Loading course…
ACE-Step 1.5 Local Text-to-Music Workflow
Created by Shaunak Ghosh
Run ACE-Step-1.5 locally with the official uv + Gradio toolchain, validate a known-good first generation, and then drive quality with structured prompts and the key inference controls. You’ll finish with a practical benchmarking mindset for comparing outputs to Suno-style proprietary systems, including what variability and limitations mean for real workflows.
Requirements
- Comfort with Python environments and CLI tools
- GPU/VRAM constraints and CUDA basics (or willingness to test CPU)
- Ability to evaluate generated audio for adherence and artifacts
- Basic music metadata vocabulary (BPM, key, time signature)
What you'll learn
- Choose a viable local ACE-Step-1.5 setup path and validate a known-good first run, distinguishing device/backend mismatch from download or initialization issues.
- Convert musical intent into a structured, non-contradictory song spec (caption + section-tagged lyrics) that supports controlled iteration.
- Make principled inference trade-offs by selecting model variants and tuning shift and thinking LM settings based on quality needs, speed, and VRAM budget.
Learning path
4 modules • Each builds on the previous one
1
ACE-Step 1.5 setup options
Set up ACE-Step 1.5 locally using the official repo, select an appropriate backend (CUDA/ROCm/MLX/Intel XPU/CPU), and verify that the DiT + optional 5Hz LM components initialize reliably for your hardware. ([github.com](https://github.com/ace-step/ACE-Step-1.5))
1 video7 min
2
Text prompts to song spec
Generate music from text by converting an idea into a structured song spec (caption, lyrics, and optional metadata like BPM, key, duration), using Simple mode to bootstrap and Custom mode to lock down intent. ([huggingface.co](https://huggingface.co/spaces/ACE-Step/Ace-Step-v1.5/blob/main/docs/en/GRADIO_GUIDE.md))
1 video6 min
3
Quality controls: steps, CFG, LM
Tune quality vs speed by choosing the right model variant (turbo/sft/base) and manipulating core controls (inference steps, guidance/CFG behavior, shift/timesteps, seeds, and LM “thinking” settings) to get reproducible improvements. ([huggingface.co](https://huggingface.co/ACE-Step/Ace-Step1.5))
1 video6 min
4
Benchmarking, Suno comparison, limitations
Benchmark ACE-Step 1.5 output quality and speed, design a fair comparison against proprietary services like Suno, and translate results into realistic workflows and known limitations (including responsible-use risks). ([arxiv.org](https://arxiv.org/abs/2602.00744?utm_source=openai))
1 video12 min
Start Learning
Begin your learning journey
Modules4
Duration30 min
Science-backed learning
In-video quizzes and scaffolded content to maximize retention.
Key concepts
Official Local Setup And Validation (Uv + Gradio)Two-Model Pipeline: Generator And Thinking LMPrompt-To-Spec Workflow: Captions, Sectioned Lyrics, Metadata Intent