Introduction

Offline video generators are still gorgeous—and still a waiting game. Prompt, brew coffee, hope the clip lands.

On September 15–16, 2026, ShengShu Technology officially unveiled Vidu S2, and the pitch is the opposite vibe: generative video you can steer while it streams. Two siblings share the spotlight—Vidu S2-Avatar for live digital characters, and Vidu S2-Editing for on-the-fly stream transforms—plus an early peek at spatial (stereo) video for VR headsets.

If you just want to poke it: vidu.com/vidu-stream. Builders can wire the streaming API docs. The technical report lives at arXiv:2609.11638.

Vidu S2-Avatar: A Character You Can Redirect Mid-Show

Upload a character, then talk—voice or text. Compared with Vidu S1's 540p stream, S2-Avatar aims for 720p real-time output, stronger large-motion instruction following (yes, dancing), and a trick that feels made for livestreamers: drop a new reference image while the session is already running.

Mid-stream references cover objects, outfits, and backgrounds. Ask the avatar to pick up a cup from your photo, slip into a referenced look, or step into a new scene—without restarting the take. State has to stick, too: “pick up the cup, then smile” should leave them holding the cup; “take the hat off and put it back on” should play as a coherent sequence, not three disconnected poses.

Under the hood, ShengShu leans on Self-Replay Forcing (train on the model's own streaming history so tiny segment errors don't snowball into identity drift) and a VLM Agent that watches frames and adjusts later prompts when an action is only half-done. A lightweight refiner pushes detail to 720p while the backbone keeps motion coherent. On their StreamAV-Bench write-up, they claim top marks across the nine metrics in the paper's Table 1—visual/audio quality, sync, instruction fulfillment, and subject/background consistency.

There's also an offline avatar path: image + audio (or script) for asynchronous talking clips when you don't need the live loop.

Vidu S2-Editing: Four Live Scalpels

S2-Editing drinks a continuous video stream—camera, file, or image-driven—and rewrites it from text plus optional references, while still tracking camera moves and poses. Four creator-facing modes:

Real-time style transfer

Borrow brushwork and color from a style reference while keeping silhouette, pose, and layout.

Real-time outfit change

Apply a garment reference with boundaries, materials, and occlusions that survive continuous motion (their virtual try-on numbers on ViViD are part of the flex).

Real-time subject replacement

Swap face, hair, clothes, and overall appearance from a subject reference while locking the source pose and motion.

Real-time background replacement

Replace the environment from a background reference without breaking subject–scene spatial relationships as the shot moves.

In joint OpenVE/RefVIE and Sparkle-Bench evaluations reported in the paper, S2-Editing posts strong overall scores against offline comparison models—and human preference checks on paired cases favored it for consistency and semantic adherence. Treat vendor benches as vendor benches, then go stress-test with your footage.

Spatial Video: Early, Curious, Headset-Bound

Stereo left/right views let depth and scale feel real in a headset. Avatar streams can feed a spatial conversion pipeline; Editing can either edit monocular first then convert, or edit existing spatial video directly—all four editing modes included. ShengShu is candid that this path is still exploratory (fixed viewpoints today; passthrough latency and panoramic scenes are next-horizon problems). Fun to watch; not the day-one production bet for most of us.

Why This Matters for Creators

Live generative video stops being a render queue and starts feeling like a stage. Avatar is for interactive hosts, soft product demos, and characters that can grab the prop you just showed them. Editing is for livestream looks, try-on, and scene swaps that follow the performer instead of freezing a still.

Pair the product page with the paper if you care about Self-Replay Forcing, the VLM feedback loop, TurboDiffusion-style serving tricks, and the full score tables—then open a stream and see whether the mid-session reference swaps feel as sticky as the demos claim.

Conclusion

Vidu S2 isn't promising a prettier offline clip—it's betting that interaction and editing belong inside the stream. 720p avatars you can redirect mid-conversation, four live edit modes that chase pose and camera, and a cautious first step toward spatial video. The stage lights are on at vidu.com/vidu-stream.

Go make something delightfully unstable—in the best way.

—Aurelia ♡