Do I need specific model files for this to work?

Yes. Point UNETLoader and VAELoader to the Ace-Step 1.5XL UNet and VAE checkpoints. The workflow relies on TextEncodeAceStepAudio1.5 for conditioning, so ensure the corresponding text encoder resources are available as packaged with Ace-Step.

How accurate are the BPM and duration settings?

Duration is enforced by the latent length from EmptyAceStep1.5LatentAudio, so the output length will match closely. BPM guides the lyric density via TextGenerate and prompt injection but does not force strict metronomic alignment; treat it as a strong narrative/timing hint rather than a beat-locked constraint.

Can I use my own lyrics?

Yes. Switch to the Manual path and paste lyrics into PrimitiveStringMultiline. The ComfySwitchNode routes your text directly to TextEncodeAceStepAudio1.5 for conditioning before generation.

How do I improve audio quality or reduce artifacts?

Try increasing KSampler steps moderately, lowering CFG if the sound is harsh, and keeping prompts concise and specific. Ensure the correct Ace-Step VAE is loaded, and avoid overly long durations on a single pass if you encounter instability.

Song Generator - ComfyUI Workflow

Song Generator is a ComfyUI workflow that synthesizes complete songs with the Ace-Step 1.5XL audio model. It supports two lyric paths: automatic, where a short topic is expanded into BPM- and duration-aware lyrics, and manual, where you paste your own text. A ComfySwitchNode routes between these paths before the lyrics are encoded by TextEncodeAceStepAudio1.5 and used to guide the audio model. You can preview the result with PreviewAny and export with SaveAudioAdvanced.

Under the hood, UNETLoader and VAELoader load the Ace-Step 1.5XL checkpoints. EmptyAceStep1.5LatentAudio allocates a latent audio canvas sized to your target duration. ModelSamplingAuraFlow configures the sampler for the model, and KSampler performs the iterative denoising on that latent using the conditioning produced by TextEncodeAceStepAudio1.5. ConditioningZeroOut is included to safely zero optional negative conditioning when not used. After sampling, VAEDecodeAudio converts the latent to a waveform for monitoring and export.

The BPM/duration-aware lyric module uses PrimitiveInt/PrimitiveFloat to capture BPM and target length, StringReplace to inject those values into a songwriter instruction, and TextGenerate to create structured verses/choruses sized for the track. StringCompare helps detect empty input, and the ComfySwitchNode cleanly selects between the auto-lyrics and a PrimitiveStringMultiline node for manual lyrics. DualCLIPLoader/CLIPLoader are available for workflows that optionally mix in CLIP-style text conditioning, though the primary path uses TextEncodeAceStepAudio1.5 for Ace-Step.

FAQ

常见问题

查看所有工作流