How to Run ACE-Step 1.5 XL Locally for Free

TL;DR

ACE-Step 1.5 XL can generate vocals, instrumentals, and multilingual songs locally for free, with at least 12 GB of VRAM when CPU offload and int8 quantization are enabled. The model supports NVIDIA, AMD, and Apple Silicon hardware, while 20 GB of VRAM is recommended for running the full model without offloading.

Transcript

Baby, you're the one I know. I don't need the stars and I never said it right. Always too late or too quiet. But you fit inside my ordinary life. Are you So Mustpping. This is now the best open- source music generator you can use right now. Heck, apparently it even beats the best closed models out there like Sununo and Udo. Not only that, but it's ... Read More

Key Insights

• ACE-Step 1.5 XL is an open-source music generator that produces complete songs locally, allowing free and unlimited generation after installation. The demonstrated outputs include vocals, lyrics, instrumentals, choir passages, multiple languages, and several distinct musical styles.
• The XL release improves audio quality and consistency compared with the previous ACE-Step 1.5 version. Its generated vocals are described as cleaner, more dynamic, and more convincing, while its instruments also sound noticeably better than those produced by the earlier model.
• The model supports a broad range of genres and languages, including soulful R&B, Italian opera, Spanish Latin trap, J-pop with Eurobeat and trance influences, children's songs, jazz, Chinese bossa nova, and tango. These demonstrations emphasize its flexibility across different musical structures and vocal styles.
• Instrumental generation works by omitting lyrics and adding an instrumental tag. Prompts can also specify arrangement changes such as when a flute continues, a harp enters, or a cello begins, enabling more deliberate control over the progression and instrumentation of a generated composition.
• The minimum stated memory requirement is 12 GB of VRAM when CPU offload and int8 quantization are enabled. Running the entire official XL model on the GPU without offloading requires at least 20 GB, while adding a language model may call for about 24 GB.
• CPU offload and int8 quantization theoretically reduce output quality slightly, but the presenter reports hearing little difference. Community-created quantized versions are already smaller than 10 GB in file size, although the stated current minimum hardware target remains 12 GB of VRAM.
• The official benchmark results claim that ACE-Step 1.5 XL surpasses Udio and Suno version 5 across measures such as song coherence, musicality, and naturalness. The same benchmark presentation claims generation can be up to 120 times faster than other models for a four-minute song.
• The available XL checkpoints serve different purposes: the base model is intended for training and fine-tuning, while the SFT and turbo models are intended for music generation. SFT uses more steps and offers somewhat better quality, whereas turbo uses fewer steps and runs faster.

Questions & Answers

Q: How do you install ACE-Step 1.5 XL locally?

Install UV first, then install Git if it is not already available. Choose a destination folder, open a terminal there, and clone the official ACE-Step 1.5 GitHub repository. Change into the cloned directory and run UV sync. UV creates a separate virtual environment and downloads the packages and dependencies required by the interface. After that process finishes without errors, download an XL generation model.

Q: How much VRAM does ACE-Step 1.5 XL require?

ACE-Step 1.5 XL requires at least 12 GB of VRAM when CPU offload and int8 quantization are enabled. A GPU with at least 20 GB is recommended if the entire official model must remain in GPU memory without offloading. Pairing the generator with a language model for thinking mode requires additional memory, with approximately 24 GB suggested as a safer capacity.

Q: What music styles can ACE-Step 1.5 XL generate?

ACE-Step 1.5 XL can generate a wide selection of styles demonstrated in the source, including soulful R&B ballads, Italian opera, Spanish Latin trap, J-pop with Eurobeat and trance influences, cheerful children's songs, jazz, Chinese bossa nova, and tango. It can produce vocal songs, fully instrumental pieces, and arrangements that combine instruments with choir sections.

Q: Can ACE-Step 1.5 XL generate songs in different languages?

ACE-Step 1.5 XL supports multilingual song generation. The demonstrations include Italian opera, Spanish Latin trap, and a Chinese bossa nova track, alongside English-language examples such as an R&B ballad and a children's song. These samples show that the model can apply different languages to distinct genres instead of being limited to one language or one vocal style.

Q: Can ACE-Step 1.5 XL create instrumental music?

ACE-Step 1.5 XL can create instrumental music when no lyrics are supplied and the prompt includes an instrumental tag. The demonstrated tango piece uses this approach. The model can also combine instrumentals with choir passages and follow arrangement directions that identify when specific instruments should enter, including flute, harp, and cello cues placed within the requested musical structure.

Q: What is the difference between the SFT and turbo models?

The SFT and turbo checkpoints are both intended for generating music, but they prioritize different outcomes. The SFT model runs more slowly and requires more generation steps, with the benefit of somewhat better quality. The turbo model uses fewer steps and generates faster, while sacrificing some quality. The separate base model is intended for training and fine-tuning rather than ordinary song generation.

Q: Does quantization reduce ACE-Step 1.5 XL quality?

CPU offload and int8 quantization theoretically cause a slight reduction in output quality, according to the explanation in the source. However, the presenter reports that the difference is difficult to hear in practice. These options make it possible to use the XL model with 12 GB of VRAM instead of requiring at least 20 GB to keep the full model on the GPU.

Q: How does ACE-Step 1.5 XL compare with Suno and Udio?

The official benchmarks shown in the source claim that ACE-Step 1.5 XL outperforms Udio and Suno version 5 across measures that include song coherence, musicality, and naturalness. They also claim that it can generate a four-minute song up to 120 times faster than other models. These comparisons are benchmark claims presented by the model's creators, not independent test results described in the source.

Summary & Key Takeaways

• ACE-Step 1.5 XL is presented as an open-source music generator with improved audio quality, vocal clarity, consistency, coherence, musicality, and naturalness. Demonstrations cover soulful R&B, Italian opera, Spanish Latin trap, J-pop, children's music, jazz, Chinese bossa nova, tango, and compositions combining instruments with choir sections.

• The XL model requires at least 12 GB of VRAM when CPU offload and int8 quantization are enabled, while 20 GB is recommended for holding the full model in GPU memory. Pairing it with a language model for thinking mode may require approximately 24 GB to provide additional memory headroom.

• Installation begins by adding UV and Git, cloning the official ACE-Step 1.5 repository, entering its directory, and running UV sync. UV creates an isolated virtual environment and installs the necessary packages and dependencies. Users then download either the higher-quality SFT model or the faster turbo model for music generation.