ACE-Step 1.5 + Chiptune Remix Setup
Target system: Windows 11, RTX 2080 Ti (11GB, Turing)
Part 1 β Prerequisites
- Update your NVIDIA driver. Use the current Game Ready or Studio driver from NVIDIA's site. Don't rely on whatever Windows Update installed.
- Install Python 3.11 or 3.12. Get it from python.org, not the Microsoft Store. Tick "Add python.exe to PATH" during install.
- Avoid 3.13 β there's a known torchaudio backend problem on Windows, plus general dependency friction across the AI stack.
- Install Git for Windows. https://git-scm.com/download/win β defaults are fine.
- Install
uv(optional but recommended, it makes the ACE-Step install much cleaner):pip install uv - Free up ~50GB of disk. Model weights plus dependencies add up fast.
Part 2 β Install ACE-Step 1.5
- Open PowerShell and pick a working folder:
cd C:\ mkdir AI cd AI - Clone the repo:
git clone https://github.com/ACE-Step/ACE-Step-1.5.git cd ACE-Step-1.5 - Install PyTorch with CUDA first. Go to https://pytorch.org/get-started/locally/, select Stable / Windows / Pip / Python / CUDA, and run the command it gives you. Getting this from the selector rather than a copied command matters β a mismatched CUDA build silently falls back to CPU.
- Install the rest:
Or, without uv:uv syncpip install -r requirements.txt - Verify CUDA is actually live before going further:
You wantpython -c "import torch; print(torch.cuda.is_available(), torch.cuda.get_device_name(0))"True NVIDIA GeForce RTX 2080 Ti. If it saysFalse, step 3 didn't take. - Launch the web UI:
Then open http://127.0.0.1:7860 in your browser. First launch downloads the model weights β expect a long wait.python app.py --port 7860
Turing-specific settings
- Use FP16 weights, not BF16. The RTX 20-series lacks native bfloat16 support and emulates it, which is slower. Where a model variant offers both, pick FP16 and set the precision flag explicitly rather than leaving it on auto.
- Skip the XL variant. It wants 12GB minimum and 20GB recommended; you have 11GB. Standard 1.5 has every editing mode XL has.
- Don't use
--halfor CPU offload. Those are for cards under ~6GB. You have headroom β running full speed is the point.
Part 3 β Source material tools
Getting the Gen 1 tracks
- foobar2000 + the foo_gep component (Game Emu Player). This plays
.gbsfiles and β the important part β lets you mute individual channels and export each one separately. That gives you clean stems (pulse 1, pulse 2, wave, noise) with no AI separation needed. - VGMusic.com β the long-running archive of MIDI transcriptions. Gen 1 is thoroughly covered. MIDI is the better starting point if you want to change tempo, key, or groove.
DAW (for re-instrumenting the MIDI)
Pick one:
- REAPER β best fit for this kind of work. Fully functional unlimited evaluation, $60 personal license if you keep using it.
- Waveform Free β genuinely free, no feature crippling, good MIDI editing.
- LMMS β free and open source, lighter weight, a bit rougher around the edges.
Supporting tools
- Audacity β trimming, normalizing, and format conversion between steps.
- A general MIDI soundfont (e.g. FluidR3_GM) β gives you usable instruments immediately so you can hear the arrangement before committing to real VSTs.
Part 4 β The actual remix workflow
- Pick a track and get both formats β the GBS rip and the MIDI transcription.
- Decide your route:
- Melody-faithful: start from the MIDI. Open it in your DAW, swap instruments, adjust tempo and swing. You control exactly how recognizable the tune stays.
- Texture-faithful: start from the GBS audio. Export the muted channel stems from foobar2000.
- Rough out the arrangement in the DAW. Even a crude version β right instruments, right tempo, right feel β gives ACE-Step something coherent to work from.
- Bounce to WAV and load it into ACE-Step's editing modes:
- Cover β re-synthesizes new timbres over the existing melodic skeleton. The main tool for "same song, different genre."
- Layering β adds complementary instruments to what you feed it. Good for fleshing out a sparse chiptune arrangement.
- Completion β builds a full arrangement from a single motif. Use when you want it to take real liberties.
- Repaint β regenerates one time-window. Use to fix a section that came out wrong without re-rolling the whole track.
- Iterate on the prompt. Describe instrumentation, tempo, and production style, not the song. The model doesn't know PokΓ©mon music β the melody has to come from your audio input every time.
- Final mix back in the DAW. ACE-Step output has a faint metallic shimmer on sustained notes and cymbals. It hides well under lo-fi, breakbeat, or distortion; for clean jazz you may want a de-esser or gentle high-shelf cut.
Troubleshooting
| Symptom | Cause | Fix |
|---|---|---|
CUDA out of memory |
Duration too long, or batching | Shorten the generation, drop batch size to 1 |
torch.cuda.is_available() returns False |
CPU-only PyTorch installed | Uninstall torch, reinstall from the pytorch.org selector |
ModuleNotFoundError: acestep |
Package not installed | pip install git+https://github.com/ace-step/ACE-Step.git |
| torchaudio backend error | Python 3.13, or missing soundfile | pip install soundfile scipy β or downgrade to 3.12 |
| Generation is very slow | BF16 emulation on Turing | Switch to FP16 weights and set precision explicitly |
Note on releasing
Nintendo owns these compositions and is more active than most publishers about fan music. Making remixes is uncontroversial; publishing them is where takedowns happen, and a takedown is usually the extent of it. Worth knowing before you build a channel around it.