The research team of the Beijing Institute of Artificial Intelligence has released a new image generation model OmniGen, which breaks the single function limitation of traditional image generation tools. Different from models such as Stable Diffusion, OmniGen integrates multiple functions such as text to image generation and image editing under a unified framework, making it an "all-rounder". The editor of Downcodes will explain in detail the power of OmniGen and its application prospects.
Recently, the research team of Beijing Institute of Artificial Intelligence launched a new image generation model called OmniGen.
All-round image generation and editing player
Compared with previous image generation tools such as Stable Diffusion, the biggest highlight of OmniGen is that it no longer just focuses on a single task, it has multiple capabilities:
It can handle a variety of image generation tasks under a unified framework: from text to image generation and image editing. It can be said to be an all-rounder.
This means that users only need to provide simple prompt words to control image generation and fine editing, and no longer need to use plug-ins such as ControlNet and IP-Adapter to adjust the details of the image!
Here AIbase is based on giving a detailed effect prompt word for creative photography with an old-fashioned camera. The overall effect generated is full of details and the effect is as follows:
Across multiple tests, OmniGen performed impressively, performing on par with the most advanced models on the market for text-to-image generation. On the GenEval benchmark, OmniGen used only 0.1 billion images for training, while SD3 used over 1 billion images.
The image editing capabilities are equally excellent, with the ability to accurately control source images and editing instructions. For example, on the EMU-Edit test set, it surpasses well-known models such as InstructPix2Pix, and is even comparable to the current state-of-the-art EMU-Edit model.
In the task of subject-driven generation, OmniGen has demonstrated extraordinary personalization capabilities and is suitable for many fields such as art creation and advertising design.
Trial address: https://huggingface.co/spaces/Shitao/OmniGen
Paper: https://arxiv.org/html/2409.11340v1
OmniGen brings new breakthroughs to the field of image generation with its powerful functions and efficient performance. Its simple and easy-to-use operation method also lowers the threshold for image generation and provides more users with convenient creative tools. It is expected that OmniGen will have wider applications in the future and promote the further development of artificial intelligence image generation technology.