This is an AI video generation comparison for image-to-video 
prompt:
Zebra in a red bomber jacket stands in the middle of a gritty urban street, flipping off the camera with a bold striped finger. Gold chain gleaming, aviator sunglasses reflecting city lights, he stares dead into the lens, standing atop chalk one-line graffiti that reads "AICREATORS.TOOLS". Suddenly, the camera begins an intense super dolly zoom-out — rushing backward revealing the city with cars and walking pedestrians, rising above rooftops and power lines, pulling past highways and satellite views until the entire continent comes into frame. The Earth spins into view, suspended in the black void of space, clouds swirling over Africa, the zebra now just a forgotten dot on the planet's surface.
Tested: July 29, 2025
Supplied zebra's image (you can see it flashing briefly in the beginning) and using cut-to prompt made the Earth zoom-in video.
Tested: July 29, 2025
Hailuo 02 consistently and reliably does this generation well. External link containes the reversed version with music and extension with a shotgun (shotgun scene is Seedance 1 Pro).
Tested: July 29, 2025
This model was never able to smoothly transition imitating altitlude but it's tried, and some results are still quite fun.
Tested: July 29, 2025
This is not too great. Transition mode with 2 keyframes.
Tested: July 29, 2025
This is using Pixverse's template 'Earth Zoom Challenge'. I think it needs some more work, it's very simplistic.
Tested: July 29, 2025
This is the 4th variation of the prompt wording but Wan 2.2 keeps 'cheating' and literally turning the city into a bubble-wrapped mini-Earth! In other generations it'd do a "circular mask transition" or "iris wipe (reverse)" transition. Doesn't help that the max prompt length is smaller than for 2.1.
Tested: July 29, 2025
Been tweaking the prompt for a while, Pro model does great super dolly out but then struggles - as many others - at the point of small city to continent on Earth transition, often simply incorporating that area onto the Earth in space. Also likes to keep the main subject slightly larger than it should naturally be when zooming out.
Tested: July 29, 2025
Higgsfield offers a handy preset called "Earth Zoom Out" so we're using it. And it's working.
Tested: July 29, 2025
Oddly enough, does poorly for this task. If you find the 'key' to this model for this kind of generation let me know in the comments below!
This test checks if AI can handle a Super Dolly Out camera move. It starts way up high and pulls back from a close-up city scene to a full view of Earth.
Some AIs mess up this kind of zoom. They might cause a rough cut where the real street scene jumps straight to a generic Earth shot. The test is to see if the AI can pull off one smooth zoom... starting in the city and moving out to orbit without messing up the scene or adding glitches.
Such viral videos can be made in different ways: sometimes you can create a super pull out video (from subject to planet Earth in space) and then reverse it in a video editor.
Or, you can prompt tools that respond well to 'cut to' like Veo 3 and get the Earth - to subject straight up, but you'll likely have a first second still flashing your reference image.
Finally, you can try start and end frames for tools that have keyframes feature and upload an Earth in space and your subject images + a prompt like this one, but in this case it's more about what happens in the middle that's tricky, so start-end frames don't help as much.
Check out the results from Veo (Veo 3 Quality) vs Hailuo AI x Minimax (Hailuo 02) vs Freepik (Seedance 1.0 Mini) vs PixVerse (Pixverse V4.5) vs PixVerse vs Wan (Online Platform) (Wan 2.2) vs Freepik (Seedance 1.0 Pro) vs Higgsfield AI (Higgsfield-v1 Lite) vs Kling AI (Kling 2.1) for similar or identical prompts side-by-side.
