Someone merged Turbo and SFT and it is a huge improvement in my opinion.

#19
by Dampfinchen - opened

https://huggingface.co/Aryanne/acestep-v15-test-merges/blob/main/acestep_v1.5_merge_sft_turbo_ta_0.5.safetensors

It has less crust, higher overall sound quality, less wrong notes. I can really recommend it over the turbo one.

Base really beats them all. It takes some more prompting and experimentation to get what you want, but when you do .. blows you away. It's exactly what Gong Junmin writes in the Tutorial:

Unlike the strong purposefulness of one-click generation, human-centered generation has more of a playful nature. It's more like an interactive game where you and the model are collaborators.

The workflow is like this: you throw out some inspiration seeds, get a few songs, choose interesting directions from them to continue iterating—

  • Adjust prompts to regenerate
  • Use Cover to maintain structure and adjust details
  • Use Repaint for local modifications
  • Use Add Layer to add or remove instrument layers

At this point, AI is not a servant to you, but an inspirer.

And this is so much more true for -base.

Can you please share your opinion of the sft version? I've been using that extensively since the beginning. How do you find it compares?
Thx.

Sign up or log in to comment