Explosion Test: The Top 8 "Image-to-Video" AI Models. Same Prompt, Same Image.

Which video model do you think performed the task most accurately?

I tested 8 of the top Image-to-Video models using the same image and prompt—trying to "blow up" the ground floor of this building for a WWII film shot I’m working on.

The still image was generated using my custom WWII Flux checkpoint (you can grab it here: https://civitai.com/user/MattHVisual).

The wild part? Wan 2.1 was done locally on my RTX 4090. Took around 15 minutes—but pretty incredible that this is possible without a commercial GPU, but you be the judge.

Prompt:
"Wide shot: A dimly lit street in a foggy, old town. The camera slowly pans across the cobblestone road, revealing vintage buildings with warm light spilling from the windows. Gentle rain falls, creating reflections on the wet pavement, enhancing the moody atmosphere. Suddenly, the ground floor of the building explodes, sending fire, smoke, and debris out through all the windows onto the street."

Models tested:
Kling 1.6 (Pro 5 + 10 sec)
Luma Dream Machine
Luma Ray 2
Minimax
Veo 2
Runway Gen-3 Alpha
Runway Gen-4
Wan 2.1 (720p)

Let me know which one captured the scene best—for realism, timing, or accuracy.

Contact Links #
Website: http://hallettvisual.com/
Website AI for Architecture: https://www.hallett-ai.com/
Instagram:   / hallettvisual  
Facebook:   / hallettvisual  
Linkedin:   / matthew-hallett-041a3881