale.py to:

• Change DEFAULT_MODEL or any key in MODEL_CONFIGS
• Add new upscaling models with appropriate config
• Tune inference parameters (tile_size, steps, noise_level, etc.) ...but you can!

Depth Map

The Depth Map plugin creates a depth map layer (white = near, black = far) (or the reverse of that) from the active layer using Depth Anything V2, running in-process like upscale. Find it under Filters -> AI.

Requirements. pip install --no-deps depth-anything-v2 torch pillow huggingface_hub plus opencv-python and matplotlib. Install with --no-deps because the package pins an old torchvision combo. A checkpoint (depth_anything_v2_vit[sbl].pth) is also needed; the Small (~98 MB), Base (~390 MB), and Large (~1.3 GB) variants match three model sizes, with the model cache avoiding reloads between runs. GPU (CUDA) is used when available, with a CPU fallback.

Model download location. Our plugins may attempt to automatically download AI models. Newer apps let Huggingface Hub manage model storage in a standard way. Our plugin finds models in the Hugging Face cache (~/.cache/huggingface/hub, honoring HF_HOME) via huggingface_hub. If you have already downloaded models manually, or through outdated apps, our plugin will automatically find them & tell huggingface_hub where they are, so apps like ours won't re-download them yet again in the future.

Simply pick a model (small, base, or large) and output mode (grayscale, colored inferno, or both) in the dialog and click OK.

Alternatively, tick Export as 3D photo (Facebook) to also write <name>.jpg plus <name>_depth.jpg for upload to Facebook.

Depth of field (Lens Blur). To simulate depth of field, Duplicate your image layer (Ctrl-D), putting it on a new tab. Then generate an inverted-grayscale depth map (black = near, white = far), with Invert (for lens blur) ticked. Switch tabs back to the original image & run Filters → Blur → Lens Blur, choosing the depth map as the "aux" layer (set the Radius higher, e.g. 5) so the far background blurs while the near subject stays sharp. If the blur lands on the wrong parts of the image, use Colors -> Curves or Burn tool on the Depth Map tab, to blacken the area(s) to be focused. See GIMP Lens Blur example.

Troubleshooting. First run can take a while because the checkpoint downloads. The plugin resolves the checkpoint from the Hugging Face cache first, silently adopts any model it finds on disk into that cache, and only downloads when nothing else matches (disable Download if missing to skip that). To fetch a checkpoint manually, use hf download depth-anything/Depth-Anything-V2-Small. Depth Anything V1 checkpoints (depth_anything_vitl14.pth) are not supported. If a model file is missing the plugin says so; install errors usually mean depth-anything-v2 was installed with its pinned old torchvision. If the 3D photo parallax looks inverted, rerun with Invert (for lens blur).

AI Image

The aiimage plugin calls sd-cli from the command line. It's a bit more complicated to use. But it offers more flexibility and of course it doesn't require a server. Find the AI Image dialog under File -> Create.

Edit aiimage.py and set MODELS_PATH and LLM_PATH to where your models are. Or just enter the full path to each model into the dialog at runtime and save preferences. We are reusing our ComfyUI models but yours will be somewhere else.

Qwen Image 2.1 Turbo (default).

Text-to-image with the same models as AI Edit (see below). It only needs 8 steps at CFG scale 1.0 with the Turbo sigma schedule. Download them with the hf download commands in the AI Edit section; the defaults are:

DEFAULT_DIFFUSION_MODEL = MODELS_PATH + "/diffusion_models/qwen-image-2.1-turbo-Q4_K_M.gguf"
DEFAULT_LLM = LLM_PATH + "/Qwen3-VL-8B-Instruct-UD-IQ3_XXS.gguf"
DEFAULT_LLM_VISION = LLM_PATH + "/mmproj-F16.gguf"
DEFAULT_VAE = MODELS_PATH + "/vae/qwen_image_2.1_vae_bf16.safetensors"

Z Image Turbo (alternative).

Z Image Turbo is for fast image generation in 4-6 steps. Enter these model locations into the dialog and save a preset:

DEFAULT_DIFFUSION_MODEL = MODELS_PATH + "/diffusion_models/z-image-turbo-Q5_K_M.gguf"
DEFAULT_LLM = MODELS_PATH + "/text_encoders/qwen_3_4b.safetensors"
DEFAULT_LLM_VISION = ""
DEFAULT_VAE = MODELS_PATH + "/vae/ae.safetensors"

AI Edit

Like aiimage the aiedit plugin calls sd-cli from the command line. It's good for simple, single-layer edits. Find the AI Edit plugin under Filters -> AI.

Scale images down to 256x256 or lower for testing AI Edit. Editing large images takes longer. Default model paths may be edited into the top of aiedit/aiedit.py or enter the paths into the dialog at run time.

Qwen Image 2.1 Turbo (default).

This is the default AI Edit model. It edits the current image from a text prompt, using the source image as a reference (-r) together with the Qwen3-VL vision LLM. It is a distilled model, so it only needs 8 steps at CFG scale 1.0; the plugin selects reference editing and the fixed Turbo sigma schedule automatically.

Download the four files with the Hugging Face CLI (pip install -U "huggingface_hub[cli]", or curl -LsSf https://hf.co/cli/install.sh | bash):

# Diffusion model: Qwen Image 2.1 Turbo, Q4_K_M GGUF
hf download unsloth/Qwen-Image-2.1-Turbo-GGUF qwen-image-2.1-turbo-Q4_K_M.gguf \
  --local-dir ~/Downloads/src/ComfyUI/models/diffusion_models

# Vision LLM and its mmproj projector: Qwen3-VL 8B
hf download unsloth/Qwen3-VL-8B-Instruct-GGUF \
  Qwen3-VL-8B-Instruct-UD-IQ3_XXS.gguf mmproj-F16.gguf \
  --local-dir ~/models/Qwen3-VL-8B-Instruct-UD-IQ3_XXS

# VAE
hf download Comfy-Org/Qwen-Image-2.1 vae/qwen_image_2.1_vae_bf16.safetensors \
  --local-dir ~/Downloads/src/ComfyUI/models

These are the defaults at the top of aiedit/aiedit.py. Set MODELS_PATH and LLM_PATH for your machine, or enter the paths in the dialog:

DEFAULT_DIFFUSION_MODEL = MODELS_PATH + "/diffusion_models/qwen-image-2.1-turbo-Q4_K_M.gguf"
DEFAULT_LLM = LLM_PATH + "/Qwen3-VL-8B-Instruct-UD-IQ3_XXS.gguf"
DEFAULT_LLM_VISION = LLM_PATH + "/mmproj-F16.gguf"
DEFAULT_VAE = MODELS_PATH + "/vae/qwen_image_2.1_vae_bf16.safetensors"

Pick the Default preset in the dialog to load these, along with 8 steps, CFG scale 1.0 and the Turbo sigma schedule.

Boogu Edit.

This is a Kontext model. Kontext mode should be set automatically, and Guidance will be set to 0.0. You can change them but it is not recommended. You can also create new images with these. This model has worked well for adding characters and text to images.

DEFAULT_DIFFUSION_MODEL = MODELS_PATH + "/diffusion_models/boogu-edit-dit-Q4_0.gguf"
DEFAULT_LLM = LLM_PATH + "/Qwen3-VL-8B-Instruct-UD-IQ3_XXS.gguf"
DEFAULT_LLM_VISION = LLM_PATH + "/mmproj-F16.gguf"
DEFAULT_VAE = MODELS_PATH + "/vae/ae.safetensors"

Flux 2 Klein.

This is an older & smaller img2img model. But it should work. Img2img mode should be set, and Guidance set to 3.5. Download these models and enter their locations into the image editing dialog and save it.

diffusion model: flux-2-klein-4b-Q4_K_S.gguf
clip_l: clip_l.safetensors
llm: Qwen3-4B-Instruct-2507-UD-Q4_K_XL.gguf
vae: flux2-vae.safetensors"

Customizing.

Presets are saved in ~/.config/gimp-plugins/aiedit/presets.json but you can load, save, and rename settings at runtime. You may also modify these plugins to show different defaults if you want.

License

This software is Copyright (2026) by Henry Kroll III (@themanyone) under GNU General Public License v3.0

This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.

This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
[GNU General Public License](LICENSE) for more details.

You should have received a copy of the GNU General Public License
along with this program.  If not, see <https://www.gnu.org/licenses/>.

For any other uses, contact the author.

Author

Henry Kroll III (@themanyone)

Thanks for trying out gimp-plugins!

• GitHub https://github.com/themanyone
• YouTube https://www.youtube.com/themanyone
• Mastodon https://mastodon.social/@themanyone
• Linkedin https://www.linkedin.com/in/henry-kroll-iii-93860426/
• Buy me a coffee https://buymeacoffee.com/isreality
• TheNerdShow.com https://thenerdshow.com