ACE-Step 1.5XL Turbo: Text to Music
Generate high-quality music from text prompts using the distilled 4B ACE-Step model. Produces commercial-ready audio in just 8 inference steps without CFG.
试试一个示例
关于此工作流
This workflow turns a plain-language text prompt into finished music using the ACE-Step 1.5 XL Turbo model. Your prompt is encoded by TextEncodeAceStepAudio1.5 (backed by DualCLIPLoader) into audio-aware conditioning. EmptyAceStep1.5LatentAudio creates a blank latent audio canvas at your chosen length, and UNETLoader loads the distilled 4B ACE-Step checkpoint (acestep_v1.5_xl_turbo_bf16.safetensors). ModelSamplingAuraFlow configures the sampler schedule tuned for ACE-Step, and KSampler performs just 8 diffusion steps to synthesize the latent audio.
The audio is then reconstructed with VAELoader + VAEDecodeAudio and saved via SaveAudioMP3.
What makes this workflow practical is its “turbo” configuration: no classifier-free guidance (CFG) is used, enabled by ConditioningZeroOut feeding an intentionally empty negative conditioning. This keeps prompts straightforward (focus only on what you want), while ModelSamplingAuraFlow and KSampler deliver fast, stable results. The graph is organized into Model and Prompt groups for clarity, and uses PrimitiveInt/PrimitiveNode for quick edits to seed, steps, and duration.
The result is a tight, teachable text-to-music pipeline you can run repeatedly for variations or integrate as a reusable subgraph.
常见问答
This workflow intentionally disables classifier-free guidance using ConditioningZeroOut for fast, stable sampling. Negative prompts are ignored in this setup. If you need negatives, replace ConditioningZeroOut with an explicit negative encoder path and enable CFG in KSampler—but that changes the workflow’s behavior.
Set the length in EmptyAceStep1.5LatentAudio (seconds or samples, depending on the node’s fields). The model interprets structure via text (e.g., “16‑bar trap loop,” “slow build and soft drop”). For longer pieces, generate multiple segments and arrange them in a DAW.
Keep KSampler at 8 steps with ModelSamplingAuraFlow connected—this is tuned for ACE-Step 1.5 XL Turbo. If you hear artifacts, try modest increases (e.g., 10–12 steps). Maintain CFG disabled to match the workflow’s design.
The graph saves MP3 via SaveAudioMP3. You can adjust its bitrate if the node exposes that option. If you prefer WAV, swap in a compatible WAV saver node and connect it after VAEDecodeAudio.