ACE-Step 1.5XL SFT: Text to Music
Generate high-quality music from text prompts using a 4B parameter SFT model. Control output with CFG for precise prompt adherence.
Try an example
About this workflow
This ComfyUI workflow turns text prompts into finished music using the ACE-Step 1.5XL SFT (4B) model. It follows a familiar diffusion pipeline adapted for audio: DualCLIPLoader and TextEncodeAceStepAudio1.5 convert your prompt (and optional negative prompt) into conditioning; EmptyAceStep1.5LatentAudio creates a latent audio canvas for your chosen duration; UNETLoader brings in the ACE-Step 1.5XL SFT model; and KSampler performs iterative denoising guided by CFG. ModelSamplingAuraFlow configures the sampler strategy tuned for ACE-Step audio, while VAELoader + VAEDecodeAudio convert the final latent back into a waveform that SaveAudioMP3 writes to disk.
Technically, CFG (classifier-free guidance) lets you balance creativity and prompt adherence: lower CFG values explore broadly, higher values follow your text more strictly. ConditioningZeroOut provides a clean unconditional branch if you don’t specify a negative prompt, ensuring stable guidance. Workflow groups (Model, Duration, Prompt) make it easy to swap checkpoints, set seconds, and iterate on prompts.
The result is a practical, reproducible text-to-music pipeline that can quickly iterate across genres, moods, instruments, and structure—ideal for ideation, temp scoring, and background tracks.
Questions & answers
KSampler performs the denoising process using your steps, CFG, and seed. ModelSamplingAuraFlow configures the sampler strategy/schedule tailored for ACE-Step audio, so you get stable, high-quality convergence without manual scheduler tuning.
Use EmptyAceStep1.5LatentAudio in the Duration group. Adjust its duration or frame parameters (often driven by connected PrimitiveInt/PrimitiveNode controls). Longer clips require more compute and memory.
Refine the prompt with clearer descriptors (genre, instruments, mood, tempo, mix notes) and add a negative prompt (e.g., "distortion, artifacts, vocals"). Try lowering CFG slightly if the audio is harsh, or increasing steps moderately if it lacks detail. Changing the seed can also help.
This workflow saves with SaveAudioMP3. If you prefer WAV, add or swap in a SaveAudioWAV node (if available in your setup) after VAEDecodeAudio and select the desired format and path.