Local generative media grows up: open video and music you can actually run
Originally announced by Hugging Face . We link the primary source so you can read it for yourself.
Local AI has always been broader than language models, but for a long time the generative-media corner of it, image, video, and audio, sat awkwardly on the “local” label. Image generation was the exception, thanks to Stable Diffusion and a deep community around it. Video and music were something you admired in a demo, then ran in the cloud. That is now shifting, and it is worth taking stock, because the models are real, and the catches are mostly about licensing rather than hardware.
Video reaches consumer cards
The headline change is video. Open text-to-video has been gated by memory: the good models wanted 80GB accelerators. The arrival of models like Wan 2.2 TI2V-5B, which fits a single 24GB card and is Apache 2.0, changes what “local video” means. It generates short 720p clips, from a prompt or an input image, on hardware a lot of people already own.
It is not the whole story. LTX-Video trades some quality for speed and a small memory footprint, while HunyuanVideo produces stronger results but still wants 45GB or more, keeping it a workstation or rented-server model. Clips remain a few seconds long. But the floor has dropped, and that is what matters. We have written up the details in a new guide, local video generation.
Music, where the licence is the story
Local music generation has a different shape. The technical bar is low, some models run in 8GB, but the licensing is where people get caught. The best-known open music model, Meta’s MusicGen, releases its weights under a non-commercial licence, so it cannot be used in commercial work at all. That single fact rules it out for a lot of would-be users.
The alternative is worth knowing about. ACE-Step is Apache 2.0, generates full songs with duration control, and runs on a consumer card, which makes it the practical choice where commercial use matters. Stable Audio Open sits in between, free below a revenue threshold and strongest at sound effects rather than songs. Our new guide, local music and audio generation, walks through the trade-offs.
The common thread
Most of these models are diffusion models, the same family as local image generation, which is why they run through the same tools, chiefly ComfyUI, and why their hardware needs scale with resolution, length, and step count rather than context. The through-line for anyone exploring this space is that the hardware question is increasingly answerable on consumer gear, while the licence question is the one that still needs checking every time.