Salsedine-DJ
I was wandering if would be possible to integrate in the application some agentic AI workflow that interacts with VDJ environment (e.g: accessing the History records, propose suggestion based on historical patterns identified, or new entires in the library or longly not played songs, etc.). The feature could be delivered with the contribution of locally deployed small LLM models to manage the orchestration and execution
Posted yesterday @ 4:41 pm
DJ VinylTouch
There is some AI functionality already involved in the software wrt track discovery ideas:
https://virtualdj.com/manuals/virtualdj/interface/database/aiprompt.html
There is (AFAIK) no local LLM or other type of bring your own Model (BYOM) type integration though (however it's already been suggested) and the feature could probably be expanded to consider more things.
https://virtualdj.com/manuals/virtualdj/interface/database/aiprompt.html
There is (AFAIK) no local LLM or other type of bring your own Model (BYOM) type integration though (however it's already been suggested) and the feature could probably be expanded to consider more things.
Posted 22 hours ago
locoDog
history suggestions, linked tracks, genius dj are already a thing.
Anything extra is just a hat on a hat.
Anything extra is just a hat on a hat.
Posted 22 hours ago
djcel
Local LLM needs resources more than stems 2.0 computing...
If you have 4GB of VRAM you are very limited in models. It will use the RAM (at least you need 16 GB) too and prediction could be slow.
Not sure you want to give more resources for your local LLM than for your VirtualDJ (especially if you use videos)
You can create a LLM local server for example with LM studio (or ollama) and llama 3.2:1B (or gpt-oss-20B if you have more VRAM+RAM) but I do advise you on another computer.
Then you can call the url of the LLM server.
Still easier and faster to use cloud server and pay tokens or a subscription to my mind.
Other small models: Phi-4-mini (3.8B), Gemma 2 2B, and Qwen 2.5 3B
If you have 4GB of VRAM you are very limited in models. It will use the RAM (at least you need 16 GB) too and prediction could be slow.
Not sure you want to give more resources for your local LLM than for your VirtualDJ (especially if you use videos)
You can create a LLM local server for example with LM studio (or ollama) and llama 3.2:1B (or gpt-oss-20B if you have more VRAM+RAM) but I do advise you on another computer.
Then you can call the url of the LLM server.
Still easier and faster to use cloud server and pay tokens or a subscription to my mind.
Other small models: Phi-4-mini (3.8B), Gemma 2 2B, and Qwen 2.5 3B
Posted 13 hours ago
DJ VinylTouch
I suspect even running local models would be out of the scope for most persons, because as you pointed out, any decent, less quantized, general-purpose model requires a fair bit of resources, and works against the realtime playback requirements (making dedicated machine use obvious), and then (unless you are only using this in a home network circumstance), all internet exposure problems apply (exposing safely behind a secure connection, maybe only accessible via VPN, etc), and that is probably beyond most most people's price range and patience (ppl here already complain about writing vdjscript and the price of the software).
LLM access as a service providers like OpenRouter would probably be the first choice for most users.
LLM access as a service providers like OpenRouter would probably be the first choice for most users.
Posted 6 hours ago