Wan Streamer v0.2: Higher Resolution, Same Latency

Overview概览

Wan Streamer v0.2 keeps the v0.1 native-streaming formulation: user and agent text, audio, and video remain on one causal timeline and are modeled by a single Transformer. Around that formulation, v0.2 raises the interactive output stream from 192×336 to 640×368 while preserving its low-latency interaction budget.Wan Streamer v0.2 延续 v0.1 的原生流式建模：用户与智能体的文本、音频、视频仍位于同一条因果时间线上，并由单个 Transformer 统一建模。在这个建模框架不变的前提下，v0.2 将交互输出从 192×336 提升到 640×368，同时维持低延迟预算。

Wan Streamer framework — **Overview of Wan Streamer.** Language, audio, and video inputs and outputs are represented on a shared causal timeline and coordinated by block-causal attention. v0.2 keeps this formulation while increasing output resolution and changing the serving topology. **Wan Streamer 总体框架。**语言、音频和视频的输入输出被表示在共享因果时间线上，并通过 block-causal attention 协同建模。v0.2 保持这一建模方式，同时提升输出分辨率并调整推理拓扑。

Against the v0.1 baseline, v0.2 changes three axes: the output stream increases to 640×368, the high-cost latent generation path moves into a Ulysses-style context-parallel performer, and the supported visual composition expands from close-up calls to scene-grounded mid-shot agents.相对于 v0.1 基线，v0.2 主要改变三点：输出提升到 640×368，高成本 latent 生成路径迁移到 Ulysses 式上下文并行 performer，并且可支持的视觉构图从近景通话扩展到具有场景感的中景智能体。

v0.1 vs v0.2v0.1 / v0.2 对比

The end-to-end streaming formulation and latency budget stay fixed, while output resolution, visual format, and serving topology change.端到端流式建模和延迟预算保持不变，变化集中在输出分辨率、视觉形式和推理拓扑。

Aspect项目	v0.1	v0.2
Output resolution输出分辨率	192×336	640×368
Frame rate帧率	25 fps	25 fps
Model-side latency模型侧延迟	~200 ms	~200 ms unchanged不变
Total latency总交互延迟	~550 ms with 350 ms network网络往返 350 ms 时约 550 ms	~550 ms with 350 ms network网络往返 350 ms 时约 550 ms
ThinkerThinker	Streaming perception, state update, KV construction, and decoding流式感知、状态更新、KV 构建与解码	Same role, kept on one GPU职责不变，仍为单 GPU
PerformerPerformer	Single-GPU performer latent generation单 GPU 生成 latent	Multi-GPU Ulysses-style context-parallel latent generation多 GPU Ulysses 式上下文并行 latent 生成
Communication通信	Thinker ↔ performer K/V and latent exchangethinker ↔ performer 交换 K/V 和 latent	Thinker broadcasts performer-compatible K/V slices; Ulysses all-to-all/gather for the latent video sequence stays inside the performer groupthinker 广播 performer 可用的 K/V 切片；视频 latent 序列的 Ulysses all-to-all/gather 保留在 performer 组内
Visual presence视觉呈现	Close-up video-call framing近景视频通话	Higher-fidelity close-up interactions + scene-grounded mid-shot agents with readable body and scene context更高保真的近景交互 + 身体与场景信息清晰可读的场景化中景智能体

Demos演示

Real-time deployment recordings. Use the arrows or horizontal scroll to switch clips.实时部署录屏。可用箭头或横向滚动切换片段。

LIVE CAPTURE WITH REAL END-TO-END LATENCY. Full deployed sessions recorded as they ran, including network time. No retiming, compositing, synthetic UI, or post-processed latency. 真实端到端延迟的现场录屏。 所有片段均为实际部署会话的原始录制，包含网络耗时；没有重定时、拼接、合成界面或后处理延迟。

1 / 4

A landscape call with clear facial detail, expression, and lip motion.横屏通话，面部细节、表情和口型更清晰。

A portrait scene with body motion, handheld objects, and stable agent presence.竖屏桌面场景，包含身体动作、手持物体和稳定的智能体呈现。

A wider scene with hands, objects, and agent feedback in the same frame.更宽的生活场景中，手部、物体和智能体反馈处在同一画面。

A street-side vertical interaction with natural turn-taking and mobile context.街边竖屏互动，呈现自然轮次切换和移动场景。

Deployment: thinker–performer部署：thinker–performer 架构

The serving challenge in v0.2 is to allocate the additional 640×368 generation cost without slowing down the interactive loop. Wan Streamer is trained as a single end-to-end model; for real-time deployment, that same model is split into a thinker–performer pipeline with a single-GPU thinker and a Ulysses-style context-parallel multi-GPU performer.v0.2 的推理挑战，是在不拖慢交互循环的前提下承接 640×368 带来的额外生成成本。Wan Streamer 训练时是单个端到端模型；实时部署时，同一个模型拆分为 thinker–performer 流水线，其中 thinker 为单 GPU，performer 为多 GPU Ulysses 式上下文并行组。

Thinker-performer overlap / thinker-performer 流水线 — **v0.2 latency-preserving serving.** The thinker stays on one GPU and owns perception, language/state update, K/V construction, and final decoding. The performer uses Ulysses-style context parallelism for the 640×368 latent generation path: K/V slices are written into pre-sharded performer caches, the high-resolution latent video sequence is split across ranks for denoising, and Ulysses all-to-all/gather communication stays inside the performer group. Audio latents are short and are not split this way. **v0.2 延迟保持型推理。**thinker 保持在单 GPU 上，负责感知、语言/状态更新、K/V 构建和最终解码。performer 通过多 GPU Ulysses 式上下文并行承接 640×368 latent 生成路径：K/V 切片写入预分片的 performer cache，高分辨率视频 latent 序列按 rank 切分并去噪，Ulysses all-to-all/gather 通信保留在 performer 组内。音频 latent 很短，不做序列分片。

The thinker hosts the causal audio/video encoders, the short token-causal Transformer pass that updates language and state, K/V construction, and the causal decoders that render the previous unit's latents into audio and video for immediate emission. The performer holds only the latent-generation path. Performer ranks receive the current K/V slice, update local cache shards, and run Ulysses context-parallel denoising for the next unit. The high-resolution latent video is the long sequence that gets split across ranks. Audio latents in a 160 ms unit contain few tokens, so they are generated without sequence sharding. The language/state part does not become a separate performer sequence; it is already represented in the K/V cache produced by the thinker.thinker 承载因果音视频编码器、用于语言与状态更新的短 token-causal Transformer 计算、K/V 构建，以及将上一单元 latent 渲染为音视频并立即输出的因果解码器。performer 只保留 latent 生成路径：各 performer rank 接收当前 K/V 切片，更新本地 cache shard，并为下一单元运行 Ulysses 上下文并行去噪。高分辨率视频 latent 是真正长的序列，因此按 rank 切分；160 ms 单元里的音频 latent token 很少，所以不做序列分片。语言/状态部分也不会作为另一条序列进入 performer，它已经体现在 thinker 产生的 K/V cache 中。

This schedule pipelines current-frame perception, previous-frame decoding, thinker-performer K/V and latent transfer, and next-frame latent video denoising across adjacent units. Real-time throughput holds as long as performer-group time plus communication fits inside one 160 ms streaming unit. Separately, the signal-to-signal path — encode → state update → latent generation → decode — remains the ~200 ms model-side latency target.这一调度将当前帧感知、上一帧解码、thinker–performer K/V 与 latent 传输，以及下一帧视频 latent 的并行去噪流水线化，分摊到相邻流式单元中。只要 performer 组计算加通信能控制在一个 160 ms 流式单元内，就可以维持实时吞吐；同时，编码 → 状态更新 → latent 生成 → 解码这条端到端信号路径仍保持 约 200 ms 的模型侧延迟目标。

Cite this work引用本文

@misc{huang2026wanstreamerv02higherresolutionrealtimeinteraction,
  title         = {Wan-Streamer v0.2: Higher Resolution, Same Latency},
  author        = {Lianghua Huang and Zhi-Fan Wu and Yupeng Shi and Wei Wang and
                   Mengyang Feng and Junjie He and Chen-Wei Xie and Yu Liu and
                   Jingren Zhou and Ang Wang and Bang Zhang and Baole Ai and
                   Chen Liang and Cheng Yu and Chongyang Zhong and Jinwei Qi and
                   Kai Zhu and Pandeng Li and Peng Zhang and Wenyuan Zhang and
                   Xinhua Cheng and Yitong Huang and Yun Zheng and Yuxiang Bao and Yuzheng Wang and Zoubin Bi},
  year          = {2026},
  month         = jun,
  note          = {Technical report}
}