New here, I have questions: 1) what are the green and red things? 2) Do you have prompt and parap examples? 3) Did you figure out how to use SFT to get good resuls?

[StableInfo]

Hi.
So in this image:

1. What are the red aras? (I bet green is for "read")?
2. Do you example on how to reprorduce the music example samples you shared in the readme repo page?
3. So far when I tried SFT or base I did nnot get good enough instrumentals, it was alwasys too aggressive, turbo was a bit better, did you figure out how to use all these models
4. I usee you are using QWEN, is it much much better than the base LM model from acestep team?
5. Anything else I should know? (in term of features for example)

Thanks

[ServeurpersoCom]

What are the red aras? (I bet green is for "read")?

The red areas simply indicate a selected area for repainting function (use base model).
It's connected to the "Src" checkbox. The player loops back to it.

Do you example on how to reprorduce the music example samples you shared in the readme repo page?

I didn't keep the JSON files because the WebUI didn't even exist; I made the music using command lines during dev!
For the lyrics, I just used Claude. But I can give you plenty more, even better ones in this discussion :)
Try the "Dice" on last version. Workflow is :

Dice -> [then "Inspire" button for the one with only Caption filled] -> Compose -> Synthesize -> Listen
Use Turbo model first, and SFT work also for experts

"Compose" launch a LLM inference, "Synthesize" launch a DiT inference.
"Inspire" -> complete metadata
"Format" -> format metadata

So far when I tried SFT or base I did nnot get good enough instrumentals, it was alwasys too aggressive, turbo was a bit better, did you figure out how to use all these models

The base model is used for Repaint, and when you open an MP3 for understanding mode (it generates metadata with more or less hallucinatory lyrics that often stick to the syllabic rhythm, the model is not a speech recognition system).
Turbo is cleaner and simpler to use. You also have three variants: "continuous," "shift1," and "shift3," which function like Turbo (read the official ACE-Step doc linked on readme). SFT has a richer latent sound, but it requires more exploration and cherry-picking of the right music.

I usee you are using QWEN, is it much much better than the base LM model from acestep team?

It's not comparable; in fact, Ace-Step's LLM is specialized in generating the audio codes that guide the DiT. It has been trained on tons of metadata versus music; it's a composer.
However, using a different standard causal LLM allows you to write much better lyrics that are more suited to your wishes, but still need the Ace-Step LLM to generate audio code !

Anything else I should know? (in term of features for example)

You really need to refer to the original Ace-Step project documentation to understand the principles; I will also improve my documentation as I go along!

[StableInfo]

Thanks a lot! No no its not a lack of documentation from your end, its just me thing, I just learned about this and thought I write a message before I surf in!
Last question: is this safe?
I don't know anything about cpp and stuff, but hope this stuff are safe for our machines.
I have a suggestion I will make in another post.

[ServeurpersoCom]

To put it simply: none. C/C++ are the fastest languages ​​and are used for anything serious and/or performance-related.

• AI calculations push the hardware to its limits: it shouldn't overheat!!!!

• Security: if you ever open the WebUI to the internet, you have to do it properly (firewall, secured container etc...), but that goes beyond normal use. I do it here: https://www.serveurperso.com/ia/music/ (my dev instance)