Shengshu Technology releases Vidu Q4 Preview: up to 15 reference images, 2K/4K output, first-release pricing from 0.09 yuan/second
According to Machine Heart (Jiqizhixin), Shengshu Technology has introduced Vidu Q4 Preview, focused on character performance, camera language and complex visual effects, supporting up to 15 reference images, 3 reference audio clips and 2K/4K output, with first-release pricing from 0.09 yuan per second; the report says image-to-video speed is about 2x that of comparable models. The performance conclusions come from that outlet's own testing. This version is not in the live KimiSeek catalog.
What changed
According to Machine Heart (Jiqizhixin), Shengshu Technology released Vidu Q4 Preview, positioned as the first preview version of Vidu Q4 aimed at creators. The report says the version strengthens three areas: character performance (continuous emotional transitions and identity stability under fixed close-up long takes), camera language (multi-shot-size switching and movement continuity), and complex visual effects (multi-object elastic deformation and dynamic lighting). On specifications, it supports up to 15 reference images and 3 reference audio inputs, plus 2K and 4K output. On pricing, the report says first-release pricing starts at 0.09 yuan per second and claims the same budget can generate up to 5x the content. On efficiency, the report cites data saying image-to-video speed is about 2x that of comparable models, with a 5-second video completing in as fast as about 110 seconds, and attributes the efficiency to Shengshu's work on attention acceleration (SageAttention, SLA), few-step generation and quantization (TurboDiffusion), saying the related technologies have already been productized in Vidu S1 real-time video generation. The report also says the model is still in preview and a formal version has not yet launched. It gives the official Vidu.cn experience entry and an invitation code, and does not say whether an API is offered; the capability, speed and quality descriptions above are that outlet's reporting and its own test conclusions, not independently verified by this article.
What it means for developers
The direct impact for developers is in the material control slots and delivery specifications, not merely in picture-quality descriptions. The report puts the reference-image limit at up to 15 and adds support for 3 reference audio clips, meaning character appearance and voice control could potentially be completed within a single call, which would change how character-consistency pipelines are designed: the common past practice of repeated retries with a first frame plus prompt could shift to submitting multi-angle reference material in one go. 2K/4K output touches the delivery stage; producing at 1080P and then upscaling has been common practice, and if 4K direct output holds up in stability, the post-processing chain can be shortened. But the report does not say whether the 4K tier's time and billing match the lower-resolution tiers, which needs to be verified independently. On efficiency, the report says image-to-video speed is about 2x that of comparable models and a 5-second video completes in as fast as about 110 seconds; if reproduced under one's own workload, batch throughput and output per unit of compute would improve, which is especially relevant for short-drama, advertising and e-commerce teams that need to shoot multiple versions in parallel. On price, 0.09 yuan per second is the first-release framing reported as of October 2026; actual cost should be weighed together with the usable-output rate: a low per-unit price with a high drift rate may have screening and rework costs that offset the price advantage, so a lower unit price alone is not enough to prove total cost has fallen. Note that the current live KimiSeek catalog contains no Vidu Q4 Preview or Vidu-series entry; this article does not describe it as a model available on this site. The report provides only Shengshu Technology's official experience entry and first-release pricing, and does not say whether this version offers an API; for interface capabilities, check with Shengshu Technology's official channels.
Integration and evaluation
Both of the following are conditional analysis and suggestions, not reported facts. First, whether to move character and voice control to a single multi-reference call: if the workflow needs to keep the same character and voice across continuous shots and you already have 5 to 15 multi-angle assets of the same character, you could try changing the existing first-frame-plus-prompt-plus-retry process into one submission of multiple reference images plus reference audio; the precondition is first confirming with Shengshu officially the actual accepted limits on reference images and audio and whether they affect billing. It is advisable to start with a fixed close-up long take of no more than 15 seconds, measuring the rate of face and clothing drift for the same character within the same video and the average number of retries needed to get a usable shot, and compare that with the current process. Second, whether to bring 4K direct output into the delivery chain: if delivery requires 4K and you currently rely on post-processing upscaling, you can compare 4K direct output against generating at 1080P and upscaling in terms of time and unit price per usable output; the precondition is confirming the 4K tier's billing method and generation time for this version, since the report gives only a starting price and a fastest speed of about 110 seconds for 5 seconds, without breaking it down by resolution tier. It is advisable to generate several clips each at 2K and 4K with the same prompt, measuring actual time and cost per usable output rather than single-request duration. On the evaluation steps, it is advisable to first build a small benchmark set: choose prompts containing long takes with strong emotion, multi-object effects and multi-shot product advertising, run fixed numbers of attempts on Vidu Q4 Preview and the existing approach, record the usable-output rate, average retries, per-clip time and the distribution of drift types, then decide the scope of migration. For pricing and quotas, refer to Shengshu Technology's official pricing page; first-release framing in historical reports should not be assumed to remain valid long term.
Limits and open questions
First, Vidu Q4 Preview and the Vidu series do not currently appear in the live KimiSeek catalog; this article does not indicate that this version can be called on this site, nor that it has launched on any third-party aggregation platform. Second, the report gives only the Vidu.cn experience entry, first-release pricing and an invitation code, and does not mention API, interface or API availability at all, so this article does not treat offering an API as part of the report; for interface capabilities, quotas and billing methods, check with Shengshu Technology's official channels. Third, the test conclusions in the report (emotional continuity, shot composition, effects stability, speed about 2x faster than comparable models, about 110 seconds for 5 seconds) come from the outlet's own testing, and sample size, prompts, control group and hardware conditions were not fully disclosed, so they cannot be treated as independently reproduced results. Fourth, this version is a preview, and it remains uncertain whether the formal version will change in stability and controllability, or whether pricing will continue the first-release framing. Fifth, the report says up to 15 reference images and 3 reference audio clips, but does not say whether different input combinations affect time and billing; price and speed differences between the 2K and 4K tiers are also not given separately. Sixth, the report attributes part of the low price to efficiency work such as SageAttention, SLA and TurboDiffusion, which is a link between research conclusions and existing productization descriptions, and does not prove their actual share of contribution in Q4 Preview. Seventh, the new-user credits and invitation code mentioned in the report are short-term promotional information and should not be used as a basis for long-term availability.