Sidestep just got a GUI that beats Lora training in AceStep by far
SideStep had a huge update. It's not only a GUI. SideStep just became SideStep-Studio. There is nothing you can't do with it. You won't miss a single function. You are rather like "Oh, I can do that too?" Too many new features to only list them all here. But the GUI comes with an introduction guide explaining everything you need to know to get started. So I will just show some of the cool new features it provides. My favorite feature is certainly AI driven Captioning that beats the captioning of AceStep by far. It simply nails it. And you don't have to care about anything at all. But later more...
SideStep has an Easy Mode. All you got to do is pull your audio files into the right folder and select it in the dataset dropdown menu. After clicking Start Training they get preprocessed and you can select the preprocessed files in the same dropdown menu and start training. That's it. For instrumentals this is the easiest way to train.
The Advanced mode allows you to control all the parameters and settings of the training. Including memory Management. The VRAM estimation at the top right helps you to find the best settings for your GPU.
When the Training starts the Monitor shows you a live view of the training powered by Tensorboard. As well as the Steps, Epochs, estimated time, VRAM consumption and more...
The Lab allows you to automatically caption and analyze all songs of your dataset. You can edit and even play each song.
Contrary to AceStep, editing your Songs to fix the Captioning is not necessary. If you use Gemini for captioning it will nail the captions and genre.
Therefore you need an API Key you can get for free here if you have a google account. But it's also possible to use a local model for the captions.
The Lyrics get automatically scrapped by Genius if you have an API key. This is also completely free. All you need to do is sign in to Genius for free and generate an API key.
I don't know if there is a proper way to get a Genius API Key. I searched quite a while, but if you can't find one, click this Link.
On that page click To get your API Token for GeniusBadge click here! and it will lead you to a page with a 404 Error. But in the address field you will see your API Key and it works. If you find a better way to get a Genius API Key, please let me know in the comments.
A few things to AceStep I figured out while testing my Lora:
When you change the Lora in AceStep and you get an error, initialize the model again.
Although I trained my Lora with the AceStep base model, I got the far best results using the turbo model together with my Lora in AceStep.
I tried almost all combinations, but initializing without LM, with pt backed and without 8-bit Quantizatin (since this is not supported when using a Lora) gave the best and fastest results.
I traiend a Lora on only one album of Slipknot, what must be total chaos for the model. Here is an example of the outcome. For those knowing Slipknot, you'll recognize at once.