UNO - ByteDance's innovative AI image generation framework

What is UNO?

UNO is an innovative AI image generation framework launched by ByteDance, breaking through the limitations of traditional models in multi-subject generation. Through a "few-to-many" generalization approach, it can generate high-quality single-subject and multi-subject images, solving the consistency problem in multi-subject scenes. UNO generates highly consistent multi-subject data based on a diffusion transformer, employs progressive cross-modal alignment technology, and trains the model in stages to gradually improve generation results. It introduces Universal Rotation Position Embedding (UnoPE), supporting image generation at various resolutions and aspect ratios.

Main functions of UNO

• Single-subject customized generationUNO can generate images from a reference image that retain the same subject features but are in different scenes, poses, or styles.
• Multi-subject combination generationUNO can accept multiple reference images as input and generate a new image that includes all the reference subjects.
• Virtual try-on and product displayUNO supports virtual try-on functionality, allowing users to place specific products (such as clothing and accessories) on different character models to showcase their effects. Products can be placed in various scenes while maintaining their original characteristics.
• Stylization generationUNO can perform style transfer on a reference subject to generate images with different styles.
• Strong generalization abilityUNO has demonstrated strong generalization capabilities in multiple tasks, adapting to various application scenarios, such as single-subject and multi-subject driven image generation, and can generalize to scenarios such as id, tryon, and style.

UNO's technical principles

• High Consistency Data Synthesis PipelineUNO leverages the inherent context generation capabilities of diffusion transformers to generate highly consistent multi-agent paired data. It can automatically create large-scale, high-quality training data, solving the challenge of data acquisition.
• Progressive cross-modal alignmentUNO employs a progressive cross-modal alignment strategy, dividing the training process into two phases:
  • Phase 1Fine-tuning a pre-trained text-to-image (T2I) model using data generated from a single-agent context enables it to handle single-agent-driven generation tasks.
  • Phase TwoIntroducing multi-agent data for further training enhances the model's ability to handle complex scenarios. Through this progressive alignment approach, the model can better adapt to generative tasks ranging from single-agent to multi-agent models.
• Universal Rotational Position Embedding (UnoPE)UNO introduces Universal Rotational Position Embedding (UnoPE), which effectively solves the attribute confusion problem when expanding visual subject control. UnoPE modulates the interaction between multimodal tags by assigning specific position indices to text and image tags, allowing the model to focus on extracting layout information from text features, thereby improving subject similarity while maintaining good text controllability.
• Model ArchitectureUNO is based on the open-source model FLUX.1 dev, inheriting its text-to-image generation capabilities and multimodal attention mechanism, and adopts a general-purpose customized model framework. This allows the model to be iteratively trained from text to image models. Through its unique progressive cross-modal alignment and general rotation position embedding mechanisms, it achieves both high consistency and controllability in single-subject and multi-subject driven generation.
• Data Management and Model EvolutionUNO adopts a new paradigm of "model-data co-evolution," the core idea of which is to use a weaker model to generate training data and train a stronger model. This allows the model to gradually adapt to diverse scenarios during training, effectively handling complex situations that may be encountered in real-world applications.

UNO's project address

• Project official website:https://bytedance.github.io/UNO/
• Github repository:https://github.com/bytedance/UNO
• HuggingFace model:https://huggingface.co/bytedance-research/UNO
• arXiv technical paper:https://arxiv.org/pdf/2504.02160

UNO application scenarios

• Virtual try-onUNO allows users to place different clothing, accessories, and other products on virtual character models to generate try-on effects in different scenarios.
• Product DesignIn product design, UNO allows products to be placed in various backgrounds and scenarios while maintaining their original characteristics, providing designers with more flexible design ideas.
• Creative DesignUNO can accept multiple reference images as input and generate a new image containing all the reference subjects.
• Personalized content generationUNO can generate images from a reference image that retain the same subject features but are in different scenes, poses, or styles.
• Character and scene designUNO provides powerful image generation support for game development, helping developers quickly generate characters and scenes and inspire creativity.