In this video, the creator conducts an experiment by giving eight different AI video models the exact same prompt and settings to generate video clips. The prompt describes a tall lighthouse keeper with a silver beard walking along a rain-soaked coastal cliff path at dusk, carrying a glowing brass lantern shaped like a small whale. The camera is supposed to track smoothly beside him, arc around to his front as he stops, raises the lantern, looks out at a stormy sea, and quietly says, “Keep the light on. Captain Maruin.” The goal is to test the models’ ability to follow the prompt accurately, including motion, object consistency, and audio transcription of the made-up name.
All eight models ran sequentially on the Hicksfield platform with the same settings: text-only input, 16x9 aspect ratio, 10-second duration, and audio enabled. The models tested were Cedense 2.5, Cedense 2.0, MiniAX H3 1 3.0, Broke Imagine 1.5, Flux 3, Gemini Omni Flash 1.1, and Cling 3.0. The creator noted some differences in resolution and aspect ratio depending on the model’s capabilities but kept the comparison as fair as possible. Some models required retries to meet the 10-second duration requirement.
The results varied significantly across the models. Cedense 2.5 produced a stable face but lacked the walking motion and the whale-shaped lantern, instead showing a plain cage lantern. Cedense 2.0 performed best in following the prompt with the walking, whale lantern, and camera arc, though the face aged inconsistently, and the name was transcribed as “Captain Marvel Fin.” MiniAX H3 delivered the best whale lantern with a glowing belly but skipped the walking motion and altered the name to “Captain Marathon.” WAN 3.0 offered the most cinematic frame with rain and waves but used a plain oil lamp instead of a whale lantern.
Other models showed mixed results: Broke Imagine 1.5 had a whale-like brass lantern but at a lower resolution, Flux 3 included a small whale on the lantern but had some name transcription errors, and Gemini Omni Flash 1.1 featured a fisherman with no whale lantern and a name variation similar to MiniAX. The creator highlighted that none of the models got the made-up name exactly right, and each had strengths and weaknesses in prompt accuracy, motion consistency, audio, and how natural the video looked.
In conclusion, the creator ranked the models by category for this specific prompt: Cedense 2.0 was closest to the prompt overall, MiniAX H3 had the best whale lantern, Cedense 2.5 had the steadiest face, and WAN 3.0 produced the most cinematic visuals. The video encourages viewers to try the prompt themselves on Hicksfield and share their preferred model. The creator plans to test a second prompt involving a product shot using the same lineup of AI video models in a future video.
Useful Links
