Shengshu Technology Releases Vidu Q4 Preview: Starting at ¥0.09/sec, Supports 4K Output

Shengshu Technology launches Vidu Q4 Preview, the first preview version of Vidu Q4 aimed at creators, focusing on three capabilities: more delicate character performance, more tension-filled cinematic language, and more impactful complex visual effects. The report compares it to a 'DeepSeek moment' for video models, on the grounds that performance, efficiency, and usability thresholds are simultaneously being pushed to the extreme.

This analogy itself is a media judgment, not a product fact. Whether it holds depends on whether three things can be verified at the same time: the model's capability ceiling, system efficiency, and creative threshold. The judgment proposed by the report is that only when all three hold together can video generation become a stable, scalable content production tool.

The boundary must be drawn first: Vidu Q4 Preview is currently just an early appearance before the official version launches, and the official version's performance in stability, controllability, and creative ceiling has not yet been announced. Therefore, this article discusses 'what the preview version demonstrates,' not 'what Vidu Q4 can ultimately do.'

The most eye-catching number is the price. Vidu Q4 Preview's debut price is 0.09 yuan/second and up, and the report says the same budget can generate up to 5 times more content. The accompanying input and output specifications are: up to 15 reference images, reference voice input, and 2K, 4K quality output.

On efficiency, the report says its image-to-video speed is about 2 times that of similar models, with a 5-second video completed in as fast as about 110 seconds. These numbers all come from Shengshu Technology or Machine Heart, lacking independent third-party verification; the specific comparison models, hardware, and test conditions corresponding to the about 2x speed and 110 seconds are also not disclosed.

So a more prudent reading is: this is a set of official statements with a clear direction, not a reproducible benchmark conclusion. Whether the price will be maintained long-term and what the official version pricing will be are also not explained in the report.

The report proposes a noteworthy argument: if a model continuously generates content with identity drift, broken motion, or product deformation, no matter how low the per-generation price is, creators still have to pay a lot in screening, rework, and waiting costs. Therefore, when discussing price advantage, one must also see whether it can reduce the number of attempts needed to obtain a usable shot.

This explains why Q4 Preview emphasizes expressive performance—according to the report, its significance lies in reducing whole-segment rework caused by character drift, broken motion, disordered shots, and disconnected effects. Up to 15 reference images, 3 reference audio clips, and 2K, 4K output cover practical aspects such as character control, voice continuity, and delivery quality.

But to provide these capabilities at a lower price, computing efficiency must also be solved. Video generation is a computationally intensive task: as duration increases, the model needs to maintain a longer chain of states; as resolution increases, the visual information involved in computation grows rapidly. For Transformer-based diffusion models, attention computation in long sequences is especially expensive, and inference speed directly determines the computing cost per generation.

The report connects the price to Shengshu team's past efficiency research in a line: SageAttention attempts to accelerate attention with low-bit computation; the subsequent SLA combines sparse attention with linear attention to further compress long-sequence computation; TurboDiffusion then combines attention acceleration, few-step generation, quantization, and engineering optimization.

Among them, TurboDiffusion is reported to have achieved hundred-fold acceleration under specific open-source video models and hardware conditions, and the related technology has been productized in Vidu S1's real-time video generation. This 'hundred-fold' has clear limiting conditions and cannot be extrapolated as general performance.

The significance of this technology line is to explain where the low price comes from: speed improvements shorten waiting time and also increase task throughput per unit of computing power, giving cost room to continue declining. Based on this, the report argues that the value of the price cut should not be understood only as a short-term promotion; the debut discount is just the entry point, and the long-term significance depends on whether technical efficiency can continuously translate into lower generation costs.

Machine Heart provided three sets of first-hand hands-on tests, all of which are qualitative descriptions from a single source. The first is an emotional monologue: the setting shifts from restraint to outburst, with five emotions occurring in sequence—holding back grievance, tears falling, laughing through tears, slamming the table in outburst, and slowly calming down. The only reference materials are one image of a female character and one image of a wooden kitchen environment, with a fixed close-up throughout and no cuts. The report says the completion is very high, with no obvious 'face-swap feel' at emotional transitions, and facial features, short hair, and clothing show no obvious drift under large expressions and multiple rounds of light-dark changes.

The second tests shot choreography: a 16-second ad sequentially completes a desktop wide-angle push-in, foreground occlusion by a character's gesture, a red circular transition, product matching and catch, overhead group shot, low-angle character, arc group camera movement, and final product freeze. The report says the shot scale spans widely but the rhythm does not fall apart, and the product maintains high recognizability across near and far shots and multi-person interaction.

The third tests effects: a static home image is made into an 8-second 'space assembly show,' with a sofa, floor lamp, side table, wall art, green plants, and throw pillows leaping in one after another, each piece of furniture having compression, rebound, and slight shaking when landing. The report says the shape, scale, and landing points are stable, with no obvious interpenetration or positional confusion in the frame, and the transition from the latter half's dimming environment to the floor lamp lighting up is natural.

The common point of these three tests is that they all test the model's ability to 'hold the whole picture steady,' but the sample size is small, there is no control, and there are no quantitative indicators, so they can only serve as directional reference.

The report mentions that in overseas creator communities, 'Vidu flavor' has formed a reputation: some creators originally doing abstract art style switched to anime after hearing that Vidu's anime capability is strong; a Japanese AI anime creator has expressed approval multiple times; AI anime works entirely generated with Vidu have received hundreds of thousands of views on YouTube. These are all qualitative retellings, with no verifiable quantitative indicators provided.

The report attempts to explain this style perception with a technical route: anime has high requirements for motion coherence, camera scheduling, character consistency, dynamic details of hair and clothing, particle effects, and emotional outbursts. Being able to stably generate high-energy anime often means the model has stronger control over time, motion, and spatial relationships.

This connects to Shengshu's self-positioning—a general world model company. In its five-level roadmap, Vidu mainly corresponds to the first level, 'generating the world,' responsible for learning characters, objects, motion, spatial relationships, and scene dynamics; the Vidu S series further adds real-time interaction, while Motus and Motubrain extend predictive capability to physical action. The report does not explain the landing time and product form of these subsequent levels.

The report lists Vidu's technical milestones: in 2022, it was the first to propose the U-ViT architecture; in 2024, it released China's first video model comprehensively benchmarked against Sora, and pioneered 'reference-to-video' and multi-subject consistency; Vidu 2.0 compressed generation time to within 10 seconds; Q3 achieved 16-second synchronized audio and video. These milestones form the background for understanding Q4 Preview, but they also come from one side of the report.

Separate the known from the unknown: what is known is the preview version's capability direction, debut price, speed claims, and reference material limits; what is unknown is the specific improvement of the official version in stability, controllability, and creative ceiling, whether the 0.09 yuan/second and up price will be maintained long-term, and the consistency effect of up to 15 reference images and 3 reference audio clips in actual workflows—the latter lacks independent evaluation.

For advertising, short drama, anime, and e-commerce content teams, what truly determines whether AI video can enter formal budgets is not the debut price, but the 'cost of a usable final cut': a single generation is cheaper, and with the model's more stable control over characters, shots, and reference materials, the number of attempts needed to obtain a usable asset will decrease. Whether this chain can close must wait for the official version and independent verification to answer.