We animated four drink drawings into turntables on Kling and on Seedance, then measured how much each tabletop rocked. Seedance tilted about six times as far. That settled our Kling vs Seedance default, but the bigger lesson was about when to use neither.

The short version
  • Kling held steadier. On the same turntable job, Kling VIDEO 3.0 kept the tabletop to about ±4° of tilt; Seedance 2.0 rocked about ±25°. Kling is our default for animating a still.
  • Neither holds a turntable perfectly. The model changes camera height mid-turn, and no 2D stabilisation fixes a 3D move. Generate several takes, measure them, keep the steadiest.
  • Generative video animates the subject even when you only want the camera to move. A push-in over a frozen photo is a compositing job. Render it yourself.
  • Near-still "living photo" loops read as ghostly. Prompt candid, real-time motion from a start image only, then cut the loop at matching frames instead of crossfading.
  • Generated clips arrive untagged and play brighter than their poster. Tag the colour, crop like the poster, and keep the video box no larger than the poster's.
  • We reach both through Higgsfield. As of November 2026, newer versions of both models exist that we haven't tested.
±4°
Kling tilt on a turntable
±25°
Seedance 2.0 tilt on the same job
3
takes per drink, steadiest kept
0.75×
slow motion that stayed clean on faces and hands

What we tested, and how we reach it

We use generative video for one narrow job: taking a still image, usually a drawing, a product shot or a mascot, and making it move for a website hero, a product tile or a social post. That's the only job this comparison covers.

Both models came to us through Higgsfield, a platform that puts image and video models from several vendors behind one account. Its video page lists Kling 3.0 and Seedance 2.0 among its models and describes a "first and last image reference" feature for locking a clip's opening and closing frames. Through Higgsfield we ran Kling 3.0 in its turbo, standard and pro variants, and Seedance 2.0.

What the vendors document, as of November 2026:

  • Kling VIDEO 3.0. Kling's own guide gives clip lengths from 3 to 15 seconds, native audio generated alongside the picture, and "Start & End Frames-to-Video" as a supported mode.
  • Seedance 2.0. ByteDance's launch post describes four input modalities (text, image, audio and video), up to nine images, three video clips and three audio clips in one request, and 15-second output.

Both vendors have moved on since our test. Kling has published Kling 4.0, with clips from 3 to 30 seconds and up to ten keyframes. ByteDance has published Seedance 2.5, with up to 30 seconds per generation. We haven't run either. Everything measured below is Kling 3.0 and Seedance 2.0.

The turntable test, measured

For the drinks list on our September party page with Sprizzini, each drink had a line drawing of its ingredients on a little pedestal table. We wanted each drawing to turn once, slowly, in ten seconds, with the table holding still while it spun.

The first four Kling clips were rejected on sight. The tabletop's ellipse rocked like a boat; the note on review was "bumpy ground". So we measured it. A short script tracked the left and right extremities of the tabletop through every frame and reported two numbers: how far the rim drifted vertically, and how far the ellipse tilted.

Run Rim drift Tilt
Kling 3.0 turbo 30–45 px about ±4°
Kling 3.0 pro, start and end frame set to the same still no better than turbo no better than turbo
Seedance 2.0 not recorded about ±25°

The root cause showed up once we stopped looking at the whole frame. The front rim barely moved in any take. What rocked was the ellipse's height: the tabletop got taller and flatter as the clip played. That's the camera's elevation changing mid-turn. The model isn't spinning an object in front of a fixed camera; it's re-imagining the scene from a slightly different height every few frames.

That matters because it rules out the obvious repair. Two-dimensional stabilisation, whether fitting an ellipse or pinning the rim's extremities, can't undo a three-dimensional camera move. A front-rim pin measured almost no change and fixed nothing, because the front rim was never the part moving.

Our call was "still animate, better not perfect", and that pointed at selection rather than repair. We generated three takes per drink on Kling 3.0 standard, with the start and end frame set to the same drawing and sound off, scored each take on rim drift and tilt range, and kept the lowest. Every drink improved. One drink's score fell from 13.6 to 4.6, another's from 8.3 to 6.1. The drawn objects still soften slightly mid-turn; it's the table that steadies.

Three smaller findings from the same job:

  • Direction and speed. One take in four turned 90° and came back instead of completing the rotation. A re-roll fixed it. Speed can't be tuned in the model, but direction can be fixed after the fact by reversing the clip.
  • Check at full size. We pulled frames every 1.25 seconds and read the direction of rotation at full resolution. Low-resolution grids misread it.
  • Match the prompt across the set. The same prompt, word for word, for every drink ("exactly one full 360° at constant speed, never pausing or reversing, camera locked") kept four turntables in step on one page.

If a product rotation has to be rock-steady, generative video is the wrong tool. A sequence rendered from a 3D model holds perfectly still because the camera is a number, not a guess. That's how we built the OLVR ring scroll.

Kling vs Seedance, honestly

Kling (VIDEO 3.0)

Right when: you're animating one still for a web page, the subject is an object, drawing or mascot, and you want a start frame, an end frame or both pinned.

Wrong when: the rotation or camera path must be exact, or you need the subject frozen while the camera moves.

What people underestimate: take-to-take variance. The same prompt gave a clean rotation three times and a 90°-and-back once. Plan on generating several takes and choosing.

What people overestimate: what the pro tier buys on this kind of job. With the start and end frame pinned, the pro take wobbled the same as turbo.

Seedance (2.0)

We ran Seedance 2.0 in the turntable test and haven't shipped a clip made with it, so beyond the tilt figure this section leans on ByteDance's documentation.

Right when: the job is driven by references rather than a single still. Its launch post describes taking images, video clips and audio together in one request, and planning camera language from the prompt. That's a different tool from one-still animation.

Wrong when: in our test, a locked-camera turntable from a single drawing. It rocked about six times as far as Kling.

What people underestimate: how hard a model built to plan camera moves may fight a locked-camera brief. That's our reading of the tilt, not something we've proven.

What people overestimate: how far one test generalises. Ours covers one subject type on one version. Seedance 2.5 is out and we haven't measured it.

Kling VIDEO 3.0 Seedance 2.0
Turntable tilt, our test about ±4° about ±25°
Clip length, per vendor 3–15 s up to 15 s
Inputs, per vendor text, image, start and end frames text, image, audio, video references
Used in our delivered work Yes: turntables, mascot loops, a client hero film No
Newer version we haven't tested Kling 4.0 Seedance 2.5

The ghost problem: why near-still loops fail

The hardest lesson came from people, not objects. For a bridal accessories brand's home page, we made hero films from still images. The first versions were slow, subtle, ten-second loops: the start frame set equal to the end frame so the clip returned to its first frame, and a prompt asking for calm, gentle movement with the camera still.

They were rejected as "creepy" and "ghost like", with a request for "natural movement and flow that doesn't appear as AI". The review was right. A slow blink over a frozen room is the uncanny valley. We measured it: the rejected clips changed about 0.1 from frame to frame, against roughly 1 to 1.7 for footage people accepted.

Pinning the end frame was half the problem. A model told to finish exactly where it started barely leaves. What worked:

  • Start image only. No end frame for people.
  • Ask for footage, not a photograph. "Candid wedding film footage at real-time speed, shot handheld on a 35mm film camera with a gentle natural sway."
  • Give it a sequence of actions. Glance away, laugh, tuck a strand of hair; a couple laughing and swaying.
  • Ask for "natural blinking at a normal rate" and "no slow motion".
  • Cut the loop yourself. Find the best-matching pair of frames and cut there. Our final hero used frames 3 to 123 of the raw clip.

The review then asked for a touch of slow motion on that real movement. Slowing it to 0.75× with ffmpeg's motion-compensated interpolation stayed clean on faces and hands.

Two traps. Never crossfade a loop between two different poses: it leaves a double-exposure ghost of both. Use a cut, or a blend of three frames at most. And check the late frames for things the model invents as it runs out of instructions. A veil appeared in one clip at six and a half seconds; we cut before it. When the brand wanted black and white, we graded the file locally rather than paying a model to do it.

When start-equals-end is the right call

Pinning the start and end frame isn't wrong in itself. It's wrong for people.

Teddy, our studio mascot, is a 3D vinyl-toy Bernese Mountain Dog in an olive suit. His loops for the party page were five-second Kling 3.0 clips with the start and end frame set to the same still, which return to their first frame and loop without a jump. A toy dog holding a drink is supposed to look like a toy; a little stillness reads as character, not as a ghost.

The review on those loops was about drift, not stillness. A take could go wide-eyed for a second mid-clip, which is off-model for a character whose canon eyes are heavy-lidded. So we generated two takes per pose and chose by frame strips and face crops across the whole clip, not by watching it once.

The same split runs through everything we've animated. Objects, drawings and characters tolerate a pinned loop. People need real motion and a hand-cut join.

When not to use generative video at all

The most useful thing we learned about Kling vs Seedance applies to both: they cannot leave a subject alone.

Asked for a slow push-in over a photograph of people, with the subject told to stay still, Kling 3.0 turbo and Higgsfield's own cinematic camera model both animated the people anyway. They shifted poses, hugged, changed expressions. No prompt wording or guidance setting stopped it. If the brief is "add movement to this image" and the image must not change, that isn't a generative job. It's a camera move over a frozen picture, and it belongs in compositing.

Even there, the plain tool has a catch. ffmpeg's zoompan filter snaps its crop window to whole source pixels, so slow zooms step. We measured frame-to-frame difference on a slow zoom: it alternated roughly two to one every other frame, and still did at 4× and 8× oversampling. Fast moves over busy artwork hide it; slow moves over text or UI don't. We now render slow zooms as a CSS transform in a headless browser, which composites at subpixel precision and measured as a smooth, steady rise.

Two more places we keep generative video out:

  • Screen recordings of a website. A model restyling real interface pixels warps the text. Walkthroughs are real captures, and they ship full length and uncut; every round of montage editing we tried on one made it worse.
  • Anything that has to match a real product exactly. The model redraws as it animates.

Where generative video earns its place is the opposite case: when you want the subject to come alive. A liquid pouring, a drink fizzing, a mascot raising a glass. Start from the still, prompt the motion, keep it around five seconds, and test one clip before batching.

Putting the clip on a web page

A generated clip isn't finished when it downloads. On a page, it fades in over a poster image, and three things go wrong at that handover. We measured all three in Chrome and in WebKit.

The tone jumps. Kling and Higgsfield output arrives without colour tags. Untagged H.264 rendered about eight levels brighter than the same frame saved as a WebP poster, in both engines. Tagging it as limited-range BT.709 with the sRGB transfer curve brought both within one to four levels. A gamma-2.2 tag fixed Chrome and made WebKit worse; a full-range tag made both worse. Retag the video bitstream and the MP4 container's colour box. A file already tagged in its container keeps that tag if you only fix the bitstream, and WebKit reads the container.

The crop jumps. Cut the phone version of the video to the phone frame's exact shape, centred on the poster's focal point, and use that cut's first frame as the phone poster. A 3.6% aspect mismatch alone doubled the pixel difference at the handover.

LCP moves. Google's LCP definition counts <video> elements using "the poster image load time or first frame presentation time for videos", whichever is earlier. In our measurement, when the video box was even a fraction of a pixel larger than the poster's, the largest paint moved from the poster at 0.7 seconds to the video frame at 7 seconds on a throttled phone. Make the video box one pixel shorter than its poster. If you're chasing that number, take several runs, not one.

Compositing has its own versions of this. A careless scale of a Kling clip darkened Teddy's cream background by two to four units, enough to show a visible box behind him on a cream page. Scale with accurate rounding and the colour matrix the file actually uses (BT.601, for these), and put the clip on a page background sampled from the clip itself. Don't try to cut a character out of his background with a colour-key: his white fur, shirt and glass sat inside the tolerance and got eaten.

The page markup is ordinary. MDN's <video> reference notes that browsers block autoplay of video with unmuted audio, so a background clip is autoplay muted loop playsinline with a poster. Turn generated audio off at generation time; you don't want it.

Our defaults

Kling vs Seedance turned out to be the smaller decision. The bigger one is whether the job needs a generative model at all.

  • Animating one still of an object, drawing or mascot → Kling, start and end frame set to the same still, two or three takes, the steadiest kept by measurement rather than by eye.
  • People in a hero film → use real footage if it exists. If not, Kling from a start image only, real-time candid motion, the loop cut by hand at matching frames, slowed to 0.75× at most.
  • A product rotation that must be steady → render it from a 3D model, not a video model.
  • A zoom or pan over a photo that must not change → composite it. No generative model.
  • A site walkthrough → a real screen recording, uncut.
  • Reference-heavy work with audio or video inputs → Seedance is documented for it; test it on your own material before committing, as we would.

Related comparisons: Nano Banana vs GPT Image vs FLUX: tested on brand work · AI product photography vs a real shoot: when each wins · Three.js website vs image sequence vs scroll-scrubbed video

If you're weighing motion for a launch page and want it to look real rather than generated, tell us what you're animating. Sometimes the answer is a camera move and no model at all.

Common questions

Is Kling or Seedance better for image-to-video?

For animating a single still for a website, Kling was steadier in our test. On a turntable job, Kling VIDEO 3.0 held the tabletop to about four degrees of tilt either way; Seedance 2.0 rocked about twenty-five. That's one test on one kind of subject, so treat it as a starting default rather than a verdict. Seedance 2.0 is documented as taking image, video and audio references together, which makes it a different tool for reference-heavy work. As of November 2026 both vendors have newer versions we haven't tested.

Why does my AI turntable video wobble?

Because the model changes the camera's height partway through the turn. You can see it in the tabletop: the front edge barely moves while the ellipse gets taller and flatter, which is what a shift in camera elevation looks like. Two-dimensional stabilisation can't remove a three-dimensional camera move, so tracking and pinning the rim does nothing. What helped was generating several takes per object, measuring rim drift and tilt in each, and keeping the steadiest. If the rotation must be perfect, render it from a 3D model instead.

How do I make an AI video loop without a visible jump?

It depends on the subject. For an object or a mascot, setting the start frame and end frame to the same still makes the clip return to where it began, which loops cleanly. For people, the same trick makes the model barely move, and the result looks ghostly. There, generate from a start image only, prompt real-time natural motion, then cut the loop yourself at the pair of frames that match best. Never crossfade between two different poses; it leaves a double-exposure ghost.

Can AI add a slow zoom to a photo without changing it?

Not reliably. In our tests, generative video models animated the subject even when the prompt asked only for a camera move: people shifted poses, hugged or changed expression, and no prompt wording or setting stopped it. A push-in or pan over a frozen photo is a compositing job. Render it outside the model. We found ffmpeg's zoompan filter steps visibly on slow moves even when oversampled, so we render slow zooms as a subpixel transform in a headless browser, which measured smooth.

Why do AI-animated photos look creepy?

Usually because almost nothing moves. A clip where a face blinks slowly over a frozen room sits in the uncanny valley. In our measurements, rejected clips changed about a tenth as much from frame to frame as footage people accepted. The cause was a start frame equal to the end frame plus prompts asking for slow, subtle, calm movement. Asking for candid footage at real-time speed, a handheld sway, a concrete sequence of actions and normal blinking fixed it.

Why does my video look brighter than its poster image?

Generated clips usually arrive without colour tags, and browsers then render them brighter than the same frame saved as an image. We measured about eight levels of difference in both Chrome and WebKit. Tagging the file as limited-range BT.709 with the sRGB transfer curve, in both the video bitstream and the MP4 container, brought the two within one to four levels. Also cut the video to the poster's exact shape and keep the video box no larger than the poster's.

Do I need Higgsfield to use Kling and Seedance?

No. Higgsfield is the platform we use because it puts many image and video models behind one account and one interface. Kling offers its models directly on its own site, and ByteDance's Seedance 2.0 page points to a chat app, the BytePlus playground and an API. Going through an aggregator is convenient for comparing models; going direct can give earlier access to new versions. Pricing models differ (Higgsfield, for one, charges in credits), so check each vendor's current terms.