Vidu S2: Real-Time Interactive Avatars and Video Editing

i10X Quick Take

Quick Take

The finding: Vidu S2-Avatar supports real-time 720p generation with dynamic references and stronger instruction following; S2-Editing edits a live stream (style, try-on, character/background replacement).
Preference wins: In GSB tests, S2-Avatar is preferred for overall quality in 85.7% of comparisons vs Runway Character GWM-1 and 100% vs PixVerse Image Avatar and HeyGen.
Editing too: S2-Editing wins majority preferences on video quality, consistency and semantic adherence against several commercial editors.

Study profile

Preprint
No code released
Not independently replicated

Paper: Vidu S2: Real-Time Interactive, Editable, and Spatial Video Generation
Authors: Jintao Zhang, Kai Jiang, Jintao Chen, Xu Wang, Deyuan Liu, Jungang Li, Dechuang Chen, Ming Lin, Jingjiang Zhou, Haopeng Jin, Qi Jia, Xiaohang Wang et al. (35 authors)
Institutions: Tsinghua University; Shengshu Technology
Published: Sep 10, 2026
Status: Preprint (not peer-reviewed) · arXiv (cs.CV), version 1
Models tested: Vidu S2-Avatar and Vidu S2-Editing vs Runway, PixVerse, HeyGen and editing baselines
Data / sample: GSB human preference studies for streaming character and editing
Read the paper: arXiv:2609.11638 · PDF
Code & data: Not released
Funding / conflicts: Shengshu / Vidu commercial affiliation.
Paper license: CC BY 4.0
Figures: Chart drawn by i10X from Figure 6 GSB results (CC BY 4.0)

Interactive digital characters and live video edits need real-time models, not offline renders. Tsinghua University and Shengshu Technology present Vidu S2, spanning Avatar and Editing, and explore real-time spatial video for both.

i10X chart based on data from Zhang et al. (2026), arXiv:2609.11638, Figure 6 GSB results (CC BY 4.0).

What the study shows

Clip filtering keeps false rejection under 2% while avoiding cuts. Human preference studies cover motion, expression, semantic adherence and A/V sync.

Motion and expression: Unanimous preference over Runway and HeyGen on both dimensions in reported cells; 100% expression preference over PixVerse.
Demo available: Playable demo at vidu.com/vidu-stream.

Why it matters: interaction needs milliseconds

Streaming avatars and try-on only work if generation keeps up with user input. S2 targets that product surface.

Method in brief

GSB human preference comparisons; streaming architectures for Avatar and Editing.

What remains open

Vendor study: Authors affiliated with Shengshu / Vidu product.
Preference N: Some cells use small comparison sets.

Context

What it means for interactive video products

Prioritize dynamic references: Users change outfits mid-stream.
Bundle editing with generation: Same real-time stack.
Validate with GSB: Not only automated scores.

About the analysis

Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated illustration. Read our AI Use Policy. Spotted an error? Request a correction.

This article is based on the full text of the preprint “Vidu S2: Real-Time Interactive, Editable, and Spatial Video Generation” (arXiv:2609.11638v1, submitted September 10, 2026). All numbers are taken from the paper’s text, tables and figures; i10X did not rerun the experiments. The chart was redrawn by i10X from Figure 6 GSB results of the paper, which is published under a CC BY 4.0 license.

i10x Perspective

Vidu S2’s story is product timing: real-time 720p with live edits is where digital characters leave the demo reel. The sweep of preference wins suggests the latency constraint did not force a quality cliff.