Among the visual generation models coming out of China, this is one of the names that cannot be skipped. Tongyi Wanxiang is a visual generation model built in-house by the Tongyi lab, positioned on its own site as a leading model for AI video and image generation. The version featured on the front page is the third-generation Wan, an all-in-one arrangement that folds text to video, image to video, multimodal reference and video editing into a single model.

Thirty seconds in one pass, with sound

The official documentation states that the third-generation series is an all-in-one video model, upgraded across audio-native video, multimodal reference and video editing: a single generation of up to thirty seconds, output at thirty frames per second, with dialogue, background music and sound effects produced natively. Put those numbers together and most of the old routine — generate silent footage, hire voices, sync the tracks — falls away.

Text to video needs only a prompt, with no material to upload. The documentation says it natively supports multi-shot narrative up to thirty seconds and generates matching dialogue, music and effects on its own. Single-shot and multi-shot formats are both available, each shot lasting four to six seconds and carrying a time stamp. Resolution comes in three steps up to full high definition, duration runs from two to thirty seconds, and there is an automatic duration setting where the model decides how long the piece should be.

References: images, video, audio, and documents too

This is the part of the third generation worth its own section. A single request accepts up to twenty multimodal references, spanning images, video, audio, documents and web pages. The specific limits: ten images at twenty megabytes each, five video clips totalling no more than fifteen seconds, five audio tracks totalling no more than fifteen seconds, and the three kinds can be combined freely — images with video, images with audio, or all three together.

Documents and web pages are the rarer trick. Set the reference type to a file or a link — one of the two, one item each — and it will parse the contents of a document or a page and build the video from them. A product manual, a planning document or a data table can become the structure and the visual direction of a clip.

First and last frames, and editing afterwards

On the image-to-video side, both a first-frame mode and a first-and-last-frame mode exist: two stills pin the opening and the closing picture with pixel-level fidelity to the reference, and the motion in between is left to the model. In the two-frame mode it generates the transition by itself, which lands closer to the shot in your head than a sentence of description.

The documentation also covers editing and extension. For editing, set the material type to a reference video and let the prompt carry the intent — remove, replace, turn into — and elements can be added, deleted or changed, styles transferred, lighting adjusted and dialogue rewritten. Extension works forward, backward or both ways. In other words, what you are unhappy with can be fixed inside the same model, and the footage can keep going.

The image side is not an afterthought

Video is the main stage, but images are not weak. The site describes the system as a video and image generation model, and the image side covers text to image and image to image, with multi-image blending and creative recombination, commercial-grade consistency for a subject across outputs, transfer of aesthetic elements, and precise control over lens and lighting. It can also produce mixed layouts of pictures and words, generating several images and the text together, which suits picture books and infographics where the arrangement has to make sense. Realistic portraits are an explicit direction, with composition and light tuned to reduce the giveaway look of generated imagery.

The family and how to reach it

The release page states that the Wan family now covers more than ten visual creation capabilities — text to image, text to video, image to video, voice-driven video and motion generation among them — and has produced three hundred and ninety million images and seventy million videos to date. More than twenty models have been released openly in succession, with over thirty million downloads across open-source communities and third-party platforms. In daily use the model can be tried on the official site, or called through a cloud model service for batch production.

The boundaries, stated

A few limits in the documentation are worth knowing first. Document and page links only work for public pages that need no sign-in; anything behind a login cannot be fed in. The reference caps apply as listed. And the two-frame mode cannot be mixed with other reference types — supplying an opening and a closing frame rules out an audio reference in the same request. These are not traps, they are constraints to price in before choosing.

To sum up: what the third generation closes is the gap between a script and a finished cut that used to require several tools. Thirty seconds, with sound, multi-shot, referenceable, editable — all of it inside one model. Anyone working on short drama, advertising or commerce video should put it on the trial list. Anyone who only wants the occasional picture will find the image entry point sufficient.