OmniGen2: composing one picture from several reference photos, and where to run it
Apache 2.0, weights free, nothing to accept. Self-hosting needs about 17GB of video memory, which its authors pair with an RTX 3090 or an equivalent card, plus the Hugging Face diffusers library, with CPU offload available for smaller GPUs. In your browser it runs on cnaps studio, included in the basic plan.
One model covers three jobs: generating from text, editing from an instruction, and composing a new picture from several reference images. The third is the one to come for. Write put the bird from image 1 onto the desk in image 2 and it takes the subject from one photo and the setting from another.
Two costs. It publishes no benchmark score, and at a default of 30 steps it takes many more passes than the distilled generators here. Its authors also say it works best with English prompts. Seven image generation models sit in the cnaps studio catalog, and this is the one built around combining references.
What it actually is
A 7-billion-parameter unified model built on a vision-language backbone, so it reads the reference images and the instruction together. Text and image are decoded along two separate pathways with unshared parameters, and images pass through a decoupled image tokenizer.
The settings follow the three jobs. Diffusion steps run 30 to 50, defaulting to 30, with a seed and width and height defaulting to 1024. Two guidance scales do the steering: text guidance, default 5, sets how closely the prompt is followed, and image guidance, default 2, sets how closely the output keeps to the references. The authors recommend 1.2 to 2.0 for editing and 2.5 to 3.0 for composing from several images.
Input images work best above 512 by 512 pixels, and a default negative prompt already lists the usual failures, from bad anatomy to fused fingers. Instructions work better when they say which element comes from which image.
Try OmniGen2 in cnaps studio →
The published numbers
None on the model card. The authors built their own benchmark for multi-reference work, OmniContext, and name it on the card without listing a score, and their efficiency figures appear only as a chart image.
The paper was accepted at CVPR 2026, so the method has passed peer review, though acceptance is not a measurement of output quality. Adoption is modest at 3,545 downloads in the last month.
With no number to lean on, the useful evidence is what it does with your own references, and the image guidance setting makes that easy to probe: the same composition at two values shows how far it will drift from the source photos.
How it compares to the other image generators here
No two of these columns share a test, so the table maps what each model accepts. The four overlap heavily: all of them generate, edit and take references.
Reach for OmniGen2 when the job is combining elements from several photos and you want a dial for how faithfully each reference is kept. Reach for FLUX.2 KLEIN 4B when speed and a small card matter, DeepGen-1.0 when the instruction needs reasoning, and GLM-Image when the result must carry legible text. For a text-only prompt, Z-Image-Turbo is the more widely used choice.
What to chain it with
Images and text in, one picture out, so it sits in the middle of a chain more often than at either end.
Two chains suit it. BiRefNet (Subject) cuts a product or person out of its photo cleanly, and that cutout makes a better reference than a cluttered original. Afterwards, QWEN-Inpaint fixes one region without regenerating the composition, and PiSA-SR or SMFANet+ enlarges the 1024-pixel result for print.
Open OmniGen2 in cnaps studio, run one two-image composition at image guidance 2.5 and again at 3.0, and see which keeps your subject closer to the original.
FAQ
What does OmniGen2 need to run locally?
About 17GB of video memory, which its authors pair with an RTX 3090 or an equivalent card. CPU offload brings that down for smaller GPUs, and a sequential offload mode goes further at a much slower speed. In cnaps studio it runs in the browser and is included in the basic plan.
How do I combine elements from two photos?
Upload both images and name each element's source in the instruction, for example: add the bird from image 1 onto the desk in image 2. A vague instruction such as add a bird to the desk works less well. Keep image guidance between 2.5 and 3.0, the authors' recommendation for this mode.
Does OmniGen2 work with prompts in other languages?
Its authors say it performs best with English prompts and name no other language, so English is the safest choice for instructions. For pictures that need legible English or Chinese text inside them, GLM-Image is the stronger option in the cnaps studio catalog.
