The site introduces itself plainly, as a next-generation generator of video and images. The top navigation splits the product in two, a creative studio for people who make things and an interface for developers. One detail is worth flagging: the international home page carries a notice saying the international site has moved to a new address, with a button to follow it. The studio aimed at mainland users keeps its own tool shelf, listing image generation, video generation, motion control, a canvas, an agent and a second-generation digital human, and the membership banner there advertises a price under six yuan a month.

One family, five members

The line currently pushed is the third generation, sold under a slogan that rhymes in Chinese and means everything in one, one for everything. It is not a single model but a family: video, video omni, image, image omni, and an element library. The official framing says the family sits on a rebuilt foundation and places video, stills and element material into one process, so a team that needs both clips and pictures can reuse one style reference across modalities.

Audio generated with the picture, not after it

The most consequential upgrade in this generation is sound. The documentation says the video models parse multimodal instructions deeply and fuse tasks across them, rebuilding how light and sound carry a story, and describes the result as a dual binding of visual identity and vocal tone, from accurate shot breakdowns in very long clips down to the decoupling of audio and image features. What that means in practice is simple: nobody has to render a silent clip and then dub it to match the mouth, because voice and picture are bound at generation time.

The native 4K claim, and where it comes from

An entry on the official blog is titled the world's first native 4K video model, and its opening paragraph dates the release to 23 April, when the third generation model brought native 4K generation. Native has a specific meaning here: the company says users generate genuine 4K directly, and quotes a partner explaining the difference from upscaling tools, which tend to alter a character while enlarging it, whereas this model builds at true 4K from the start and so avoids the problem. The same article gives the measurement standard, namely native output at 3840 by 2160 produced during generation.

Making a body move, and making a face speak

Beyond turning text or a picture into footage, two capabilities get their own billing. Motion control rests on motion capture, reproducing complex movement and accurate facial expression, and the typical use is a reference video plus a portrait, so the result inherits the body language and the expression of the original. The digital human tool takes a portrait and a voice recording and returns a convincing performer with natural delivery and lip sync. Both sit in a popular section on the home page, and each has documentation on the developer side.

The tool shelf itself

The studio groups its tools into six: image generation, video generation, motion control, a canvas, an agent and the digital human, with omnimodal generation and a catch-all entry beside them. The video list runs longer, covering text to video, image to video, several images to video, extension, lip sync and effects, plus text to audio and video to audio. The image side covers text to image, image to image, multiple images to image and outpainting, along with vertical uses such as virtual try-on. The platform spread is broad too, with mobile clients on both major systems and desktop builds for Windows and for Macs on either chip family.

One unit name, two different rates

The developer pricing page is unusually clear about its unit, and the clarity is the point: video and image share one name but not one price. For video, one unit is 14 cents. For images, one unit is a fraction of a cent, to the tune of a thousand units for 3 dollars 50 cents. Two reader-friendly rules appear on the same page: a failed generation is not charged, and unused balance in a resource pack expires rather than rolling over.

What a second costs

Video is billed by duration, laid out in tiers. The turbo variant of the third generation leads the main table at about 11 cents a second in 720p and 14 cents in 1080p, with native audio listed as a feature. The standard variant splits further: in 1080p without native audio it is about 11 cents a second, with native audio about 17 cents, and at 4K both configurations come to 42 cents a second. Motion control runs between roughly 13 and 17 cents a second, the digital human between 6 and 11 cents, all inside the same per-second system. Images are far cheaper: about 3 cents a frame at 1K and 2K, doubling to roughly 6 cents for 4K output.

Four boundaries

Every price above belongs to the developer interface and is charged by the second or the frame, which is a different ledger from a studio membership; the membership figure visible on the China site is a starting price, and tiers and entitlements should be read from the page on the day. 4K is the expensive rung of this generation at 42 cents a second, roughly four times the 1080p rate without audio, and audio pushes it higher still, so tune parameters at 720p before scheduling real work. The native 4K wording is the company's own claim and should be attributed as such, and figures here come from the site and the blog and will move with promotions. Finally, using a real person's face or voice for generation requires the appropriate permission and respect for platform content rules.

The short version: this round is about binding picture and native audio into one pass and lifting the quality ceiling toward production work with native 4K. The per-unit rates, 14 cents on the video side and a fraction of a cent on the image side, are the more practical card it plays for developers.