1. What we tried
In 2025-2026, AI video generation saw explosive growth, with new models appearing all the time. Rather than judging from specifications and marketing claims, we decided to try them ourselves — we took the same task and the same set of prompts and ran four scenarios across six platforms.
Platforms tested:
Web interface: Kling 3.0, HappyHorse 1.0, Veo 3.1 Fast
API access: Seedance 2.0, Dreamina Video 3.0, Xiaoyunque
Supporting tool: Dreamina 5.0 Lite (to generate the character reference image).
Note: in third-party benchmarks, the models covered here rank Seedance 2.0 > HappyHorse 1.0 > Kling 3.0 > Veo 3.1 Fast. The ranking reflects the overall benchmark score; real-world experience varies by scenario.
━━━━━━━━━━━━━━━━
2. Test design
All prompts were sent in English; the Chinese descriptions below are only there to make them easier to follow.
◆ Test A: Static scene — wasteland at dusk
A small settlement in the ruins of a nuclear war, at dusk. The outlines of broken towers loom faintly in the pink-purple glow of radioactive dust, and a rusted water-purification tower throws a harsh reflection in the setting sun. The ground is covered in grey-white radioactive dust, and a few figures wrapped in old cloth move slowly among the ruins. Cinematic widescreen composition, post-apocalyptic aesthetic, hyperrealistic style.
What we looked at: atmosphere, composition, and accuracy of detail (whether elements such as the radioactive dust, the rust, and the water-purification tower are reproduced faithfully).
Results:
Kling 3.0 — strong sense of immersion
Veo 3.1 Fast — good texture
Seedance 2.0 — built-in camera movement
Dreamina Video 3.0 — adequate
HappyHorse 1.0 — adequate
Xiaoyunque — a consistently stronger sense of story (thanks to its 15-second generation length)
◆ Test B: Dynamic action — wiping a glass in a tavern
Inside the dim bar of a wasteland tavern, a man of about 35, “Galen Ash”, looks down as he wipes a metal goblet with an old cloth. He wears a worn dark-brown leather jacket; his face is weathered and his eyes are grey-green. He lifts the goblet to an oil lamp and turns it to inspect it, sets it down, then raises his eyes to the camera, the corner of his mouth curling slightly, his gaze wary and calculating. Close-up shot, slow push-in, cinematic lighting.
What we looked at: character stability (whether the face deforms), continuity of motion, gaze and expression, and the quality of the lighting.
Results:
Kling 3.0 — the micro-expressions fall short
Veo 3.1 Fast — high-contrast texture, but the micro-expressions ignore instructions completely
Seedance 2.0 — the expression passes, but the clip is too short for the performance to finish
Dreamina Video 3.0 (mislabeled as seedance3.0 in the video) — the motion is a little stiff
HappyHorse 1.0 — the most natural micro-expressions; adequate overall
Xiaoyunque — a strong sense of story
◆ Test C: Image-to-video — character consistency
Starting from the uploaded reference image of Galen Ash, generate the character turning his head to look off to the right, then slowly turning back to face the camera, while his left hand brings an old pistol out from under the bar and sets it on the tabletop. Keep the character's face, clothing, and hairstyle exactly the same as in the reference image.
What we looked at: how closely the character's face, clothing, and hairstyle match the original image, the turning motion, and the hand movement.
Reference image:
Results:
Kling 3.0 — character consistency OK and motion natural, but it never ends up facing the camera
Veo 3.1 Fast — not completed (free quota exhausted)
Seedance 2.0 — not completed (reference image rejected)
Dreamina Video 3.0 — did not take part
HappyHorse 1.0 — character consistency OK with built-in camera movement; adequate
Xiaoyunque — did not take part
◆ Test D: Complex narrative — escape and betrayal
A street in the wasteland ruins: Galen bursts out of a collapsed building and runs fast across the rubble. He glances back over his shoulder (the camera follows), his expression changes, and he speeds up. He leaps over a fallen concrete pillar, rolls once as he lands, picks up a rusted metal box from the ground, then ducks into the shadow of the building beside him, back against the wall, breathing hard, his eyes full of fear and anger. Long take, handheld camera style, several actions in continuous motion.
What we looked at: execution of several actions in sequence, expressiveness of camera movement, physical collisions (running / jumping / rolling), and emotional expression (fear and anger).
Results:
Kling 3.0 — follows instructions; adequate
Veo 3.1 Fast — not completed (free quota exhausted)
Seedance 2.0 — smooth overall
Dreamina Video 3.0 (mislabeled as seedance3.0 in the video) — the character's running looks unnatural, with obvious AI artifacts
HappyHorse 1.0 — the character does not roll
Xiaoyunque — adequate
Overall observation: none of the videos delivers the camera follow of the “glance back over the shoulder”, which is a real shame.
━━━━━━━━━━━━━━━━
3. The ranking from the tests
For all six models, a single generation took between 3 and 5 minutes. Based on their actual performance across the four test scenarios, the ranking is as follows:
No. 1 【Seedance 2.0】(ByteDance, Volcano Engine API)
API barrier to entry: top up from 200 yuan, or buy a token pack for about 190 yuan. It is the underlying model of Xiaoyunque.
Performance: built-in camera movement in A, expressions that pass in B, smooth overall in D. In C the reference image was rejected for being “too much like a real person”, which is not a matter of capability.
No. 2 【HappyHorse 1.0】(Alibaba, web interface)
No free quota; about 4.5 yuan for 5 seconds at 720p. It completed all four tests.
Performance: the best micro-expressions in the whole field in B; in C, consistency OK with built-in camera movement. In D the character does not roll, which costs it points.
No. 3 【Kling 3.0】(Kuaishou, web interface)
A free quota is available, and the videos come with their own sound effects. It completed all four tests.
Performance: strong immersion in A; in C, consistency OK with natural motion; in D, follows instructions. Micro-expressions fall short in B.
No. 4 【Xiaoyunque】(ByteDance / CapCut, Volcano Engine API)
A free quota is available. A content-creation agent, built on Seedance 2.0. Minimum clip length 15 seconds (other platforms start at 5), at 2.13 yuan per second, so trial and error is expensive.
Performance: a strong sense of drama in A and a high sense of story in B; adequate in D. The 15-second minimum gives it a generally stronger sense of story.
No. 5 【Veo 3.1 Fast】(Google DeepMind, web interface)
A free quota is available; the subscription is $99 per month. It completed only A and B (C and D went untested once the quota ran out).
Performance: good texture in A; in B, the micro-expressions ignore instructions completely.
No. 6 【Dreamina Video 3.0】(ByteDance, Volcano Engine API)
A free quota is available.
Performance: adequate in A, slightly stiff motion in B, and in D the character's running shows obvious AI artifacts — the weakest of the whole field.
These results are broadly consistent with our earlier ranking of large AI video generation models:
External video-model reference ranking (Artificial Analysis, June 2026; text-to-video with audio)
- 1. Seedance 2.0 — 1215; 2. HappyHorse-1.0 — 1121; 3. SkyReels V4 — 1106.
- 4. Kling 3.0 1080p Pro — 1104; 5. Kling 3.0 Omni 1080p — 1099; 6. Kling 3.0 Omni 720p — 1094; 7. Kling 3.0 720p — 1094.
- 8. Veo 3.1 — 1093; 9. Sora 2 (closed 26 April 2026) — 1089; 10. Veo 3.1 Fast — 1086.
- This is an external blind-test ranking, separate from the article’s own six-model test. Source: artificialanalysis.ai/video/leaderboard/text-to-video.
━━━━━━━━━━━━━━━━
4. Tips
1. Send your prompts in English
All the models understand and follow English more reliably, so write your prompts in English; a Chinese scene description can be added as a supplement inside the English prompt.
2. Don't queue online for Seedance 2.0
The queue on the web version is a poor experience — after a full day in the line, the waiting time had barely moved. Use the Volcano Engine API instead.
3. Calling the API takes code
Using the API means writing the integration code yourself, so it helps a lot to keep an AI coding assistant on hand. For the details, see each platform's official documentation.
━━━━━━━━━━━━━━━━
A record of our tests in June 2026. The platforms iterate so often that the next attempt could be completely different.
All of the above reflects a personal testing perspective and is for reference only.