We put Seedance 2.5, Kling 3.0 Turbo, MiniMax H3 and Wan 3.0 against each other across three image to video scenes, all run on EmpirioLabs. Every model animated the exact same starting image, one generation each, so any difference you see is the model, not a different starting picture.
Watch the four way comparison
Specs at a glance
| Seedance 2.5 | Kling 3.0 Turbo | MiniMax H3 | Wan 3.0 | |
|---|---|---|---|---|
| Inputs | Text, images, video and audio | Text and images | Text, images, video and audio | Text, images, video and audio |
| Highest resolution | 1080p | 1080p | 2K | 1080p |
| Clip length | 4 to 30 seconds | 3 to 15 seconds | 4 to 15 seconds | 2 to 30 seconds |
| Vertical 9:16 output | Yes | Yes | Yes | Yes |
| Price as tested | $0.462 per second at 720p | $0.18 per second at 720p | $0.18 per second at 768p | $0.14 per second at 720p (standard tier) |
How we ran it
For each scene we first generated one starting image with Seedream 5.0 Pro at 1584 by 2816, a vertical 9:16 frame. Then every model animated that same image, one generation each, 5 seconds per clip at 720p (MiniMax H3 at 768p, its lowest setting), in 9:16, with otherwise default settings. The three scenes were a snow leopard leaping across a mountain ravine, a rocket lifting off from a seaside pad at dusk, and a giant wave curling over at sunrise. Each model got the same motion prompt per scene, asking for a single continuous shot with no cuts.
What to look for
Because every model starts from the same frame, this comparison is about motion. Watch how the snow leopard lands and keeps its shape, how the flame and exhaust move as the rocket climbs, and how the water breaks as the wave curls over. Look at how closely each clip follows the motion prompt and how stable the subject stays from frame to frame. We are not declaring a winner. Run the clip, see which look fits the kind of video you make, and pick from there.
Generate any of these on EmpirioLabs
All four models serve the same image to video endpoint: send a starting image and a motion prompt, then swap the model id to compare.
curl https://api.empiriolabs.ai/v1/videos/generations \
-H "Authorization: Bearer $EMPIRIOLABS_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "seedance-2-5",
"image": "https://example.com/your-vertical-frame.png",
"prompt": "The wave pitches forward and curls over, single continuous shot, no cuts.",
"resolution": "720p",
"duration": 5
}'Change model to kling-3-0-turbo, minimax-h3 or wan-3-0 and send the same image and prompt again. For MiniMax H3, set resolution to 768p or 2k. You can also try them in the playground.
Frequently asked questions
Which video models were tested?
Seedance 2.5 from ByteDance, Kling 3.0 Turbo from Kling, MiniMax H3 from MiniMax and Wan 3.0 from Alibaba. All four run on EmpirioLabs.
Did every model animate the same image?
Yes. We generated one starting frame per scene with Seedream 5.0 Pro, then gave that same frame to all four models, so the comparison isolates the motion from the image.
What settings were used?
5 seconds per clip at 720p (MiniMax H3 at 768p), in 9:16, otherwise default settings, one generation each, no edits or retries. Each model got the same motion prompt per scene.
Can I generate vertical video?
Yes. All four models produce 9:16 video. Set aspect_ratio to 9:16 on Kling 3.0 Turbo, MiniMax H3 and Wan 3.0. With a starting image, Seedance 2.5 takes adaptive (its default) and follows the image's shape.
Can I use my own starting image?
Yes. The image to video endpoint takes any starting image plus a motion prompt. Swap the model id to compare the same input across models.
