UniSpark
蚂蚁技术研究院 Ant Research

UniSpark

Exploring Hybrid Attention for
Multimodal Understanding and Generation

Understand, create, and edit within one shared framework.

Technical reportExplore the gallery
Portrait of a woman in a black turtleneck
Flower-lined canal with covered market walkways
Red fox surrounded by frost
Textured embroidered map of a river

Selected generations from the UniSpark paper.

View the full gallery

Hybrid attention.
Shared visual context.

Can hybrid attention support both understanding and visual generation? UniSpark brings full attention and recurrent linear attention into a shared Mixture-of-Transformer-Experts backbone.

Full attention Explicit token interactions
+
GatedDeltaNet Recurrent context states
UniSpark-9BShared multimodal contextApproximately 9B parameters per branch
UnderstandVisual question answering, perception, and reasoningCreate & editText-to-image generation and instruction-guided editingContinue & adaptInterleaved sequences and video generation and editing

Autoregressive text prediction and visual flow matching connect language and images. Private noisy targets remain separate from reusable clean content, so completed outputs can provide context for later predictions.

Training progresses from image resolution scaling up to 2K to unified supervised refinement, followed by video adaptation with mixed image and video tasks. The understanding branch stays frozen during video adaptation.

Explore the model architecture
UniSpark architecture: understanding and generation branches with full attention and GatedDeltaNet
Shared context connects text, semantic visual features, and visual latents.

Small instructions.
Meaningful changes.

Reference-conditioned editing changes selected attributes while retaining the scene. Choose an example and drag the divider to compare.

Image editing

Descriptions summarize the requested edits.

Edited image: emerald green coatSource image: red coatSourceEdited

Drag to compare · Use arrow keys when the slider is focused

Beyond a single image

Interleaved prediction carries visual context across successive outputs. These four-image sequences illustrate scene continuity as actions unfold.

UniSpark in action

Generation, editing, and visual continuity in one showcase.

72-second showcase

Understanding, generation, and efficiency

Explore benchmark breakdowns, evaluation settings, and the trade-offs behind each result. Expand a table to see all model entries reported in the paper.

View the benchmark overview
Selected understanding, image generation, and image editing benchmark results
Selected results from Figure 1. ImgEdit scores use different judges; see the editing tab for details.

Source: Technical report, Experiments. Values follow the manuscript tables; benchmark papers are linked in each evaluation-settings panel.

Reference

Technical report citation
@misc{pan2026unisparkexploringhybridattention,
      title={UniSpark: Exploring Hybrid Attention for Multimodal Understanding and Generation},
      author={Jiancheng Pan and Yuzheng Liu and Hengyuan Cao and Qiyuan Zhang and Yangyan Li and Biao Gong and Zheng Zhang and Ka Long Cheng and Xinyang Wang and Yujun Shen},
      year={2026},
      eprint={2611.16517},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2611.16517},
}

Authors