Image
Bagel
ByteDance's open-source unified multimodal model for understanding, generation and image editing.
★ 6.2k starsApache-2.0
View guide →OmniGen: one unified model for text-to-image, editing, subject-driven and identity-preserving generation
OmniGen is an open-source image project by VectorSpaceLab (VectorSpaceLab/OmniGen on GitHub). It has 4,343 GitHub stars and 364 forks, is released under the MIT license, and was last updated on GitHub on 4 December 2025.
Create the documented Conda environment (Python 3.10.x) — OmniGen inference expects a CUDA GPU.
Clone the repo and install it in editable mode after installing the pinned PyTorch build.
Load Shitao/OmniGen-v1 weights from Hugging Face through the OmniGen pipeline (or Diffusers).
Write a prompt, pass input images with guidance scales, and save your first edited or subject-driven image.
Text-to-image, editing, subject-driven, identity-preserving, and image-conditioned generation in one model.
Tasks that usually need ControlNet or IP-Adapter are handled natively by the framework.
Reference images inline with <|image_1|>-style placeholders and combine multiple inputs.
Separate guidance_scale and img_guidance_scale parameters balance text and image influence.
Use the model through Hugging Face Diffusers or run the included Gradio demo (app.py).
LoRA and full-parameter training scripts (train.py with accelerate) adapt the model to your data.
Clone the project and enter its directory.
git clone https://github.com/VectorSpaceLab/OmniGen.git
cd OmniGenCreate the Conda env, install the pinned CUDA PyTorch build from the README, then install the package.
conda create -n omnigen python=3.10.13
conda activate omnigen
pip install torch==2.3.1+cu118 torchvision --extra-index-url https://download.pytorch.org/whl/cu118
pip install -e .Load the pretrained model and generate with an image reference, following the README's examples.
from OmniGen import OmniGenPipeline
pipe = OmniGenPipeline.from_pretrained("Shitao/OmniGen-v1")
images = pipe(prompt="Edit <|image_1|>", input_images=["input.png"], guidance_scale=2.5, img_guidance_scale=1.6, seed=0)VectorSpaceLab, the research team behind the OmniGen paper (arXiv 2409.11340).
The code is MIT-licensed and the weights are downloadable from Hugging Face; check the model card for the weights' own terms.
No — it runs locally on your GPU with downloaded weights.
OmniGen folds editing, subject-driven, and conditioned generation into one model with in-prompt image references, instead of a base model plus separate plugin networks.
Create the Python 3.10 Conda env, install the pinned PyTorch build, `pip install -e .`, then run the OmniGenPipeline example from the README.
ByteDance's open-source unified multimodal model for understanding, generation and image editing.
Infinite text-free AI image generator with a searchable multilingual vault
MMagic — OpenMMLab's generative-AI toolbox: diffusion, text-to-image, restoration and a model zoo.
Tencent's HunyuanImage-3.0 — a powerful native multimodal model for image generation.
How Python does AI. Agents, realtime voice, image generation, embeddings. Every model, every interface, typed end to end.
Run LLMs plus image, video and music generation locally on your phone
The most-starred other Image projects in the LibreAI catalogue — the realistic shortlist if OmniGen is not the right fit.
Stable Diffusion web UI
The node-graph powerhouse for precise, professional AI image and video generation.
High performance self-hosted photo and video management solution
Focus on prompting and generating
🆙 Upscayl - #1 Free and Open Source AI Image Upscaler for Linux, MacOS and Windows
Free and open — grab the source, star the repo, and start creating.