Vidu Q4 Preview Targets Character Performance With 0.09 Yuan-per-Second Video Generation

ifanr reported that it received early access and tested the model as if it were an actor auditioning for a role. The preview supports up to 15 reference images and three reference audio clips, along with optional 2K and 4K ultra-high-definition output. New users who register at vidu.cn with the invitation code APPSOQ4 can receive 500 credits, ifanr said.

The tests focused first on reactions, which ifanr described as the smallest unit of a good performance. It used a Dune scene in which a queen moves from shock to fear and anger inside a deep-blue sand palace, ending with a single look; the scene has almost no action or blocking and is carried by close-up shot-reverse-shot coverage over 16 seconds. ifanr said the AI result did not match the original, but the sequence of hearing, pausing, moving the eyes and slowly turning the head resembled how a human actor plays the realization of a truth. A second test used a scene from King of Comedy that ifanr said is taught at the Beijing Film Academy: a character calls after a woman in the rain, she turns back with a smile, says he should support himself first, leaves, and then breaks down in the back seat of a taxi. ifanr found that Vidu Q4 Preview could perform the flow from laughter to tears, but its overall rhythm was not yet natural enough. That fits the preview's purpose, ifanr said: putting the model in creators' hands to expose problems and collect feedback before a formal release.

When the performance did not require a sharp emotional turn, ifanr said Q4 Preview held character states more steadily. In another Dune case, a wounded soldier answers that it is not his vision, and the camera stays on a tired, alert face; stitched wounds around the eyes were recognizable across multiple sample points. ifanr said this more restrained acting looked more natural. In action, the model also showed Vidu's familiar style. ifanr placed Black Myth: Wukong's Destined One and Phantom Blade Zero's Soul in a ruined temple in heavy rain for a fight. The prompt specified a full combo, including a sliding dash, a head-level slash, a staff block, a twisting counter-sweep, a vault over the staff and a landing pursuit slash. ifanr said the model presented the fight with different camera moves and did not rely on frequent close-ups to hide the two characters' positions, a common trick in earlier AI fight videos.

Camera movement was another focus. ifanr described Vidu's signature as knowing why the camera moves. It tested FPV shots, including an escape from inside a sandworm's mouth; after a large roll, the model still found the same bright gap between teeth, and the exposure shift from dark to bright was blocked by the real exit. Another shot imitated an ancient ship escaping a cyclops in The Odyssey. ifanr said the giant's presence affected the scene's lighting, and the video cut from a sky-covering arm to an open golden sea, making clear that the danger had passed. A more complex Dune shot moved from a crowd panorama to riders on top of mechanical sandworms, combining tracking, orbiting and elevation. ifanr said direction and subject stability were strengths, while style control and the hallucination problems common to AI still need optimization.

For effects, ifanr tested waves, black holes, ice canyons, explosions and mechanical sandworms. In an FPV shot, a kilometer-scale tidal wave rises on a sunset sea; water streaks stretch along the wave wall and reflections change as the view tilts. A space-station escape video with explosions and a black hole was clean, and an ice-canyon video moved from falling ice to water to an underground sea, where a blue fluorescent ray-like giant beast emerged from silhouette into the foreground. With 2K and 4K options, ifanr said firelight could fall on people and smoke could affect sightlines, so effects interacted with the world. ifanr added that these results were not generated from a single prompt. Reference images worked like a stack of pre-shoot photos, fixing identity, spatial layout and action targets. Vidu was one of the first video models to offer reference-to-video, and Q4 Preview supports up to 15 reference images; ifanr said assigning each image a clear role matters more than the number.

ifanr framed the preview as a two-way audition. Shengshu Technology is giving the model to creators to see how it performs in real workflows, while creators are judging whether it belongs in future projects. The report said 0.09 yuan per second changes the economics by allowing multiple attempts instead of one expensive generation. Different resolution tiers lower the barrier, letting creators test at low cost and finish at higher quality. ifanr compared this with film production, where the expensive part is not only the few seconds kept but the dozens or hundreds of takes discarded to find them. Vidu Q4 Preview is still being optimized, but when expensive, hard-to-repeat shots can be retried for cents, AI video can let creators audition a costly good shot at lower cost.

The article opens and closes with an analogy from Marlon Brando's audition for The Godfather, where he won the role of Corleone by stuffing cotton in his cheeks and reading lines on camera. ifanr said the model is now auditioning for its next scene. The preview is available at vidu.cn, and the company's invitation code APPSOQ4 offers new users 500 credits, according to the report.