← All model news

Vidu S2 debuts in AI-hosted livestream: assessing the real-time video model before integration

According to a September 26, 2026 report by Xinzhiyuan, Dream Dimension and Shengshu Technology had AI host "Ziying" run her first livestream at Mid-Autumn, under a rule that missing the target meant resetting memory and starting over; the report says about 20,000 people arrived within five minutes and that, after several restarts, gift targets were met. The report frames the stream as a public stress test of the real-time interaction and editing model Vidu S2, whose self-reported benchmarks and millisecond-latency claims have not been independently verified.

Published Updated

What changed

According to a September 26, 2026 report by Xinzhiyuan, during Mid-Autumn Dream Dimension and Shengshu Technology had AI host "Ziying" begin her first "endless business" livestream. The rule was blunt: if she could not earn 80,000 business points from real viewers within two hours (the report's conversion is roughly 8,000 RMB), the stream would be cut off, her memory reset, and the run restarted. The report says about 20,000 people flooded the room within five minutes; the first round missed the target, and after several restarts she earned the money herself, with the report citing over 10,000 in first-day revenue across five hours, a final business-point total of 24,184, a near-doubling in the second round, and a third round that exceeded the second round's total before it ended. The report frames the stream as a public stress test of Shengshu Technology's newly released model Vidu S2. The capabilities described in the report include: Vidu S2-Avatar supporting real-time voice interaction and complex motion control, able to combine reference images for object interaction, outfit changes and background switches; Vidu S2-Editing supporting continuous real-time editing of an input video stream based on style, clothing, subject and background reference images; and a retained offline path where one image plus one audio clip generates a digital-human video. The report says real-time frames are computed within a few hundred milliseconds, and that compared with Vidu S1, S2 can perform larger full-body motions such as turning, walking and dancing. Full-stack R&D is led by Zhang Jintao, a doctoral student of Tsinghua professor Zhu Jun and Shengshu Technology's head of streaming video generation and inference — a claim that also comes from the report, not an independent source.

What it means for developers

Separate the verifiable from the unverifiable first. The livestream did publicly happen, and the vendor did describe Vidu S2's two forms; but the stream's numbers, the millisecond latency and the reported benchmark scores all come from the organizers and the vendor. The reported StreamAV-Bench AIF audio-instruction completion of 2.985 and AVAlign of 0.353, Sparkle-Bench global instruction 4.00 and foreground instruction 3.98, and the GSB semantic-following preference rates of 71.4%–100% against three commercial systems come with no evaluation set, list of compared systems or reproduction script. GSB in particular is a preference rate, not an accuracy rate, and without knowing the compared systems and sample composition it is hard to map onto your own tasks. The following two points are conditional recommendations, not report content. First, treat the "real-time interaction path" and "offline generation" as two delivery forms to choose between, not two ways of calling the same interface. The report gives both a real-time voice interaction path and an offline path of one image plus one audio clip, and the two differ in cost structure and acceptance testing. The judgment call is whether your scenario tolerates post-editing: livestreams, real-time NPCs and video customer service involve viewers who interrupt, change their minds and throw in new demands, so only streaming works; short videos, marketing assets and pre-rendered digital-human announcements are easier offline. The premise that would change this choice is the end-to-end latency budget — real-time voice interaction must chain speech recognition, generation, speech synthesis and network round trips, while the report only says frames are computed within a few hundred milliseconds, without saying whether that is first frame, a single frame or end to end. The measurable validation step is to run a fixed script on your own network and region, recording P50 and P95 end-to-end first-response times and the success rate of resuming the previous segment after an interruption; if P95 exceeds the tolerance you set for conversational scenarios, downgrade that scenario to offline delivery first. Second, decide whether to enable dynamic references in long sessions — changing outfits or backgrounds mid-stream, or adding reference images on the fly. The report says dynamic references can add reference images at any moment while keeping a coherent before-and-after relationship, and the real risk this maps to is consistency drift in long sessions. The precondition for choosing it is that you have a character reference-image asset set under version control, because every outfit or scene change is a new reference input, and an error in the reference image shows up directly on screen. The measurable validation step is to generate multiple rounds with the same character, the same audio and the same reference images, sample frames at fixed intervals, record the deformation ratio of facial features, clothing and lighting, and the share of frames needing manual intervention, and use the share of broken frames as a launch gate rather than judging by a single polished demo. The third point is closer to risk pricing: the report describes consistency as a system property jointly constrained at five levels — training, data, architecture, inference and evaluation — and proposes SRF training technology to suppress drift in long streaming generation. If such claims hold, consistency is an engineering problem that decays with session length, not a parameter you tune once. Long continuous sessions are best tracked as a separate test line from short-clip tests.

Integration and evaluation

The first step before integration is confirming the form, not comparing prices. The report gives entry points to a demo, API platform docs and a technical report, but does not state the scope of availability, whether enterprise qualifications are required, whether private deployment is supported, concurrency and quota limits, or data-retention and regional policies. These determine whether you are calling a streaming video interface or need to stand up your own inference cluster — very different demands on the team. Second, test consistency with your own content, not official samples. Use your own character images, your own venue and styling, and run a long session covering the motions described in the report, including at least outfit changes, background switches and resuming the previous segment after viewer interruption, and write down every visible break in character. Third, calibrate the benchmark framing. Self-reported StreamAV-Bench, Sparkle-Bench and GSB scores without public evaluation sets and scripts can only be treated as the vendor's internal comparison; first confirm whether these benchmarks' languages and task forms match Chinese real-time voice and livestream scenarios before using them for selection. Fourth, keep the offline path as a fallback. The real-time path may be unavailable at peak times or during incidents, and offline generation from one image plus one audio clip keeps asset output from stalling; the cost of this backup path should also be part of the evaluation. Finally, put cost on your own usage: real-time interaction is usually billed per second or per duration, offline generation per call or per asset, and the report gives no billing method, so derive unit cost from the vendor's pricing page and your measured usage.

Limits and open questions

The report does not say whether Vidu S2's weights are open, what license applies, or the parameter count and training-data scale, so public information cannot establish whether it can be deployed locally. Open weights and an open-source license are two different things, and here not even weight availability is stated; having an API should not be read as self-hostable. SRF training technology and frame-aligned sparse attention get only one summary sentence in the report, so their engineering meaning and applicable bounds cannot be assessed. Livestream figures — 20,000 arrivals, 24,184 business points, over 10,000 RMB in first-day revenue across five hours — all come from the organizers and the report; business points are an in-platform count, the report gives only one conversion example, and none of it can be independently checked. The accompanying arXiv technical report and vendor self-reported benchmark scores have not been third-party reviewed and should not be cited as confirmed technical facts. The passages on delivery-form choice, conditions for enabling dynamic references and validation metrics are conditional analysis and recommendations, not report content, and are not judgments on the model's actual performance. Vidu S2 is not currently in KimiSeek's model catalog; the API availability mentioned in the report refers to Shengshu Technology's own Vidu platform and is unrelated to this site; the prior existence of other versions in the family does not substitute for this note. If the model is listed later, availability and pricing should follow the site's pricing page, and prices change over time and by time slot.

Three minutes to your first working call

Create an account and an API key. Calls use your account balance. 30% off your first top-up. Top up ¥500, pay ¥350, receive ¥500 in credit. Top up ¥1000, pay ¥850, receive ¥1000 in credit You can still start with a ¥10 trial top-up.

Sources and editorial process

AI主播「紫樱」首播实测:生数科技实时互动模型 Vidu S2 的公开压力测试

AI assisted the analysis and translation of the cited sources, followed by automated quality checks. Recommendations are not test results. Consult the current catalog for availability and pricing.