I’ve officially released the first finetuned Stable Audio Open sample generator on HF. Bringing with this release I’ve also updated the Gradio to take full advantage of the model. So lets dive in, see what I’ve done and figure out where we may be headed. [Long Thread🧵]

ICML in Vienna is coming to a close! 🇦🇹 Here are the top-10 general (and audio) trends from ICML 2024. A thread 🧵 1. Open vs. Closed AI: The debate was very present, notable in Soumith Chintala's keynote or by the release of Llama 3.1 (among others). icml.cc/virtual/2024/p…

🎵 Stable Audio Open 🎵 just landed into diffusers, be ready to get: -> a whole lot of fun 🎹🎺🎷🥁🎼 🎸 -> easy installation: `pip install diffusers` -> easy usage: 5 lines

Flipping a Stable Audio AI Sample into glitchy liquid texture.

This paper title really undersells the fact that this only works because the CLAP model *didn’t* use audio-only data.

call me Kermit the way I'm bout to muP it

decrescendos >>>>>>>

AI training can and should be beautiful

Getting the Skrillex meal at McDonald’s (a nice Sprite)

I've been trying this out today, it's incredible. This is the Flash Attention moment for contrastive model training. Super easy to use the code, it handles all the distributed stuff for you. This should unlock a lot of new research! Great work!

Prompt Jockey (n) - it's DJing but harder. With DJing you're playing usually other peoples' tracks. But with this, the tracks don't even exist yet. You're prompting them live with a neural network. lyra bubbles~ ♪❀ zach 🏔🎶 encanti Mr. Bill

📢ATTN Producers & Musicians📢 Today is the release of the most capable FREE open source sample generator tailored for EDM production. Its the largest model yet with HIGH musicality & VERY robust AI Style transfer. ft. a link to get started with it RIGHT NOW😎 Lets dive in! 👇

Do you want to do cutting edge research on generative music production tools? Do you want to publish papers and also release open weights and code? Apply to be a research intern on our team!

What if you could turn everyday sounds into songs? Our Audio Researcher CJ Carr shows you how with Stable Audio. You can now access Stable Audio via the Stability AI API — or, as always, at StableAudio.com. Learn more: bit.ly/4iI0V4q

Super excited to launch Stable Audio Open Small today! It was great working with Arm on this model to make sure it runs efficiently on Arm CPUs. I'm also now an Arm Ambassador! I'm looking forward to helping the Arm developer community integrate this new model.

Releasing Stable Audio Open Small! 75ms GPU latency! 7s *mobile* CPU latency! How? w/Adversarial Relativistic Contrastive (ARC) Post-Training! 📘:arxiv.org/abs/2505.08175 🥁:arc-text2audio.github.io/web/ 🤗:huggingface.co/stabilityai/st… Here’s how we made the fastest TTA out there🧵

got stable audio open small training in <12gb VRAM at batch size 8 & default sample size everyone with 16-24gb cards who wanted to locally tune SAO 1.0 but couldn't (27.6gb vram) should be very happy now

sample gen models =/= song gen models you can't compare the two, they do very different things

Landed a feature in stable audio tools that should make it easier to fine-tune your own custom text to music models - especially if you're GPU poor. Pre-encoding latents ahead of time reduces GPU memory + helps keep your GPU hot🔥 Documentation here: github.com/Stability-AI/s…

“Y’all can have your vibe coding, I’m doing vibe science” - zach 🏔🎶