ACE-Step 1.5 XL — Inference Meta
A practical guide to making ACE-Step XL actually sound good.
Settings
| Setting | Value |
|---|---|
| Model | XL Base/Turbo 0.3 Merge, Q8 GGUF |
| VAE | ACEStep-XL Regrind V10b BF16 |
| UI | HOT-Step-CPP |
| Sampler | MD Storm 4 (DPM++ 3M + STORK 4) |
| Schedule | Linear |
| Guidance | APG |
| Steps | 50 |
| CFG | 12–20 (I'm on 20 most of the time) |
| LM | Disabled |
"The LM setting is the one people get wrong"
Most of ACE-Step was trained without the LM. The devs suspected that's how Udio works, since DiTs are more creative than LMs. In practice the model is just better most of the time with the LM off. The exception is very complex prompts where you need literal accuracy, but even then, the DiT will sound more musically impressive, and it follows complicated prompts fine if you push the CFG up.
"Don't skip the VAE"
The Regrind VAE cleans up artifacts well beyond what the merged model can do alone. It's what I used for the Yousei Teikoku generations below.
Which Model To Use
ACE-Step 1.5 XL Base/Turbo 0.3 Merge — Q8 GGUF
Merged models are far and away the biggest breakthrough the ACE-Step XL community has produced. This one is 70% Base, 30% Turbo. The Base half keeps all the creativity and range; the Turbo half patches up most of the things Base gets wrong on its own.
The result is a model with better sound quality than either Base or Turbo alone, while staying as versatile as Base. It's the closest anyone has gotten to Suno/Udio territory with this architecture. If you want crisp vocals and instruments that don't sound squashed and over-compressed, this, combined with the improved VAE, are doing a lot of the heavy lifting.
UI
Start with acestep.cpp and get a feel for what it can do. Once you're comfortable, move to HOT-Step-CPP.
I'd point you at HOTStep.cpp specifically because it has the MD Storm 4 sampler I recommend above.
Why the .cpp UIs and not ComfyUI? Speed, mostly. They're significantly faster than the ComfyUI implementation, which feels sluggish and inefficient by comparison. Note that the official Gradio implementation is to avoided at all costs, as it's vibe-coded slop riddled with even more issues than ComfyUI.
The Three Released Models (and why the merge exists)
ACE-Step 1.5 XL shipped as three separate models. Here's the honest rundown.
XL Base — 50 steps
The foundation. Easiest of the three to finetune on basically any concept, and it picks up the most detail. It's also the most creative of the bunch.
The catch: vocal quality is rough, long-form coherence can fall apart, and it likes to cut off early or come out way too quiet. Most of that has been fixed by the community, which is why we're still using it, just in modified form.
XL Turbo — 8 steps
Somewhat creative, but its genre and instrument knowledge is a fraction of what Base has, and nowhere near Suno or Udio. Outputs are noisy and full of artifacts. Useful as an ingredient in a merge, not really as a standalone.
XL SFT — 50 steps
A lot of people use this one because the devs claim it picked up detail better during training.
That's bullshit.
SFT is arguably the worst of the three. Treat it as a generic music model with an extremely narrow creative range, it'll do pop, hip hop, and a handful of other styles fine, but with a limited palette of instruments even inside those genres. Step outside what it knows (a subset of what Base and Turbo know) and the sound quality gets worse, not better. It's also by far the hardest to finetune.
I genuinely don't know why this one was released. The previous 1.5 SFT was a real improvement over 1.5 Turbo. The XL version isn't.
Example Generations
These are all from LoRAs, using roughly the settings above.
Yousei Teikoku LoRA
ZUTOMAYO LoRA
Initial D LoRA (slightly worse settings with inference on 0.5 merge, I've since improved them, so expect a few more issues here)
Why This Guide Exists
ACE-Step 1.5 XL was a great update. But if you follow the official tutorials, which are, frankly, poorly maintained, you'll hit a wall and walk away convinced the model is bad.
Honestly, you wouldn't be blamed for thinking that. Go listen to the official 1.5 (4B) demos:
https://ace-step.github.io/ace-step-v1.5.github.io/
If you're going in excited, like I was, it's easy to miss that those demos are riddled with artifacts and the vocals sound drowned-out and synthetic next to real music. There's also no discussion anywhere of what settings produced them. I'll save you the time: they used the Turbo XL model.
To be fair to the team, the prompting guide is genuinely good, and releasing the training code was a real gift (MiniMax never did). But the guide was never updated as the community improved the model, and the devs have gone quiet since release. Their attention is on the cloud offering, ACE-Studio. However, the philosophy of ACE-Step devs is superior to others, since they prioritize customizability and trained the model with this in mind, and it shows, ACE-Step XL is much easier to prompt and much more customizable than Minimax Music 3.
The result: the model never took off, and plenty of people wrote it off as never having caught up to Suno, despite the dev team claiming it matched Suno v4.5/v5.
So this guide is my attempt at filling that gap. Follow it and you should be able to get generations that hold up on a decent pair of headphones next to Suno or Udio, at least sound quality wise. It does not fully bridge the gap with Suno or Udio since their base models were trained on a larger Youtube dataset (similar to Minimax H3), but with a LoRA trained on either an artist or genre, ACE-Step XL bridges that gap.
Note on future improvements: ACE-Step XL is currently undergoing lots of exciting developments, and future samplers from devs such as MDMachine may fix even more issues with vocals and sound quality.
Improving Generations
Not fully satisfied with current results? No issue, you can improve the results by mastering them with Matchering 2 (available on HOTStep.cpp UI), or downloading Ultimate Vocal Remover. Find a well-mastered song in the same genre as your target song, and simply use this.
Training Guide
Refer to https://rentry.co/s8fg8ber
Though it is slightly outdated. I now train my LoRAs at rank 256 with an LR of 0.00009 since the higher rank picks up most details, as seen on Yousei Teikoku or Initial D LoRAs (usually just 20 songs is enough, but sometimes more may be needed at high rank depending on dataset).