Vidu Q4 Preview Lands Third on Image-to-Video Board

Shengshu Technology has released a preview of Vidu Q4, its next-generation video model, and an early third-party leaderboard puts it among the top image-to-video systems available. The release matters less for a single ranking than for what it signals: Chinese video-model developers are competing on expressive performance, audio, and price, and the gap to the leading models appears to be narrow.

What Shengshu announced

According to a republished copy of the company's October 7 press release, Vidu Q4 Preview is "available now." The company describes Tsinghua University's AI labs as its origin and pitches the model at short-drama producers, advertising agencies, social content creators, and film teams. Its headline claims are lifelike expressiveness, cinematic image quality, faster generation, and lower costs. The copy we could open was truncated and contained no specifications or price list, so the details below come from other sources and should be read with that in mind.

Analysis published by AlphaSignal describes the model as supporting clips of 3 to 16 seconds at 24 frames per second, with resolutions from 540p up to 4K and 10-bit output at 2K and 4K. Audio is generated at the same time as the video. The same write-up says users can supply up to 15 image references and three audio references, and that the model preserves the aspect ratio of the source image.

One notable limit: the preview offers only image-to-video and reference-to-video. There is no text-to-video mode yet, so a creator needs a starting image or a set of reference images. The 16-second ceiling also means longer scenes must be stitched from several clips.

How it ranks, and how much weight to give that

AlphaSignal reports that Vidu Q4 Preview debuted at number three on the Artificial Analysis image-to-video leaderboard with an Elo score of 1,179. That is up from number 19 and 1,056 for the earlier Vidu Q3 Pro, a jump of 16 places and 123 points. MiniMax H3 Max leads at 1,195, followed by MiniMax H3. The same article notes that Runway does not appear in the top 28 on that board.

The caveats are important. Elo scores on such boards come from pairwise human preference votes, and positions can shift as more votes arrive. AlphaSignal also points out that the 16-point gap to the leader is smaller than the board's roughly 20-point confidence intervals, which means the top three are not clearly separable on current data. A leaderboard ranking reflects image-to-video quality under the benchmark's test conditions, but does not necessarily indicate how well a model performs in real-world production.

Where sources disagree

Details differ between sources. AlphaSignal cites up to 15 reference images; the API listing for the model on WaveSpeedAI, a third-party inference platform, says its reference-to-video endpoint accepts up to 12 images and up to three audio clips. The difference may reflect platform-level limits rather than the model itself, but it is a reminder to check the current specification with the provider before building around a number.

Pricing and access

Price is a central part of the pitch. AlphaSignal reports that Artificial Analysis measures 1080p output with audio at $0.12 per second, or $7.20 per minute, compared with $9.60 per minute for Q3 Pro. The company-advertised minimum is $0.014 per second, with a company-listed 540p rate of $0.045 per second and a 4K rate of $0.39 per second. A 30 percent discount on image-to-video and reference-to-video is reported to run through November 30, 2026, though it is unclear whether every quoted rate already includes it. Shengshu's claim of up to five times more output per budget is measured against an unreleased full-price baseline, so it cannot be verified independently.

Through WaveSpeedAI, both endpoints start at $0.2125 per call after a 15 percent platform discount from $0.25, with the final cost depending on resolution, duration, and number of references. That is a different billing unit from the per-second figures above, and the two should not be compared directly. Access is asynchronous: a developer submits a job, receives an ID, and polls for the result. WaveSpeedAI lists durations as whole seconds from 3 to 16 and defaults to 720p with generated audio switched on.

Why it matters

For studios and agencies, the practical value of a video model rests on three things: whether characters stay consistent across shots, whether speech and movement line up, and what each usable second costs. Shengshu says Q4 improves coordination of facial expressions, body movement, speech and voice, and handles fast action and effects such as explosions more coherently. Those are company statements. The only outside comment in the reporting we reviewed is an early tester's remark about strong Chinese-text rendering, which AlphaSignal notes was not a controlled benchmark.

The reference-driven design is also telling. Allowing multiple image and audio references is aimed at keeping a character's look and voice stable, a common pain point for short-form drama and advertising work where the same person must appear across many clips. Starting from images rather than text trades flexibility for control, and for commercial users that is often an acceptable bargain.

The pricing pressure is the second signal. If a top-three model can be reached at roughly $7 per minute at 1080p with audio, higher-priced competitors face a harder argument, particularly for volume work such as social video and localised ads. Whether Q4 holds that position once the preview ends and discounts lapse remains to be seen.

What to watch next

Several gaps remain. AlphaSignal notes that public materials do not specify endpoint names, SDK support, codecs, file-size limits, deterministic controls, rate limits, queue latency or data-retention terms, all of which matter for developers wiring a model into an automated pipeline. Text-to-video is absent from the preview, and it is not clear when it will arrive or whether the final release will keep the preview's pricing.

The model's real-world performance remains uncertain. Current evidence is limited to a single leaderboard, where the scores of leading models are too close to establish a clear winner, and performance claims published by the company itself. As more user votes and independent comparisons become available, the rankings may change. Teams considering the model should test it with their own prompts and reference materials, verify current specifications and pricing with the provider, and review licensing and data retention terms before using it in production.

Taken together, the launch shows how quickly the image-to-video tier is churning: a model that sat at 19th place a generation ago now sits within a few points of the top. The next few weeks of leaderboard movement will show whether that is a durable shift or a debut-week effect.