YuE2-3B runs on cards with as little as 4.5GB VRAM

@7117 And the full BF16 on ~7GB

This free site is good and can quickly add an image to an .mp3 for free so you can upload it to Civitai: https://www.neuralframes.com/tools/audio-to-video

@J1B Thanks, just what I was looking for. 👍

That would help for my Music Focus collection https://civitai.com/articles/35308/the-music-focus-collection

Although mp3? 🧐
Don't forget FLAC, Opus, m4a (AAC).

@bno84 This site allows you to convert audio formats to FLAC and M4A. For instance, NeuralFlame supports M4A audio generation. The site also includes basic editing features, such as adding fade-ins and fade-outs, which is highly useful since AI-generated music files often end abruptly. 😂Forgot the link:https://mp3cut.net/

Ofcourse downloading Audacity or som other apps is the best way. But that site is easy and quick to use, for fast work.

@J0hn_D0e The conversion isn't an issue (LLMs and ffmpeg exists). But allowing it to be uploaded to Civitai. MP3 is imo completely obsolete so if audio uploads get supported, I think they should consider:
- Opus as the gold standard for lossy
- M4A/AAC for broad compatibility lossy
- FLAC for lossless

@bno84 Sorry I misunderstood your initial post. I agree, lossless formats should be a option. But the cost of hosting them would probably need to be covered somehow.

@J0hn_D0e Yes the size of lossless is gnarly. But the site also has video. My 15 second video clips tend to be about the same as a 2½min flac track ~30MB or so. But with Opus it's 10x smaller and indistinguishable. So maybe they should put a length-cap on the lossless formats if they want to allow but keep costs down. Say 30s for lossless?

Have you never installed a custom node for Github before? or are you saying there is a security issue with that one?
I have over 200 custom node packs and never really had a problem, although it is not unheard of/without any risk.

i agree, but generally will just go check out the actual node on github or whatever and then manually install it. that's all its doing anyway

@J1B I have absolutely no trust in that blackbox of a node that forces downloads.
But please, tell me the nodes you're using to load the int8_convrot model and the audio encoder separately...
@makiaeveli same question.

@GlowingGuardianGirl Who created the models don't have to obey on what you desire but on their. If they created an AIO model there will be a rerason and maybe if you read into their repository you'll understand why. INT8 convorot is natively supported by ComfyUI. If you don't like it noone force you. Just a little respect on other work especially if they can let you use it for free.

@GlowingGuardianGirl I'm just using the old school checkpoint loader and it can load the int8 and bf16 models, not sure what I did for the audio encoder, I think that was a custom node.

@J1BIf you use the template from Comfyui u'll understand. No need to use a custom node for the audio encoder. Audio encoder is for remixing audio, not to generate a new one.

Yes, kind of. Check out their website.
You actually get scores as an artifact, you can have an LLM edit the scores for you (they have a skill) and re-render the audio from it. Or have the model complete a short transcription.

You can feed your own music into the Sheetsage2 node, and it will generate an ABC music score. Alternatively, you can simply use any existing ABC score or generate one using an LLM.
Also, when generating, leave the "Style" field blank if you need to adhere strictly to the ABC score, and specify your vocals in the "Vocal" field.

@BibaBobacivitai There are gotchas, they require a simplified ABC dialect.
2 tracks + chord progression, monophonic. Your style prompt needs to influence whether it's likely that monophonic melody becomes power chords, single notes, etc.

There is a instrumental only lora here: https://huggingface.co/Mothersuperior/YuE2-instrumental-cot-full-loras

https://github.com/multimodal-art-projection/YuE/issues/18