Image edit · macOS · Windows · Linux

OmniGen2 EditScore

This program edits an image from an instruction, on your GPU. It uses OmniGen2 with the EditScore7B-v1.1 LoRA: OmniGen2 after more training with the EditScore-7B reward model. Download the program for your computer. Each file that you download contains the weights. The program does not connect to other computers.

macOS

Apple silicon · 5.3 GB

Download for macOS

Apple did not notarize this program. Unzip the file. Remove the quarantine flag one time. Then start the program:

cd omnigen2-editscore-macos-arm64
xattr -dr com.apple.quarantine .
./omnigen2-editscore

SHA-256 05d03a8a0412ac2f5203f8a7d7924d3a9d683b4a86982402f02008a5e6e2635d

Windows

x86-64 · NVIDIA · 8.1 GB

Download for Windows

Unzip the file. Start omnigen2-editscore.cmd.

SHA-256 808e61634ea2ecefe6ff5ca7f1e43c69a6e04c8d0cc786133a71e5b4e9691808

Linux

x86-64 · NVIDIA · 9.1 GB

Download for Linux

mkdir omnigen2-editscore
tar xf omnigen2-editscore-linux-x86_64.tar -C omnigen2-editscore
cd omnigen2-editscore
./omnigen2-editscore

SHA-256 926842080cdb1ed5200ad0d95c5394a2958eef7d4ca9b683ff714d63abcf3ef5

Weights

In each desktop program · 4.6 GB

The folder w of each desktop program contains the weights: OmniGen2 from chibifire/omnigen2-base-df5dca8a, with the LoRA from OmniGen2-EditScore7B-v1.1 in them. They are 4-bit weights. The program does not download files.

All checksums: SHA256SUMS. The files are in chibifire/omnigen2-editscore-desktop.

Edit an image

Give an image and an instruction

In the browser

Start the program with no option. The first start unpacks the Python environment into the folder env. Then the program loads the model and opens the edit page in your browser. Give an image and an instruction, then click Edit. The page and the edit stay on this computer.

From a terminal

Edit one image and save the result as a PNG file.

omnigen2-editscore --image in.png \
  --prompt "Change the background to a classroom." \
  --out edit.png

--steps N	Steps of the sampler (default 50)
--seed N	Seed of the noise (default 0)
--max-side N	Longest side of the result in pixels (default 1024)
--device	auto, cuda, mps or cpu

Example

Change the background to classroom

Edit by OmniGen2 with the EditScore7B-v1.1 LoRA

Linux package, NVIDIA RTX 3090, 50 steps, seed 0, 672 × 1024 pixels, approximately 2 minutes for each edit. The input image is from the OmniGen2 repository.

Status

Test results

Linux package on an NVIDIA RTX 3090 (Linux in WSL), with no connection to other computers. An edit was complete after 119 seconds (50 steps, 672 × 1024 pixels).
Windows package on an NVIDIA RTX 3090. We stopped all connections from the program to other computers. An edit from the terminal was complete after 121 seconds. We gave the image and the instruction on the edit page in Microsoft Edge, then clicked Edit. The page showed the edit after 122 seconds. The browser loaded all 62 files of the page from this computer.
macOS on an Apple M2 Pro with 32 GB of memory. We used the 4-bit weights with the same program code, not the macOS package file. The same edit was complete after 45 minutes. We did not stop the connections on macOS.
We compared 4-bit and BF16 on 33 image edits that were not used in training. The EditScore-Qwen3-VL-8B model gave each edit a value from 0 to 10. We added the values and divided by 33. The result was 6.92 for 4-bit and 6.76 for BF16. BF16 with a different seed got 7.16. Thus, the change from 4-bit is smaller than the change from a different seed. 2-bit weights without more training got 0.26.

Credits

Open source parts

OmniGen2 by VectorSpaceLab (Apache License 2.0) is the image model and its code. The weights are OmniGen2/OmniGen2 from chibifire/omnigen2-base-df5dca8a, with the LoRA in them. They are 4-bit weights.
OmniGen2-EditScore7B-v1.1 by VectorSpaceLab (Apache License 2.0) is the LoRA. EditScore (Apache License 2.0) is the reward model of its training.
Qwen2.5-VL-3B-Instruct by the Qwen team (Apache License 2.0) is the part of OmniGen2 that reads the image and the instruction.
PyTorch, Diffusers, Transformers and Gradio operate the model and the edit page. They have open source licenses. pixi-pack (BSD 3-Clause License) makes the Python environment.