AI music video generator test: MiniMax H3 vs Wan 3.0 vs Kling 4.0 on a two-singer rap

Most AI music video generators on Google’s first page take a song and give you visuals that move to it. We wanted something narrower: two people, a man and a woman, each rapping their own lines of a song we already had, with the mouths on the words. So on 2026-11-11 we ran one 15-second rap duet and two portraits through four video models. First try, one clip with the whole duet as reference audio, failed in both models that accept a song. MiniMax H3 and Wan 3.0 both kept our audio exactly, but they did not know who sings which line, so the woman’s mouth moved on his words and the other way round. What fixed it was splitting the song by voice. We found each singer’s lines from the pitch of the vocal (his around 130–200 Hz, hers around 230–380 Hz), cut the audio there, and made one 8-second clip per singer. The prompt for each clip says only that person raps and the other keeps their mouth closed. Then we joined the clips and laid the full song over them. With that, MiniMax H3 had the best lip sync of the four. Wan 3.0 was close but looser. Through the APIs, each 8-second clip cost between $0.19 and $0.32, depending on the model and resolution. The surprise was the light. Both got the same prompt, which said warm amber walls. H3 made a dim, moody studio. Wan made it bright and sunny. Neither prompt asked for either. If you want H3’s mouths with Wan’s light, write the light into the prompt. Two models can’t take your song at all. H3 Max Turbo has no audio input. Kling 4.0’s Omni reference accepts audio, but only as a voice reference: it borrows the voice and sings its own melody and timing. Both rapped our lyrics in their own way, so once our track went on top, the mouths were off. Their upside: starting both clips from the same first frame kept the room the same across the cut. H3 and Wan rebuilt the room for each clip. We read that as a camera change, and it works fine in a music video.