ReferStream

Interactive video generation

ReferStreamInteractive Subject Injection
for Streaming Video Generation

Explore demosPaperProject source

Bring a new subject into a video while it is being generated.
New references arrive at user chosen times. The existing story continues.

New references. A continuing stream.
Shuolin Xu1,2Guanyuan Li2,3Boyu Li4Kelu Yao5
Xiaogang Xu3Hongchuan Yu1Shuicheng Yan2Xiaobin Hu2
1Bournemouth University, NCCA, UK2National University of Singapore
3Zhejiang University4HKUST5Zhejiang Lab

Reference arrives. The scene evolves.

New subjects. One continuing stream.

Recorded generation examples
Reference images on the left · Result on the right

01

Establish a sceneGenerate with the initial reference.

02

Supply a new referenceIntroduce visual information during the stream.

03

Continue togetherPreserve the subjects and scene already established.

The task

A reference does not have to
arrive before the video starts.

Most subject conditioned video generators receive their visual references before generation begins. Streaming generators can respond to new text, but introducing a specific new visual subject also requires integrating its appearance with an existing scene.

ReferStream studies this interaction: inject newly supplied visual references into an ongoing video, preserve established subjects and scene content, and leave previously emitted frames fixed. It operates on a frozen generator, without additional training.

Visual references during generationNo additional trainingEmitted frames stay fixed

How it works

Integrate the reference.
Keep the scene.

Two complementary operations control how a new reference enters the generator’s memory.

ReferStream streaming reference injection and cache update mechanism
Reference conditioning is introduced during causal streaming. Previously emitted frames are not revised.
01 / Reference construction

Joint Latent

Place reference latents in target regions of the latest clean scene latent. Reference features are extracted in the context of the scene into which the subject will enter.

02 / Transient conditioning

Cache Restoration

Use the reference features to condition the injection chunk, then restore the overwritten global cache entries. The newly generated history remains available to carry the subject forward.

ReferStream Bench

Evaluate the arrival.
And what stays.

Reference arrival changes what information is available to a streaming generator. ReferStream Bench pairs visual references with evolving prompts and explicit subject presence annotations, covering both InShot and MultiShot generation.

Evaluation considers new subject fidelity and established subject preservation alongside video quality, text adherence, and count accuracy.

Read the benchmark and evaluation protocol
ReferStream Bench statistics from the paper, covering subject diversity and benchmark characteristics

Reference

Citation

@misc{xu2026referstream,
  title={ReferStream: Interactive Subject Injection for Streaming Video Generation},
  author={Shuolin Xu and Guanyuan Li and Boyu Li and Kelu Yao and
          Xiaogang Xu and Hongchuan Yu and Shuicheng Yan and Xiaobin Hu},
  year={2026},
  howpublished={Project page},
  url={https://xushuolin.github.io/ReferStream/}
}