#8080th of 89 in Image and videoFor image, video and design, image and design

omnigen2

Multimodal model for image generation, editing and in-context composition

Stars
4.1k
License
Apache-2.0
Last commit
Mar 2026
Language
Jupyter Notebook
Stars, last 29 weeks+127
25 Apr 2026: 3,995 stars2 May 2026: 4,000 stars9 May 2026: 4,007 stars16 May 2026: 4,026 stars23 May 2026: 4,031 stars30 May 2026: 4,032 stars6 Jun 2026: 4,039 stars13 Jun 2026: 4,048 stars20 Jun 2026: 4,057 stars27 Jun 2026: 4,060 stars4 Jul 2026: 4,064 stars11 Jul 2026: 4,071 stars18 Jul 2026: 4,074 stars25 Jul 2026: 4,080 stars1 Aug 2026: 4,082 stars8 Aug 2026: 4,086 stars15 Aug 2026: 4,091 stars22 Aug 2026: 4,092 stars29 Aug 2026: 4,093 stars5 Sept 2026: 4,097 stars12 Sept 2026: 4,098 stars19 Sept 2026: 4,099 stars26 Sept 2026: 4,102 stars3 Oct 2026: 4,106 stars10 Oct 2026: 4,111 stars17 Oct 2026: 4,114 stars24 Oct 2026: 4,114 stars31 Oct 2026: 4,117 stars7 Nov 2026: 4,120 stars14 Nov 2026: 4,122 stars

3,995 on 25 Apr 20264,122 on 14 Nov 2026

From GitHub's daily star counts.

Checked on GitHub.

Overview

OmniGen2 is a multimodal generation model built on a Qwen-VL-2.5 foundation, with separate decoding pathways for text and images. The repo provides PyTorch inference scripts, Gradio demos (image-only and chat), fine-tuning code and the OmniContext benchmark. It covers visual understanding, text-to-image, instruction-guided editing and in-context generation from reference images.

Who it is for: Researchers and self-hosters doing image editing and generation on NVIDIA GPUs

Strengths

  • One model covers understanding, text-to-image, editing and in-context generation
  • Model CPU offload cuts VRAM by about half; sequential offload goes under 3GB
  • TeaCache (~30% speedup) and TaylorSeer (up to 2X) for faster inference
  • Ships fine-tuning code, X2I2 training dataset and OmniContext benchmark

Weaknesses

  • Needs about 17GB VRAM (RTX 3090 class) without offloading
  • Sometimes ignores instructions; README suggests generating several images and picking one
  • In-context generation can alter reference objects; README admits a gap versus GPT-4o
  • Output size defaults to 1024x1024; English prompts are recommended

What it needs

  • GPU required
  • Needs PyTorch 2.6.0, Python 3.11, flash-attn (optional), Gradio
  • Models: OmniGen2 (Hugging Face), OmniGen2-EditScore7B LoRA, Qwen-VL-2.5 base

Also in Image and video

See all 89
Also in Image and video
RankProjectScore
1frigateLocal NVR with realtime object detection for IP cameras37.6k starsMIT93 out of 100
2MoneyPrinterTurboGenerates short videos from a topic, with script, footage, voiceover and subtitles144.3k starsMIT88 out of 100
3toonflow-appInfinite-canvas tool for AI short drama and comic video production22.1k starsMIT88 out of 100
#4InvokeAIWeb UI for diffusion image generation with inpainting canvas and node workflows29.1k starsApache-2.087 out of 100
#5ComfyUINode-graph editor for running diffusion image, video, audio and 3D workflows147.4k starsGPL-3.085 out of 100