Kuaishou's prompt-first flagship: cinematic video up to 4K, and Director Mode that composes up to six shots — cuts included — in a single generation. Type a prompt below and try it.
Developer
Kuaishou
Type
Prompt-first T2V / I2V
Duration
3–15 s per clip
Max output
Up to 4K
Audio
Native · multilingual lip-sync
Director Mode
Up to 6 shots per gen
Tier here
Paid credits
Our one-line verdict: the director pick. When the brief is a scene with cuts — not a single shot — this is the only model here that edits while it generates.
Kling 3.0 — the prompt-led half of Kuaishou's 3.0 generation, rolled out at the end of January — competes on a different axis than the long-take models: instead of one unbroken 30-second shot, Director Mode composes up to six distinct shots, cuts included, inside a single generation. Add the highest resolution ceiling we host (4K) and lip-synced dialogue across five languages, and its lane is clear: finished scenes, not raw clips. Its reference-driven sibling Kling O3 gets its own page.
Judged from our own renders and the published record, revised as the facts move — that's a feature of this page, not a disclaimer.
• Director Mode.
Up to six shots — wide, medium, close, cuts between them — composed in one generation with consistent characters and setting. The closest thing to editing-while-generating in our lineup.
• 4K ceiling.
The highest output resolution we host. When the deliverable is a broadcast spot or a big screen, this is the only lane.
• Multilingual lip-sync.
Native audio with synced dialogue in English, Chinese, Japanese, Korean and Spanish — localized versions of the same ad without re-shooting.
• In-video text rendering.
Signs, labels and on-screen titles come out legible more reliably than the field — a quiet superpower for ads and product work.
• Family naming maze.
V3 vs O3 vs 2.6, Standard vs Pro — Kuaishou's lineup takes a diagram to explain. Rule of thumb: prompt-driven work here, reference-driven work on Kling O3.
• 15-second ceiling per generation.
Six shots share at most 15 seconds. For one long unbroken take, Seedance 2.5 and Wan 3.0 reach 30.
• Reference control is the sibling's job.
You can feed it an image, but locking a character across many generations is what O3's MVL architecture was built for — use the right twin.
• 4K costs real money.
Per-second pricing at 4K adds up fast. Draft at 1080p, upgrade only the keeper.
Copy into the generator above, swap the brackets, render.
"Director Mode, three shots: (1) wide establishing shot of a [rainy Tokyo café] at night, (2) medium shot of a barista pouring latte art, (3) close-up of the finished cup as a hand reaches in. Warm interior vs cold street light, rain and espresso sounds throughout."
Numbering the shots is the syntax Director Mode rewards — each number becomes a cut. Name the through-line (the cup, the light) so the shots read as one scene. The demo beside the generator is this exact prompt.
"A [skincare] spokesperson in a bright studio holds the product and says, lip-synced in Japanese: '[tagline]'. Then regenerate the identical prompt with the line in Spanish."
Same prompt, swapped language — the lip-sync tracks each. This is how teams cut localization costs: one creative direction, native mouths in five markets.
"Slow push-in on a weathered storefront at golden hour. The hand-painted sign reads exactly: '[MORELLI & SONS — EST. 1962]'. A cat sleeps in the window. Street ambience, distant traffic."
'Reads exactly' plus quotation marks is the instruction that engages the text-rendering strength. Keep it under ~6 words per sign; verify in preview before a 4K spend.
Three different theories of what a generation should produce: a scene with cuts, a realistic shot, a long take. Same prompts across all three, judged by what we'd ship. All three selectable in the generator above.
Two models in the same generation, split by workflow. Kling 3.0 (V3) is prompt-first: you describe, it directs — up to 4K, with Director Mode composing up to six shots in one generation. Kling O3 (Omni) is reference-first: you supply images or video of a subject and it locks that identity. Prompt-driven work here, reference-driven work on the O3 page.