Realism vs control: the most realistic AI video model isn't always the right pick

The most realistic AI video model is not always the right pick, because realism and control are two different skills. Realism is whether a clip passes for real phone footage: a face that moves like a person talking, coffee that pours like coffee, objects that keep their shape. Control is whether the clip does what you asked: the actions in your prompt, in your order, in the room you approved in the still.

This post shows that trade-off on real takes instead of a ranking. On November 14, 2026, we ran three shots through four video models in Hook Studio: Kling 3, Kling 4, Seedance 2, and Seedance 2.5. Each model got the same approved still and the same prompt, for one 5 second take. Every clip below is the take as it came out, with what Hook Studio's realism check said about it. Camila, Marco, and Nia are AI people from Hook Studio's premade cast.

• Realism: the clip passes for real footage. A face moves like a person talking, not acting. Hands, liquids, and objects behave. Nothing appears, vanishes, or changes shape. The camera moves like a hand that holds a phone.
• Control: the clip does what you asked. The actions in your prompt happen, in your order. The words land in sync with the mouth. The framing and the camera stay as you wrote them. The props stay where the still put them.

The two overlap. A face that mugs through a line the prompt called calm fails both: it ignores the direction, and it reads as AI. A shelf that grows in the middle of a clip fails both too. But one model can be strong at one test and weak at the other, and our takes show it.

What a leaderboard measures

Public leaderboards rank models by blind votes: people see two clips and pick the one they like more (Artificial Analysis). On November 14, 2026, Seedance 2.5 was first on Artificial Analysis's image-to-video board and Kling 3.0 was 23rd. In our talking test the order flipped: Kling 3 made the only takes that passed. A ranking tells you which clips voters liked on average. It does not tell you which model does your shot.

The makers pitch the split too. Kling lists stable characters and physics-based motion among Kling 3.0's strengths (Kling). ByteDance sells Seedance 2.0 on "director-level control" of camera, performance, and lighting (ByteDance Seed). For who makes each model, see who makes Kling, Seedance, and Veo.

• Three shots, each from one approved still: Camila says one line at a café table, in English and in Spanish. Marco pours coffee into a mug. Nia does three actions in a row in her living room.
• Same input for every model: the same still and the same prompt, 5 seconds, one take each, no retakes.
• Resolution: Kling 3 renders only at 1080p. Kling 4, Seedance 2, and Seedance 2.5 ran at 720p.
• Seedance 2 sat out the talking line. It ran on the two action shots.
• The verdicts: each clip went through Hook Studio's realism check, which flags what looks fake. It does not score whether a clip followed the prompt, so the direction verdicts below are ours, read frame by frame.

Camila sits at a café table in Mexico City and says one line to the friend who films her. We ran the line on Kling 3, Kling 4, and Seedance 2.5, in English and in Spanish, with the speech in the clip's own voice. This is the English prompt, word for word:

Seated at the café table, she is already talking to her friend across the table, calm and friendly, with a small natural nod; her hand stays near the cup. She says: "Same photo, same line, different AI model. Which one looks real?"

Camera: her friend, sitting across the table, eye level, handheld phone, slight natural hand sway. 5 seconds, with speech.

Kling 3 was the only model whose talking takes passed, in both languages. Kling 4 kept every sentence in sync and still failed on the face. Seedance 2.5 failed on both.

One round is a sample, not a rule. On November 6, in the session for our guide on making AI videos, four Kling 3 talking takes in a row failed the same check for over-acting. Judge each take, not the model's name.

Three actions, in order

Nia holds a mug in her living room. The prompt asks for three actions in a set order: one sip, the mug onto the wall shelf, then a two-hand wave.

Three actions, in this order, starting at once: first she takes one sip from the white mug; then she steps toward the narrow wooden wall shelf at waist height and sets the mug down on it; then she turns back to the camera and waves with both hands, smiling. No talking.

Camera: her friend, standing about two meters in front of her in the living room, eye level, handheld phone. 5 seconds, no sound.

All four models did the three actions in the right order. The difference was what each one broke on the way.

The shelf tripped every model. In the still it shows one strap, and all four rebuilt it as a longer shelf with two straps once Nia stood at it. The check failed only Kling 3 for it, the take where the mug also disappeared.

The coffee pour

Marco pours coffee from a glass carafe into a mug until it is about half full, then sets the carafe down. Calm, no talking, filmed by his roommate across the counter.

Seedance 2 shows the trade-off inside one clip: the most believable coffee of the four, and a camera that went its own way. Kling 3 shows it across the whole test: the faces you would post, and the scenes you would not.

We call it the two-pass check. Run it on every take before you keep it.

1. 1Pass 1, realism, with the sound off. Watch the take at phone size. Look at the face first: would a person make that face at that moment? Then the hands, any liquid, and anything that grows, appears, or vanishes. Then the camera: does it move like a hand, or drift on its own?
2. 2Pass 2, control, with the prompt next to you. Tick each action in your prompt, in order. Turn the sound on and check that the mouth matches the words. Then put the still next to the last frame and compare the props one by one.
3. 3Check the last second. Three of the problems in our test landed there: the Spanish lip-sync slips on Kling 3 and Seedance 2.5, from 4.0 to 4.9 s, and the mug that vanished from Kling 3's shelf.
4. 4Don't stop at the automatic check. Hook Studio's realism check passed the Kling 4 take in which the painting changed. Two frames side by side caught it.

• Sort the shots. Mark each shot of your reel as talking (a face says a line) or doing (hands, a product, an action).
• Talking shots go on the model whose faces pass. In our test that was Kling 3, at 143 credits for 5 seconds with speech.
• Doing shots go on the model that does the steps and keeps the scene. In our test that was Seedance 2.5, at 197 credits. Kling 4 passed the check on both action shots for 63 credits less, with a painting that changed.
• Mix models in one reel. Each shot is its own clip from its own still, so the hook can run on one model and the product shot on another. In Hook Studio you pick the video model under the message box, and you can change it before any clip.
• Price doesn't predict the result. The talking takes that passed cost 143 credits on Kling 3. The ones that failed cost 134 on Kling 4 and 197 on Seedance 2.5.

For the full method, see how to pick an AI video model. For every model side by side, see which AI video model to use and Kling vs Seedance.

What this test can't tell you

• One take per model per shot. Takes vary. The same Kling 3 that passed both talking takes here failed four in a row on November 6.
• Four models. Gemini Omni, Wan, MiniMax, and the other makers' models were not in this round.
• Two resolutions. Kling 3 ran at 1080p, the only resolution it renders. The other three ran at 720p.
• Our read of direction. The direction verdicts come from frames at 0.8, 2.2, 3.5, and 4.8 seconds, not from an automatic score.