New AI Makes Video With Matching Sound in Just 4 Steps

Most AI video tools make silent clips. Adding sound is usually a separate, slow job bolted on afterward, and if the two don't line up, the result feels fake. Salt++, from a team of researchers, generates the video and its audio together, and does it in only four steps — far fewer than the dozens of passes typical systems run. Fewer steps means less computing, which means lower cost and faster results.

The clever part is how they trained it. Teaching an AI to generate something is a bit like coaching someone who practices on clean, perfect examples but then faces a messy real world. The team deliberately fed their model messy, noisy history so it learned to work with real, imperfect situations instead of only ideal ones. They also made sure the part that creates the video and the part that judges its quality were looking at the same context — like making sure a chef and a food critic are tasting the same dish.

The results are notable. At 480p resolution (roughly standard-definition quality), Salt++ improved picture quality by 57% and motion smoothness by 45% compared with a competing system called OmniForcing, using the same four-step setup. The team then scaled it up to a much sharper 1664x960 widescreen format, where it beat a well-known rival called LTX-2 on six of seven measures. Those are research benchmark scores, not your everyday viewing experience.

So what's the catch? This is an academic paper, still under review, with no public code or product. The tests ran on one specific benchmark, so real-world performance is unproven. Still, the direction is clear: AI that produces video and synced sound quickly and cheaply would matter for creators, advertisers, game makers, and anyone dubbing content into other languages.