daVinci-MagiHuman drops open-source human video model

This is the kind of release that makes human-video generation feel less like a demo and more like a system other teams can actually build on. The architecture is the real story: simplifying multimodal fusion may matter more than any single benchmark number.

• –Single-stream self-attention removes cross-attention plumbing, which should make training and debugging simpler.
• –The speed claims are real but hardware-bound: 5 seconds of 256p in 2 seconds is strong, but 1080p still needs a much heavier second stage.
• –Pairwise wins over Ovi 1.1 and LTX 2.3 suggest competitive quality, though the evaluation is still early and likely curated.
• –Releasing the base, distilled, SR, and inference stack makes this much more useful to researchers than a paper-only announcement.