Kling 4.0 or Seedance 2.5 for animating a face photo: the beauty mark test
Kling 4.0 and Seedance 2.5 are the two most talked-about video models for bringing photos of people to life. I ran the same request on both, using a beauty mark on the lip as a detail check, and I explain why Seedance may not even accept your photo.
Kling 4.0 or Seedance 2.5 for animating a face photo: the beauty mark test
If the photo is of a real person, go with Kling 4.0. Not because it draws faces better, but because Seedance 2.5 restricts photos of real faces by policy, and sometimes refuses even faces made by AI. For an illustrated character or a scene with lots of references, Seedance 2.5 is in the game, at more than double the price per second.
The headline of an article on the Russian site vc.ru promises that both "didn't even lose the beauty mark" when animating a photo of a young woman. I wanted to see that with my own eyes: I ran the same request on both models, with a beauty mark on the lip as a detail check, and worked out what each one charges.
What each model is and when it came out
These are the two big video launches of the half year, one from each Chinese giant. Seedance 2.5 is from ByteDance; Kling 4.0 is from Kuaishou.
- Seedance 2.5. Introduced in June at the Volcano Engine conference in Beijing, with clips of up to 30 seconds in a single generation and up to 50 references per request (image, video, audio and text), according to The Next Web. In August, BytePlus opened the model to businesses through the ModelArk platform. How to call it through the API and pay in Brazilian reais is in the post about the Seedance API.
- Kling 4.0. Officially launched on October 15, according to Kuaishou's announcement: also up to 30 seconds per generation, plus the Omni Reference feature, which combines up to 15 references and up to 10 keyframes. 36Kr adds 4K output and 1080p with 10-bit HDR.
Both accept a photo as the first frame, which is what everyone calls "animating a photo." And that's where the difference lies that no feature table shows.
Does Seedance 2.5 accept photos of real people?
Usually not. BytePlus, the ByteDance arm that sells the model, says in the Seedance 2.5 announcement that it applies "technical controls to restrict videos made from images or videos that contain real faces." The exception is for enterprise customers who go through real-person verification and image authorization, and even then only with faces they have authorized themselves.
In practice, the filter errs on the side of caution. On the Adobe forum, Firefly users have reported since September that Seedance 2.5 blocks even AI-created characters, mistaking them for real faces. On November 6, an Adobe community manager confirmed that Seedance failed for them too, while Kling 3.0 generated the video from the same image, and attributed the difference to each model's rules.
If you want to animate your own selfie, a client's photo or a hyper-realistic persona, this matters more than any quality difference: it doesn't help that the model keeps the beauty mark if it refuses the photo. Kuaishou doesn't disclose an equivalent restriction for Kling 4.0, which doesn't mean it accepts everything: faces of famous people and sensitive content are still blocked by the filters of practically every model.
Real test: the same request on both
A heads-up first: the tool I use for the blog's tests generates from text, not from a photo. So I didn't animate a photo. I ran the closest test: I described a face with a small, easy-to-check detail, and asked for a movement that forces the model to show that face from two angles. It's the same challenge as animating a photo, minus the photo.
The prompt, in English, on both models, in 9:16, 720p and 5 seconds, through the Cut.Pro Generator: a vertical close-up of a woman in her early twenties with a small, dark beauty mark just above the left corner of her upper lip and light freckles on her nose, window light in a small apartment. She starts in profile, turns her face to the camera, smiles, laughs and tucks her hair behind her right ear. Phone in hand, realistic skin, no makeup. Both were ready in under 5 minutes.
Kling 4.0, 5 seconds. It cost 105 credits.
Seedance 2.5, 5 seconds. It cost 222 credits.
What I saw in the 1-second frame of each:
- Side of the beauty mark. Seedance 2.5 put the mark above the lip, on the left side of her face, as the prompt asked. Kling 4.0 put it on her right side (the viewer's left). For an invented face, it doesn't matter. For animating someone's photo, it's exactly the kind of swap that gives the video away: anyone who knows the person notices right away.
- The mark itself. In Kling, it's round, black and too sharp, like a sticker. In Seedance, it's smaller, lighter and sits in the skin, closer to a real beauty mark.
- Movement. At 1 second, Seedance is still in profile with her eyes closed, and Kling is already facing the camera. Kling shortened the turn; Seedance followed the order of the request.
- Video look. Kling nailed the "phone in hand" feel: T-shirt, white wall, side light, selfie framing. Seedance delivered a cinematic close-up, with a blurred background and golden backlight, beautiful and nothing like phone footage.
- Skin and freckles. Both made subtle freckles and textured skin. Seedance even shows the fuzz on the chin against the light; Kling has slightly smoother skin on the cheeks.
Both videos are above: watch to the end and notice whether the beauty mark changes place when she laughs and touches her hair, which is where models tend to slip.
Test summary: on fidelity to detail, Seedance 2.5 was better. The problem is that, with a real photo, it may not even generate, for the reason in the previous section.
How much it costs to animate 5 seconds on each
About 105 credits on Kling 4.0 and 222 on Seedance 2.5, at 720p. The table comes from the Cut.Pro Generator on 11/11/2026, for a photo as the starting frame, with the amount in Brazilian reais (R$) on the cheapest credit pack:
| Model | 720p per second | 1080p per second | 5 s clip at 720p | Duration |
|---|---|---|---|---|
| Kling 4.0 | 21 credits (R$ 1.25) | 31 credits (R$ 1.86) | about R$ 6.20 | 3 to 30 s |
| Seedance 2.5 | 44 credits (R$ 2.66) | 83 credits (R$ 4.99) | about R$ 13.30 | 4 to 30 s |
At 1080p the gap is more than two and a half times. For anyone who tests five versions of each idea before picking one, that alone decides it. The price follows the model's cost, so the current table is at cut.pro/en/models.
Which one to choose
| Situation | Best choice | Why |
|---|---|---|
| Animating your photo or a real person's | Kling 4.0 | Seedance 2.5 restricts real faces by policy |
| Illustrated character, animation, cartoon | Seedance 2.5 | No filter problem, and up to 50 references |
| Generating lots of test variations | Kling 4.0 | Costs less than half per second at 720p |
| Long scene with several characters and references | Seedance 2.5 | Up to 50 references versus Kling's 15 |
| Recurring persona with the same face across many videos | AI persona, then either one | The face comes from the persona, not from the model's luck |
The last row is the one I see go wrong most. Animating a photo solves one video. Anyone who wants the same person in thirty videos needs a fixed reference, and the path I recommend is in the post about which photos to upload so your persona looks like you.
Common mistakes when animating a face photo
- Small or compressed photo. The model doesn't invent the beauty mark the photo doesn't show. A recompressed story screenshot becomes smooth skin in the video.
- Asking for too much movement. Turning from profile to front forces the model to imagine the side of the face that isn't in the photo. The more it imagines, the more the detail changes. To preserve the face, ask for small movement: breathing, smiling, blinking, the camera moving closer.
- Forgetting the hand. A hand near the face is still where models fail most. If the scene calls for the person to fix their hair, check finger by finger.
- Using someone else's photo without permission. Beyond the filter, there's the question of image rights. Animating a third party's photo for a commercial post without consent is your problem, not the model's.
Where to start
Both models are in the Cut.Pro Generator, with the option to upload a photo as the starting frame. I'd start with Kling 4.0 at 720p and 5 seconds, with small movement, and only pay for higher resolution on what already turned out well. If the photo is of an illustrated character or the scene calls for lots of references, that's when it's worth testing Seedance 2.5.
The finished video goes straight to the editor, where you can add captions and schedule it to social networks without downloading a file. That's the argument of the post on model convergence: three months from now there will be a Kling 4.5 or a Seedance 3, and what stays is the path from the photo to the post. Just don't forget that realistic AI-made video of a person requires the AI-generated content label on TikTok, YouTube and Instagram, and that genuinely recorded footage is gaining weight, as Celso showed in provenance as a signal.
