OpenFill 2.0.0-beta.6

AMD graphics cards. I don't have an AMD card, so this is untested on real AMD hardware; please tell me how it goes on Issues.

Added

• On Windows, Download ComfyUI gets ComfyUI's own version for AMD cards (ComfyUI_windows_portable_amd.7z, ROCm 7.2.1) for a Radeon RX 7000 or RX 9000, Radeon PRO W7000, or Ryzen AI 300 or Max chip, from the same ComfyUI release and checked against its size and SHA-256 like the NVIDIA one. An older Radeon gets ComfyUI's AMD instructions and I already have ComfyUI.
• On Linux, a ComfyUI with PyTorch for ROCm works as any other ComfyUI folder. For an RX 6700 or 6600 (and the other RDNA 2 cards except the 6800 and 6900) OpenFill starts ComfyUI with HSA_OVERRIDE_GFX_VERSION=10.3.0, and for a Radeon 780M or 760M with 11.0.0, unless you set it yourself.
• The Models tab reads an AMD card's name and graphics memory without AMD's tools, so Recommended for this PC fits it. On an AMD card the GGUF models say they have been reported to give only noise on some Radeon cards and are left out of Recommended.
• The steps are in docs/install.md, which now also says where the model manager is (Filters > OpenFill > Generative Fill..., Models tab).

Updating from beta.5: install it the same way. Your settings and models stay where they are.

File	SHA-256
openfill-2.0.0-beta.6.zip	d448c4c0986f91d8f97069f66219c46b2a7fafcbc80b26186002d9bef5954ccd

OpenFill for PhotoCraft 0.2.2

AMD graphics cards, as asked for on Reddit. I don't have an AMD card (both my laptops are NVIDIA), so this is untested on real AMD hardware. If you try it, please tell me how it went on Issues, good or bad.

If you have 0.2.1, download this zip and run Install OpenFill for PhotoCraft.cmd again (Linux: sh install.sh). Your settings and models stay.

Added

• Download ComfyUI for AMD cards (Windows). With a Radeon RX 7000 or RX 9000, Radeon PRO W7000, or a Ryzen AI 300 or Max chip, Download ComfyUI gets ComfyUI's own version for AMD cards (ROCm 7.2.1, 1.6 GB) instead of the NVIDIA one. It comes from the same ComfyUI release and is checked against its size and SHA-256 the same way. For an older Radeon (RX 6000 and earlier) the panel says to follow the AMD steps in the install guide and choose the folder.
• AMD on Linux. A ComfyUI with PyTorch for ROCm works through Choose folder, and How to install opens the AMD steps. For an RX 6700 or 6600 OpenFill sets HSA_OVERRIDE_GFX_VERSION=10.3.0 when it starts ComfyUI, as ComfyUI's README says.
• Models knows AMD cards. It shows the card and its memory ("RX 7900 XT · 20 GB VRAM · ...") without AMD's tools. On an AMD card the GGUF models (FLUX.1 Fill GGUF, Qwen Image Edit) say they have been reported to give only noise on some Radeon cards. LaMa, MAT, SD 1.5 and SDXL are plain PyTorch and should be fine.
• The README says where the model manager is: the gear icon in the OpenFill panel, Models section.

Install on Windows

1. Install PhotoCraft 0.12 to 0.16 (the .msi from its releases page; the newest, 0.16.0, is fine).
2. Install Python from python.org if you don't have it (the default options are fine).
3. Download openfill-photocraft-0.2.2.zip from the bottom of this page, unzip it and double-click Install OpenFill for PhotoCraft.cmd.
4. Open PhotoCraft with OpenFill from the Start menu or the desktop.
5. The first time, press Download ComfyUI (or Choose folder if you have it), then Open model manager and get LaMa (0.2 GB) and DreamShaper 8 Inpainting (2 GB).

On Linux: unzip it and run sh install.sh (on Fedora first sudo dnf install python3-pillow python3-tkinter). The full steps, AMD included, are in docs/install.md.

SHA-256 of openfill-photocraft-0.2.2.zip:

fb278e55f33f890a4de55fe046f09bb2a47843489a22423455811a2d539e5618

All changes are in photocraft/CHANGELOG.md. Released under Apache-2.0.

Bug reports and feature requests are very welcome on Issues!

OpenFill 2.0.0-beta.5

A small fix-up for beta.4.

Fixed

• Remove with LaMa failed the first time on a ComfyUI that did not have OpenFill's node pack yet (one you set up yourself, say), with "needs the custom node pack(s) openfill-nodes". OpenFill now puts its own node pack in place before ComfyUI starts, so the first Remove works.

Updating from beta.4: install it the same way. Your settings and models stay where they are.

File	SHA-256
openfill-2.0.0-beta.5.zip	0076430ffaed17d765697b925c4baa883ea86f178fbc4f8be53b2e299496391f

OpenFill 2.0.0-beta.4

Remove object got a lot sharper on big photos, there are two new models made just for erasing, and OpenFill now runs in PhotoCraft too.

What's new

• Remove object at full resolution. LaMa used to squeeze every area into a 256 px square, so removing something from a large photo came back soft. OpenFill now comes with its own small ComfyUI node that runs LaMa at the image's own pixels. A 2700 x 2000 px selection on a 2900 px photo was cleared at full size in 12 seconds on my 8 GB laptop card. The node installs itself into ComfyUI the next time OpenFill starts it.
• Two erasers to try. RORem (Apache-2.0, 6.9 GB), an SDXL model trained to remove objects, and the FLUX.1 Fill object removal LoRA (0.3 GB). The LoRA runs on whichever FLUX.1 Fill you already have. The LoRA itself is MIT, but FLUX.1 Fill is non-commercial, so it is for non-commercial use only.
• Any model can remove. The model you pick for Remove object is used as picked. LaMa stays the default, and MAT, the erasers and every fill model are in the menu.
• The Models list says what each model can do, like "Fill, expand, create, remove" or "Remove (erase-trained)".
• New home. The repo is now beckettharriman/openfill. Old links still work. It also holds OpenFill for PhotoCraft, which uses the same engine and models, so a ComfyUI and models you already have work for both.
• Apache-2.0. OpenFill is now under the Apache License 2.0 (it was GPL-3.0).
• Fixes. A removal's layer is now always called Object Removal (with RORem it was named after its trigger words). A busy ComfyUI that didn't answer for a moment no longer fails the job with "The AI backend is not available". ComfyUI in ~/AI/ComfyUI is found too.

Install

1. Install GIMP 3.2 if you don't have it. Open it once, then close it.
2. Download openfill-2.0.0-beta.4.zip from the bottom of this page and unzip it.
3. Add the plug-in to GIMP:
   • Windows: double-click Install OpenFill.cmd in the unzipped folder. Or drag the gimp-openfill folder into %APPDATA%\GIMP\3.2\plug-ins.
   • Mac and Linux: set up ComfyUI once in Terminal first (copy-paste steps in docs/install.md). Then run sh install.sh in the unzipped folder.
4. Restart GIMP and open Filters > OpenFill > Generative Fill....
5. On Windows, if ComfyUI isn't found, click Download ComfyUI and OpenFill sets it up for you (about 2 GB). If you already have ComfyUI, pick I already have ComfyUI.
6. When the model manager opens, download LaMa (0.2 GB) and DreamShaper 8 Inpainting (2 GB). The first request takes about a minute to start. After that it's seconds.

Updating from an earlier beta? Do the same steps. Your settings and models stay where they are.

Good to know

• It's a beta. The new Remove was tested in GIMP 3.2 on Windows with an 8 GB laptop card.
• I haven't compared RORem and the FLUX LoRA against LaMa on real photos yet. Try them and tell me what you see.
• FLUX is slow: the removal LoRA took about 4 minutes on an 8 GB laptop card.
• FLUX and Qwen fills, Generative Expand and Text to Image still work at about 1024 px and are scaled up, so on big images they come back softer. Remove is not affected.

SHA-256 of openfill-2.0.0-beta.4.zip:

87c8f93e0d5dbd7ce5c7b3fec4fc7e2f9822c096cee97c972cc55f5bc04ff16f

All changes are in CHANGELOG.md. Released under Apache-2.0.

Bug reports and feature requests are very welcome on Issues!

OpenFill for PhotoCraft 0.2.1

A small fix-up of the first release. If you have 0.2.0, download this zip and run Install OpenFill for PhotoCraft.cmd again (Linux: sh install.sh). Your settings and models stay.

Fixed

• Keep could lose the variations. Clicking in PhotoCraft (to deselect, say) while Keep was still working made it fail with "select.modify.feather is not available right now: no selection", and the variations were gone. Keep now makes the kept layer from the job's own selection first and only then takes the variations away, so a click can't get in the way. If Keep still fails, the variations stay as they were and you can press Keep again. The same goes for the result a Fill, Remove or Expand adds at the end.
• Remove failed the first time on a ComfyUI you already had. With a ComfyUI picked through Choose folder, the first Remove with LaMa stopped with "needs the custom node pack(s) openfill-nodes; press Install nodes in the Models tab" until another job had started ComfyUI. OpenFill now puts its own node pack in place before ComfyUI starts, as it already did for the ComfyUI it downloads for you.

Install on Windows

1. Install PhotoCraft 0.12 to 0.16 (the .msi from its releases page; the newest, 0.16.0, is fine).
2. Install Python from python.org if you don't have it (the default options are fine).
3. Download openfill-photocraft-0.2.1.zip from the bottom of this page, unzip it and double-click Install OpenFill for PhotoCraft.cmd.
4. Open PhotoCraft with OpenFill from the Start menu or the desktop.
5. The first time, press Download ComfyUI (or Choose folder if you have it), then Open model manager and get LaMa (0.2 GB) and DreamShaper 8 Inpainting (2 GB).

On Linux: unzip it and run sh install.sh (on Fedora first sudo dnf install python3-pillow python3-tkinter). The full steps are in docs/install.md.

Everything in the 0.2.0 notes still applies.

SHA-256 of openfill-photocraft-0.2.1.zip:

d6db8d004eb2d0d3aebb00f198098747e9b73445eb2b419471f0963dcdad0692

All changes are in photocraft/CHANGELOG.md. Released under Apache-2.0.

Bug reports and feature requests are very welcome on Issues!

OpenFill for PhotoCraft 0.2.0

The first release of OpenFill for PhotoCraft: Generative Fill, Remove object, Generative Expand and Text to Image in PhotoCraft, running on your own graphics card through ComfyUI. Free, local and open source. Nothing leaves your computer.

It uses the same engine as OpenFill for GIMP, so the models and the model manager are the same, and a ComfyUI and models you already have work for both.

What you get

• Fill. Select an area, type what you want, press Enter. With no prompt it continues the surroundings. Up to 5 variations, with thumbnails to compare them in PhotoCraft, then Keep, Keep all or Discard.
• Remove. Select it and press Enter. No prompt. LaMa erases it at the image's own resolution. MAT, RORem, the FLUX.1 Fill removal LoRA and every fill model can be picked too.
• Expand. Drag the frame's edges, or pick a ratio (1:1, 5:4, 4:3, 3:2, 16:9, landscape or portrait), and it paints the new edges.
• Create. Text to Image as a new image, or as a layer that fills your selection.
• One step in the history. Every result is a new layer with a mask, and one Undo takes a fill or a whole expand back. Your selection and the clipboard are left alone.
• A panel next to PhotoCraft, with a live preview while it works. It folds into a small strip, and on Windows it hides when PhotoCraft is minimized. Filter > Plug-ins > OpenFill brings it back.
• Download ComfyUI on Windows: the first time, the panel downloads and sets up ComfyUI for you (about 2 GB), checked against its SHA-256.

Install on Windows

1. Install PhotoCraft 0.12 to 0.16 (the .msi from its releases page; the newest, 0.16.0, is fine).
2. Install Python from python.org if you don't have it (the default options are fine).
3. Download openfill-photocraft-0.2.0.zip from the bottom of this page, unzip it and double-click Install OpenFill for PhotoCraft.cmd.
4. Open PhotoCraft with OpenFill from the Start menu or the desktop.
5. The first time, press Download ComfyUI (or Choose folder if you have it), then Open model manager and get LaMa (0.2 GB) and DreamShaper 8 Inpainting (2 GB).

On Linux: unzip it and run sh install.sh (on Fedora first sudo dnf install python3-pillow python3-tkinter). The full steps are in docs/install.md.

Good to know

• It's a first release, and it works with PhotoCraft 0.12 to 0.16. I tested it end to end on Windows 11 with PhotoCraft 0.12.0 and an RTX 4070 Laptop GPU (8 GB), and on Linux (Fedora 44, RTX 2070 Max-Q) with PhotoCraft 0.16.0. macOS is not supported yet.
• You need an NVIDIA card with 8 GB or more for the SDXL and FLUX models. LaMa and the SD 1.5 models run on less. On the processor it's very slow.
• Rough times on the 8 GB laptop card: LaMa about 5 s, RealVisXL 25 to 45 s, FLUX.1 Fill about 3 minutes.
• FLUX and Qwen fills, Expand and Create work at about 1024 px and are scaled up, so on big images they come back softer. Remove is not affected.
• Pixels travel as 8-bit PNG files, so on a 16-bit document the new layer is 8-bit.
• Open PhotoCraft with "PhotoCraft with OpenFill". A PhotoCraft opened another way can't talk to OpenFill, and the panel tells you to reopen it.
• MAT and everything that runs on FLUX.1 Fill are non-commercial. The licenses are in docs/models.md.
• OpenFill is not affiliated with PhotoCraft.

SHA-256 of openfill-photocraft-0.2.0.zip:

69ccce79596b9f44f5232872b5e9d928b29e44b3fc0778dd0c6eae328777b28d

All changes are in photocraft/CHANGELOG.md. Released under Apache-2.0.

Bug reports and feature requests are very welcome on Issues!

OpenFill 2.0.0-beta.3

A few of you asked whether OpenFill can use models you already have. Now it can.

What's new

• Use your own models. OpenFill now finds the SD 1.5 inpainting and SDXL checkpoints in your ComfyUI checkpoints folder, and FLUX.1 Fill files in unet or diffusion_models, even ones it doesn't know, and puts them in the model menu marked "your file". No download, no setup.
• Safe to point at a messy folder. It only reads .safetensors and .gguf files, and only their headers, to work out what each model is. Pickled .ckpt files are never opened. Files it can't use (regular SD 1.5, SDXL inpainting, FLUX.1 dev, SD 2) are skipped, and the log says why.
• Your model gets the right setup. A local SDXL checkpoint runs with the same Fooocus inpaint patch as the SDXL models in the list, and a local FLUX.1 Fill file uses the same text encoders and VAE (OpenFill downloads those once if they're missing).
• Pick models from the search in the dialog. Models you got through the Hugging Face and Civitai search can now be chosen right in the Fill, Expand and Text to Image dialogs, not only set as active in the Models tab.

Install

1. Install GIMP 3.2 if you don't have it. Open it once, then close it.
2. Download openfill-2.0.0-beta.3.zip from the bottom of this page and unzip it.
3. Add the plug-in to GIMP:
   • Windows: double-click Install OpenFill.cmd in the unzipped folder. Or drag the gimp-openfill folder into %APPDATA%\GIMP\3.2\plug-ins.
   • Mac and Linux: set up ComfyUI once in Terminal first (copy-paste steps in the README, under macOS or Linux). Then run sh install.sh in the unzipped folder.
4. Restart GIMP and open Filters > OpenFill > Generative Fill....
5. On Windows, if ComfyUI isn't found, click Download ComfyUI and OpenFill sets it up for you (about 2 GB). If you already have ComfyUI, pick I already have ComfyUI.
6. When the model manager opens, download LaMa (0.2 GB) and DreamShaper 8 Inpainting (2 GB). The first request takes about a minute to start. After that it's seconds.

Updating from an earlier beta? Do the same steps. Your settings and models stay where they are.

Stuck? The README has the full steps and troubleshooting.

Good to know

• It's a beta. Tested end to end on Windows and Linux (including the GIMP Flatpak), and on macOS (Apple M3) in beta.1.
• An NVIDIA graphics card is recommended. 8 GB runs every default model. Apple silicon works but is slower. Without a supported card, SD 1.5 runs on the processor in minutes.
• FLUX is slow: about 2 to 3 minutes per fill on an 8 GB laptop card. Qwen is slower still.
• Not tested yet: AMD cards and Intel Macs. Your own Qwen-Image-Edit files aren't picked up yet.
• Replacing an object is hard for FLUX. In a scene that suggests the old object, like a car-sized spot on a road, it often paints a new one back, even after removing the old one first. Qwen-Image-Edit is better at swaps, but much slower.

SHA-256 of openfill-2.0.0-beta.3.zip:

9b363c9c0c8b8e99b3ae004b84533b2d132d5f605c4ac9a07f45739358961462

All changes are in CHANGELOG.md. Released under GPL-3.0.

Bug reports and feature requests are very welcome on Issues!

OpenFill 2.0.0-beta.2

Thanks to everyone who tried the first beta and told me what went wrong. This release fixes the things you ran into most: older graphics cards that failed every job, OpenFill quietly running on the processor, FLUX jobs that looked stuck, and FLUX prompts that painted the wrong things.

What's new

• Older NVIDIA cards work. On a GTX 10-series card (like the GTX 1060), every job failed right away with "The GPU ran out of memory". It wasn't memory: the ComfyUI that OpenFill downloads only runs on RTX 20 cards and newer. OpenFill now checks your card and downloads the ComfyUI version made for older cards. If you already have ComfyUI and it can't use your card, OpenFill tells you which version to get instead.
• Real error messages. "Out of memory" now only shows up when it's actually memory, and it includes ComfyUI's own error text.
• It tells you when ComfyUI is on the processor. If ComfyUI isn't using your graphics card, OpenFill now says so, and why. Every job would be many times slower. Settings > Check backend also shows what ComfyUI is running on.
• Open log folder. A new button in Settings opens OpenFill's log and ComfyUI's own log, so you can see what the hidden ComfyUI is doing.
• Live previews for every model. The progress window shows the image forming while it works, including the long FLUX jobs. The small preview files (about 5 MB) come with the models. If you downloaded a model before this release, OpenFill gets its preview file in the background before your next job.
• FLUX for 12 GB cards. A sharper FLUX.1 Fill download (Q6_K) for cards with 12 GB or more. The Q4 version stays the pick for 8 GB.
• A FLUX prompt tip. FLUX paints everything you name, so the Fill dialog now shows a short tip when a FLUX model is picked.
• Fixed: stuck on the processor. If OpenFill set up ComfyUI on a PC without an NVIDIA card and you later pointed it at another ComfyUI, it kept running on the processor. Now it switches back to your graphics card.

Install

1. Install GIMP 3.2 if you don't have it. Open it once, then close it.
2. Download openfill-2.0.0-beta.2.zip from the bottom of this page and unzip it.
3. Add the plug-in to GIMP:
   • Windows: double-click Install OpenFill.cmd in the unzipped folder. Or drag the gimp-openfill folder into %APPDATA%\GIMP\3.2\plug-ins.
   • Mac and Linux: set up ComfyUI once in Terminal first (copy-paste steps in the README, under macOS or Linux). Then run sh install.sh in the unzipped folder.
4. Restart GIMP and open Filters > OpenFill > Generative Fill....
5. On Windows, if ComfyUI isn't found, click Download ComfyUI and OpenFill sets it up for you (about 2 GB). If you already have ComfyUI, pick I already have ComfyUI.
6. When the model manager opens, download LaMa (0.2 GB) and DreamShaper 8 Inpainting (2 GB). The first request takes about a minute to start. After that it's seconds.

Updating from beta.1? Do the same steps. Your settings and models stay where they are.

Stuck? The README has the full steps and troubleshooting.

Good to know

• It's a beta. Tested end to end on Windows and Linux (including the GIMP Flatpak), and on macOS (Apple M3) in beta.1.
• An NVIDIA graphics card is recommended. 8 GB runs every default model. Apple silicon works but is slower. Without a supported card, SD 1.5 runs on the processor in minutes.
• FLUX is slow: about 2 to 3 minutes per fill on an 8 GB laptop card. Qwen is slower still.
• Not tested yet: AMD cards and Intel Macs.
• Replacing an object is hard for FLUX. In a scene that suggests the old object, like a car-sized spot on a road, it often paints a new one back, even after removing the old one first. Qwen-Image-Edit is better at swaps, but much slower.

SHA-256 of openfill-2.0.0-beta.2.zip:

60fb768c5a216add2628190fcd30037804c58c3a17d054e1a5b083494b1319e0

All changes are in CHANGELOG.md. Released under GPL-3.0.

Bug reports and feature requests are very welcome on Issues!

OpenFill 2.0.0-beta.1

The thing I missed most after switching from Photoshop to GIMP 3 was Generative Fill and the rest of Adobe's AI editing tools. So I wrote a free plug-in that adds them. It runs locally through ComfyUI on your own graphics card, and nothing leaves your computer.

This is the first public beta.

What you get

• Generative Fill: select an area, type what should be there, get up to eight variations as new layers.
• Remove object: select something and it's erased. No prompt needed.
• Generative Expand: make the canvas bigger and the scene continues into the new space.
• Text to Image: a new layer or a new image from a prompt.

Install

1. Install GIMP 3.2 if you don't have it. Open it once, then close it.
2. Download openfill-2.0.0-beta.1.zip from the bottom of this page and unzip it.
3. Add the plug-in to GIMP:
   • Windows: double-click Install OpenFill.cmd in the unzipped folder. Or drag the gimp-openfill folder into %APPDATA%\GIMP\3.2\plug-ins.
   • Mac and Linux: set up ComfyUI once in Terminal first (copy-paste steps in the README, under macOS or Linux). Then run sh install.sh in the unzipped folder.
4. Restart GIMP and open Filters > OpenFill > Generative Fill....
5. On Windows, if ComfyUI isn't found, click Download ComfyUI and OpenFill sets it up for you (about 2 GB). If you already have ComfyUI, pick I already have ComfyUI.
6. When the model manager opens, download LaMa (0.2 GB) and DreamShaper 8 Inpainting (2 GB). The first request takes about a minute to start. After that it's seconds.

Stuck? The README has the full steps and troubleshooting.

Or let an AI agent do it

Paste this into an AI coding agent on your computer:

Install OpenFill, a plug-in for GIMP 3, on my computer. Read the Install section of https://raw.githubusercontent.com/beckettharriman/gimp-openfill/gimp3-port/README.md and follow the steps for my OS, using openfill-2.0.0-beta.1.zip from https://github.com/beckettharriman/gimp-openfill/releases/tag/v2.0.0-beta.1. Trust the README over anything you already know about the project. Ask me before installing ComfyUI or downloading any models.

Good to know

• It's a beta. Tested end to end on Windows, Linux (including the GIMP Flatpak) and macOS (Apple M3).
• An NVIDIA graphics card is recommended. 8 GB runs every default model. Apple silicon works but is slower (50 to 90 s per fill on an M3). Without a supported card, SD 1.5 runs on the processor in minutes.
• Not tested yet: AMD cards, Intel Macs, FLUX and Qwen.
• The models still have rough edges. SD 1.5 often ignores the prompt on small selections (pick an SDXL model in the dialog), and very large removals and expansions get weaker. The full list is in the README.

SHA-256 of openfill-2.0.0-beta.1.zip:

3937f6d080091c2fc83392186d22f9e53865bddffe61a689a6c5562c03a2df01

All changes are in CHANGELOG.md. Released under GPL-3.0.

I hope someone finds these tools useful. Bug reports and feature requests are very welcome on Issues!