
Thirty Seconds in One Take, or How Kling 4.0 in Unitool Remembers Your Hero
Your hero looks different in every clip, and the cuts show on the first watch. Kling 4.0 makes up to 30 seconds in one pass and keeps a character consistent from photos. Here is what the model can do, what a take costs in Unitool and how we shot an ad without a film crew.
Thirty Seconds in One Take, or How Kling 4.0 in Unitool Remembers Your Hero
The most painful part of AI video happens in the edit. The first clip comes out great. In the second one your barista has a different nose, and by the third the coffee bag has changed color. You stitch them together, and viewers can tell within two seconds that they are watching three different videos.
Kling 3.0 topped out at 15 seconds per generation, so a 30-second ad had to be built from scraps. On October 15, 2026, Kuaishou, the company behind Kling, opened Kling 4.0 to everyone. The headline number is up to 30 seconds in a single pass. The second is up to 15 references in one request. References are the photos, videos and voices the model uses to keep your characters and products consistent. We have already added Kling 4.0 to Unitool, tested it on an ad, and below we show how to shoot with it.
Kuaishou’s official Kling 4.0 video. This is how the company shows off long takes and camera work, and below we test what it delivers on an everyday task.
There is no actual shoot. There is a panel on the right that hides an entire film crew.
What Kuaishou changed in version four
The model was announced on September 28. At that point a lighter Kling 4.0 Flash opened only to annual subscribers on Kling’s top plan. The full version followed on October 15. Kuaishou sums up the update in three words, realism, control and professional quality. In terms of everyday tasks, that means five changes.
Length. One clip runs 3 to 30 seconds with no cuts, twice the limit of Kling 3.0.
Motion. The company promises stable footage in fast scenes and complex camera work, including push-ins, pull-outs, pans, trucking shots, tracking shots and orbits.
References. Up to 15 per request, including up to 10 images, up to 5 videos and up to 7 characters or objects. Images can be ordinary photos, three-view character sheets, storyboards or even interface wireframes.
Keyframes. Up to 10 anchor images the video has to pass through. They let you decide where the hero starts, where they are halfway through and how the scene ends.
Sound. According to Kuaishou, audio is now stereo, and lip movement matches speech more closely.
One more useful detail is that a scene description can run up to 8,000 tokens. A token is a chunk of a word that the model uses to count text. That leaves room for a full script timed to the second.
All of these numbers come from Kuaishou’s announcement. The company has not published independent tests or comparisons with other models, so you will still have to judge every take with your own eyes.

What is promised but not here yet
Some features in the announcement are marked as coming soon. Extending a video to two minutes does not work yet. 10-bit HDR output is also listed for later. It gives richer color in dark and high-contrast scenes. In Unitool you can currently choose 720p or 1080p.
Kuaishou also lists the languages the model can render as text inside the frame. They are Chinese, English, Japanese, Korean, Spanish, Portuguese, German, French and Hindi. For a label or sign in any other language, upload a photo with the finished text instead of asking the model to draw it.
A film crew in the right-hand panel
Open Kling 4.0 from the model list in Unitool. The scene description box sits in the center, and the settings panel is on the right. It helps to read the panel as a list of people you would otherwise have to hire.

The camera operator handles the aspect ratio. You get 16:9 for YouTube, 9:16 for Reels and Shorts, square 1:1 and the cinematic 21:9 that is new for Kling. The Auto option picks the ratio from your attached photos.
Casting is the Characters and objects block. You can add up to 7 of them, no more than 3 from video. Each one gets a name and 2 to 4 photos. The first photo should face the camera, and the rest show the hero from other angles. A single-shot clip of 3 to 8 seconds works instead of photos. There is also a short description field of up to 100 characters.
The sound engineer is the Voices block. You can add up to 3 voices built from a sample recording. A voice can also be attached directly to a character, in their card.
In the scene description you call everyone with an at sign, like @Mia. Photos attached to the input box become @image_1, @image_2 and so on. You can attach up to 10 images in PNG, JPG, JPEG or WebP, up to 10 MB each. That includes images you have already generated in Unitool.
The editor lives in the Mode menu. Text to video is the default, and there are two more modes worth knowing.
Two more modes for when you already have footage
Frames to video is for when you know exactly how the clip should start and end. Attach the first frame with the paperclip, and it becomes @first_frame in your description. Below that you can add up to 8 middle frames (@keyframe_1, @keyframe_2) and a last frame, @last_frame. The model will take the scene through every point in order. There is one limit, since the number of frames cannot exceed the number of seconds plus one. All 10 frames are therefore available in clips of 9 seconds or longer.
Video to video reworks clips you already have. Upload up to 5 of your own videos, 2 to 30 seconds each and no more than 30 seconds in total. Then describe what to change, such as “repeat the camera move from @video_1, but make it a winter evening.” Or take an ad that already works and swap the car in it for the one in your photo. The Keep original audio toggle preserves the source sound. Switch it off, and the model will create the soundtrack itself.
Pricing works differently in this mode. The seconds of your own videos are billed too, never less than the length of the result. That is why 5 seconds cost 128 tokens here and 108 in Text to video.
One ad from storyboard to final take
We tested Kling 4.0 on the task where people usually stumble, a short ad for a small coffee roaster. It needs the same barista in every shot and a bag you recognize at first glance. We kept the clip to 10 seconds, which is enough to show the product and deliver one line.
We started by casting. If you do not have photos ready, you can make them right here. We generated every reference for this clip, three angles of Mia and the coffee bag, in Nano Banana 2.1 in Unitool, at about 12.5 tokens per image.
In Kling 4.0, under Characters and objects, we created @Mia and uploaded three photos, front, three-quarter and profile. In the short description field we wrote “Barista, dark hair in a ponytail, freckles, olive-green apron.” We did not add a voice, and the model voiced her on its own. We attached the bag photo to the input box, and it became @image_1.

The settings were Text to video, 720p, 10 seconds and Auto aspect ratio. We wrote the description like a storyboard, with timing.
@Mia picks up the coffee bag @image_1 from a wooden shelf in a bright, cozy café and smiles. 0 to 4 seconds: medium shot, the camera slowly pushes in as Mia sets the bag on the counter next to a cup, steam rising from the cup. 4 to 10 seconds: the camera smoothly orbits the counter and stops on a close-up of the bag and cup, the label text ‘MORNING ROAST’ and ‘with Unitool AI’ in focus and readable. Mia looks into the camera and says: ‘Roasted on Monday, in your cup by Wednesday.’ Warm morning light from the window, quiet café background noise, no music.
The price shows up under the input box right away, 215 tokens. You pay for every second, so it is easy to estimate in advance.
The clip worked on the first try. Kling chose a wide 16:9 frame and added stereo sound. Mia is recognizable in every shot, even with her back to the camera. The label text is clean, including the small “with Unitool AI” line, and the model filled the shelf with more bags just like ours on its own.
If you would rather not write the description at all, you can go a step further. Turn on Unitool MCP in a Claude or ChatGPT chat and describe the task in plain words, such as “make a 10-second café ad from this photo of the bag.” The model picks a video model, asks about the details and shows a card with the settings, prompt and price. Nothing is generated until you press Run. MCP is a way to connect Unitool’s models to a chat or to an agent like Claude Code so they can call the right model on their own. Text, images and video all come out of the same balance.

How not to burn tokens on retakes (H3)
A long clip is not expensive by itself. Failed attempts are. This order keeps your balance in check.
Shoot a 3 to 5 second test at 720p first. For 65 to 108 tokens you will see whether the hero’s face holds and whether you like the light.
Change one thing per test, whether the light, the angle or the line of dialogue. Otherwise you will not know what fixed it.
Switch to 1080p only for the final take. Three seconds at 1080p cost 96 tokens, compared with 65 at 720p.
If there is a person in the shot, give the model photos from several angles. A single selfie is not enough for the hero to stay recognizable when they turn.
A 65-token test is almost always cheaper than reshooting a full 30 seconds. We recommend at least one test before every long take, even when you are itching to see the finished ad.

Why pay for 4.0 when Kling 3.0 Turbo exists?
It is a fair question. In Unitool, Kling 3.0 Turbo starts at 70 tokens for 5 seconds, while 5 seconds of Kling 4.0 cost 108. For one short shot with no recurring character, Turbo will do just fine.
The difference shows up once a video runs longer than 15 seconds or needs the same person in every scene. You can stitch 30 seconds together from two Turbo clips, but your barista may look slightly different at the seam, and then you redo both halves. Kling 4.0 shoots the whole story in one pass and keeps the hero tied to their character card.
Unitool also has Seedance 2.5. It too makes up to 30 seconds in one take with sound, and it accepts up to 50 references. Seedance rejects real faces, so it shines in product and landscape videos. For a story with a recognizable host, Kling 4.0’s character card is handier. We would pick the model for the scene, not for the version number.
Your First Take This Week
Any Unitool plan gives you access to Kling 4.0, and one token balance covers every model. It is worth planning your spending ahead. A 30-second clip at 720p costs 646 tokens, so for regular video work the Advanced plan or a one-time Tokens purchase makes more sense. Unused tokens do not expire at the end of the month.
Do not start with a 30-second ad. Start with a test. Upload three photos of your hero, attach the product, shoot 5 seconds for 108 tokens and see whether you recognize your character.
Shoot your first take in Kling 4.0
If you do, the remaining 25 seconds are just a matter of storyboarding.