Posted
We spent the last months researching several new ways to improve Factories with new AI/ML capabilities. Generate progressions and fragments that fit a certain mood, or that continue a development you already have in your arrangement. Maybe you sketch something on a MIDI keyboard and a Factory suggests phrases that match. Things like that.
This became more attractive now that so many building blocks for small locally run AI models are available. Unfortunately none of the current music models do the job. We are experimenting with synthetic/derived/meta data instead, because Factories merely need to generate a few meaningful parameters and abstract figures - not finished music. Another approach is to use a GAN for a model to learn how to use a Factory. Much harder.
Our stance is still that we will never generate songs or arrangements that imitate or reshuffle other people's music. We will offer new ways to come up with inspiring fragments that you can build on. If you don't build the music yourself, it's not your music.
I'll post a few results here when they eventually get better than "totally embarrassing" ;-)
Do., 27.08.2026 - 18:59 Permalink
Have you seen the new Ableton Extensions that have been coming out? I was looking at some of the things they can do, and it got me thinking that something along those lines would be really cool in Synfire Pro.
The ability to extend Synfire with more generative/compositional tools—especially something that could generate and manipulate musical ideas, rhythms, harmonies, and MIDI—could be incredibly powerful.
I think Synfire already has such a strong foundation for this kind of thing, so having an extension system that lets users build or add specialized generative tools could take it to another level.
Just curious if you've seen what Ableton is doing with Extensions and whether you've ever considered something similar for Synfire Pro.
Fr., 28.08.2026 - 12:26 Permalink
Yes, Ableton Extensions are great. They could be used to feed Synfire output into a Live project, for example. Although the export of snippet grids is already very powerful (you get an entire set of matching Live clips in no time).
Could something like JavaScript/TypeScript inside Synfire be useful? The parameter that makes most sense for generative synthesis is Figure. That's deeply structured non-linear data - segments, symbols, anchors, types - where segments are overlapping each other. Building figures with JavaScript is not trivial. The best home for scripting is probably a Factory.
So., 06.09.2026 - 12:19 Permalink
How much VRAM is everyone willing to spend on local AI?
I'm currently running a Qwen3.6 35B-A3B model on a Mac Studio and it seems clever enough to interact with Synfire in meaningful ways. Claude or ChatGPT are not really needed. It uses ~28GB of VRAM. Windows PCs don't typically have unified memory, so you'd need a dedicated GPU with 32GB VRAM (which is ridiculously expensive at the moment). I think nobody would spend $6k to chat with Synfire.
The "Factory" models we are building are much smaller though. They might fit into a single audio plug-in. We don't want to host inferencing for our users because that would require a subscription to cover the hosting. The goal is to run everything locally.
Mo., 07.09.2026 - 05:41 Permalink
Sounds good, just release it. Nvidia has CUDA, which lets everyone use their CPU as a GPU.
Mo., 07.09.2026 - 09:11 Permalink
CUDA doesn't change the fact that models need to fit entirely into VRAM (GPU). An open source LLM that can do MCP (use tools) is huge. NVIDIA consumer GPUs have only 8GB. That's why the DGX Spark with a 128GB GPU exists in the first place.
We're trying to squeeze our own custom models into less than 1G parameters. You can't chat with those. They are hard-wired to serve dedicated Factories for parameter/figure generation.