On July 30, Kunlun Tech officially launched and open-sourced Skywork UniPic, a multi-modal unified pre-training model that follows the autoregressive approach. The model integrates three core capabilities - image understanding, text-to-image generation, and image editing - within a single architecture. It demonstrates excellent versatility and portability through end-to-end pre-training on large-scale high-quality data.
The model uses MAR encoder and SigLIP2 backbone network, overcoming the limitations of traditional multi-modal models that rely on VQ or VAE encoders, which often result in insufficient semantic information retention. This allows for deep collaboration across tasks. Users can perform complex operations such as image understanding, generation, and style transfer by simply inputting prompts, for example, generating images of specific scenes or performing stylistic editing on images.
In terms of performance, Skywork UniPic achieves near-large unified model efficiency with a parameter scale of 1.5B. In benchmark tests such as instruction following, complex instruction image generation, and image editing, the model reaches industry-leading levels, especially running smoothly on consumer-grade graphics cards, significantly reducing the technical application threshold.
The success of the model is attributed to a refined data construction system, dedicated Reward Model optimization, and a progressive multi-task training strategy. Through an efficient corpus and hierarchical resolution training mechanism, Skywork UniPic balances model performance and training efficiency, avoiding the technical bottlenecks of capability trade-offs in traditional methods.
This open-source initiative is an important step by Kunlun Tech to promote AI technology accessibility. Previously, the company has open-sourced multiple large models in various fields, and the addition of Skywork UniPic will further help AI become an accessible creative tool. Users can access core resources such as model weights, technical reports, and code repositories through the specified links.
Model Weights:
https://huggingface.co/Skywork/Skywork-UniPic-1.5B
Technical Report:
https://github.com/SkyworkAI/UniPic/blob/main/UNIPIC.pdf
Code Repository:
https://github.com/SkyworkAI/UniPic