Britney Spears lora:

https://voca.ro/16blE7la2Ff8
https://voca.ro/1eLHswbE9ZHK
https://voca.ro/1me4VBIkzHfK

To that one anon that is claiming "lora training on SFT doesn't work": this is a Lora trained on SFT =)

based thread! i don't use ACE Step but the best wisheS!

>>108164777
Nakadashee AceStep-chan

File: default.jpg (23 KB, 586x275)

23 KB JPG

>>108165006
it is...ok, can hear her singing style from time to time.
my internets is cutting out for half a day and more lately since i live in a third world country (australia), so yesterday i had time to sit and test default settings as per developer instruction, results were meh via gradio.
must test comfy nodes since i got better results than via gradio interface.
one note, manual captions, removing redundant stuff like too many attributes llm gives (in case it does detect correct instruments), help quite a bit.

enya anon done it really well via overfit, and if he sees this post;
what was your overfit setting?
high lr low rank low small-medium dataset?

>>108165607
>enya anon done it really well via overfit, and if he sees this post;
>what was your overfit setting?
>high lr low rank low small-medium dataset?

I used the default 0.0003 LR at 800~1000 epochs (I can't remember where I stopped), my dataset consisted of 24 songs

>>108165764
As of the rank, I used rank 128 I think, it's the maximum my card supports without going OOM

>>108165776
>>108165764
ty, will try it.

i used 12 songs, same genre and different bands, results are meh and sometimes ok.

comfy gens are better, if i crank lora strength to 2 there is that fm radio but super high can+static noise, yet it does replicate training set song at around 80% of the content.

>>108166294
>fm radio
AM -.- radio

>>108166294
>comfy gens are better, if i crank lora strength to 2 there is that fm radio but super high can+static noise
I said that in last thread and I am going t say again, you are probably undertraining your models. Use a high enough LR, train longer, and use a high rank.

>>108166359
Also DO NOT USE THE LLM.
The LLM tends to weaken the Lora effect, sometimes it even changes the voice/singing style

does anyone have access to the suite Sony has produced to compare?

>>108166444
>suite Sony has produced
What even that is?

>>108165776
In my experience 128 has been good. I think people need to cook their LoRAs a lot longer though. At 1500 epochs 2000 or more is probably better.

>>108166395
Yeah do not use the LLM if you are using a LoRA. It straight up makes it own song and the LoRA is just antagonistic to it.

>>108164777
I'm a vramlet with 12GB VRAM and 32GB RAM. Can I use this and make loras? Also, I have a degree in music. Can this help me with anything on this matter, for example "make a bimodal pop song" or "make a kpop song using the hungiran minor scale"?
Also I'd like to train LoRAs of classical composers and mix them with other musical aesthetics.
Can I make a LoRA of my own musical compositions to figure out what resources or patterns I usually use?

>>108166694
>I'm a vramlet with 12GB VRAM and 32GB RAM.
Maybe, but you won't be able to use a higher lora rank on a regular trainer. You can try to vibe code with Claude or something a training script that uses Ramtorch or something like it to offload everything to RAM and train on GPU one demand, but it will probably be very slow.

>Can this help me with anything on this matter, for example "make a bimodal pop song" or "make a kpop song using the hungiran minor scale"?
If you create a properly tagged dataset, maybe. And btw, the model natively already requires you to tag songs by keyscale and bpm, besides the captions.

>Can I make a LoRA of my own musical compositions
Yes

>to figure out what resources or patterns I usually use
You can already "figure out" that by captioning the songs on Gemini Pro, you can explicitly ask it