🚀

Image Generation - OmniGen2

🛠

Unified Text-to-Image, Image Editing & In-Context Generation

One model for three different jobs - generate images from text, edit existing images with simple instructions, or combine multiple reference images into a new scene. All driven by natural-language prompts.

What it does

OmniGen2 is a single AI model that can do three creative tasks that usually require three separate tools: turn text prompts into images, edit existing images by typing instructions ("change the sky to sunset", "remove the person on the left"), and create new images that combine elements from multiple reference photos ("put the bird from image 1 onto the desk in image 2"). It is a 7B-parameter unified multimodal model accepted at CVPR 2026, built on top of a Qwen-VL-2.5 visual-understanding backbone so it can read images and prompts together. Instead of stitching multiple specialist models into a pipeline, OmniGen2 handles all of these workflows in one place.

Problem it solves

  • One model, three workflows – Text-to-image, image editing, and multi-image composition all in a single model — no need to switch between specialized tools
  • Edit images with plain English – Describe the change you want ("make the dress red", "add a hat") instead of using masks, layers, or selection tools
  • Combine multiple references – Pull a subject from one image and a background or style from another, controlled entirely by text
  • Visual understanding built in – Because it inherits Qwen-VL-2.5's understanding ability, the model actually "reads" the input image and instruction together rather than treating the image as a passive reference
  • State-of-the-art editing among open models – Strong instruction-guided image editing performance compared to other open-source models
  • Reasonable hardware footprint – Runs on a single ~24GB consumer GPU (e.g., RTX 3090) at native settings, with CPU offload available for less-capable machines

Input/Output

  • Input:
    • Text to Image: A plain-language description of the image you want
    • Image Editing: One reference image + a text instruction describing the desired change
    • In-context / Multi-Reference Generation: One or more reference images + a text instruction explaining which elements should come from which image
      • Tip: be explicit — e.g. "Add the bird from image 1 onto the desk in image 2" works much better than "add bird to desk"
    • Recommended input image size: greater than 512 × 512 px for best quality
    • Best prompt language: English
  • Options:
    • Diffusion steps (30 ~ 50, default 30)
    • More steps = better quality, but slower generation
    • Random seed (-1 for random, or set a number to reproduce the same image)
    • Output width (px) — default 1024
    • Output height (px) — default 1024
    • Advanced Options:
      • Text CFG scale (default 5) — how strictly the output follows the text prompt
      • Image CFG scale (default 2) — how strongly the output stays faithful to the reference image
        • Higher = sticks closer to the reference image's structure and style
        • Lower (~1.5) = gives more weight to the text prompt
        • Recommended: 1.2 – 2.0 for image editing, 2.5 – 3.0 for in-context generation
      • Negative prompt — what you don't want in the image
        • Default: (((deformed))), blurry, over saturation, bad anatomy, disfigured, poorly drawn face, mutation, mutated, (extra_limb), (ugly), (poorly drawn hands), fused fingers, messy drawing, broken legs censor, censored, censor_bar
  • Output: Generated image
    • High-fidelity, aesthetically pleasing output
    • Same model used whether the task is generation, editing, or multi-reference composition

Accuracy & Speed

Generation time
~15 – 30 seconds per image at default settings
Hardware requirement
~17 GB VRAM (e.g., NVIDIA RTX 3090)
Lower-VRAM options
CPU offload reduces VRAM by ~50% with negligible speed impact; sequential offload allows running in <3 GB VRAM at much slower speed
Text-to-image quality
Competitive with leading open-source unified models
Image editing
State-of-the-art among open-source models on instruction-guided editing benchmarks
In-context generation
Strong performance on the OmniContext benc

Technical Details

Architecture
7B-parameter unified multimodal diffusion model
Model Design
Two decoding pathways (text + image) with unshared parameters, plus a decoupled image tokenizer
Visual Backbone
Qwen-VL-2.5 (inherited visual understanding)
Capabilities
Text-to-image generation, instruction-guided image editing, in-context (multi-reference) generation, visual understanding
Key Hyperparameters
text_guidance_scale, image_guidance_scale, max_pixels, cfg_range_start / cfg_range_end
Inference Optimizations
TeaCache, TaylorSeer, model CPU offload (~50% VRAM saving), sequential CPU offload (<3 GB VRAM), optional flash-attn
Supported Frameworks
Hugging Face Diffusers (OmniGen2Pipeline), Safetensors, ComfyUI (official integration), Gradio demo
Training & Benchmarks
Training dataset: X2I2 — In-context benchmark: OmniContext — Training & fine-tuning code public
Paper
"OmniGen2: Exploration to Advanced Multimodal Generation" — arXiv 2506.18871 (CVPR 2026)

Compliance & Provenance

Provider
Open-source (OmniGen team)
Provider type
Specialized
License
Apache 2.0
EU AI Act risk class
Limited Risk
Art. 50 transparency
Required — outputs are marked. See AI Policy §2.
Region availability
Available globally
Training data summary
Provider model card / training data

For more on how we classify models and mark outputs, see our AI Policy.

Model Source