What Is OmniGen2?
OmniGen2 is VectorSpaceLab’s open unified generator. It can create from text, follow an edit instruction or combine people, objects and scenes from as many as three ordered images through one exact endpoint.
What Is OmniGen2 Best For?
- In-context generation
- Combining people and objects
- Instruction-guided editing
- Open-model workflows
How to Use OmniGen2
- Upload up to three references.
- Number every image in the prompt.
- Balance text and reference guidance.
- Generate one output before increasing the batch.
OmniGen2 Settings Explained
Reference Guidance
Use 1.3–2.0 for editing and 2.0–3.0 for stronger in-context reference following.
Text Guidance
Higher values prioritise the written instruction over variation.
The panel above shows only the parameters OmniGen2 actually supports. Switch versions and unsupported controls disappear rather than sitting there doing nothing.
OmniGen2 Prompt Examples
Multi-reference
Put the bird from image 1 on the desk in image 2, use the room lighting from image 3, preserve the bird markings
Edit
Keep the person and pose unchanged; make the dress cobalt blue with a matte silk texture
OmniGen2 Test Results
Testing in progress
Every model runs the same eight-category prompt set — portrait, product, poster text, anime character, interior, multi-reference, image editing and character consistency — recording generation time, success rate, text accuracy and identity retention. That run has not finished for OmniGen2, so this page shows no scores. We do not publish numbers we have not measured.
OmniGen2 Limitations
- 1K output
- Slower 50-step default
- Reference-heavy prompts must identify every image by number
OmniGen2 Pricing and Credits
Credits scale with resolution and image count. The Generate button always shows what this run costs.
1K
7 credits / image
Failed generations cost 0 credits.
Commercial Use and Image Rights
Commercial use is subject to the OmniGen2 model licence and the selected service terms.
- Model used: OmniGen2
- Model developer: VectorSpaceLab
- The inference route is managed internally and may change without changing the selected model
- Uploads are sent only to the configured inference service needed to complete the generation
- You can delete any generation from your history at any time
Frequently Asked Questions
Does it need a separate edit endpoint?+
No. The same OmniGen2 operation accepts zero to three input images.
Why does it cost more than draft models?+
The verified route bills per output megapixel at a materially higher rate.
Model Updates and Changelog
- 2026-08-25 —— Added exact text, edit and three-reference request mappings.
Related Models and Tools
AI Image Generator
One prompt box, every major model. Describe it, pick a model, generate.
AI Image Editor
Upload an image, describe the change. Only the part you named moves.
Image to Image
Start from an image. You decide how much of it survives.
HiDream O1 Image
Text To Image · Image To Image · Image Editing
HiDream I1 Full
Text To Image
BAGEL Image
Text To Image
Nano Banana 2 Lite
Text To Image · Image To Image · Image Editing