Sampler Nodes

The Sampler category contains the core execution engine for ACE-Step generation, heavily modified from the base implementation to explicitly support conditioning masking, inpainting, and precise guidance scaling natively.

---

1. ScromfyAceStepSampler

File: nodes/sampler_node.py

The Scromfy-exclusive sampler node. While it functions similarly to standard ComfyUI KSampler under the hood, it intercepts the MODEL and our assembled CONDITIONING bundles to parse audio_codes, timbre_tensor, and lyrics_tensor mathematically before executing semantic diffusion.

Crucially, it is natively aware of audio_mask attributes attached to the conditioning objects, allowing for perfect audio inpainting out-of-the-box (e.g. replacing seconds 15-20 of a song while freezing the rest).

Inputs

• model (Required, MODEL): Base DiT model from CheckpointLoader.
• positive (Required, CONDITIONING): Full positive tensor bundle.
• negative (Optional, CONDITIONING): Negative prompt bundle (usually zeroed-out).
• audio_codes (Optional, LIST): Explicitly pass 5Hz tokens instead of attaching them to the conditioning bundle.
• latent_image (Optional, LATENT): Optional input latents for Img2Img or continuation geometry.

Options

• seed
• steps: Number of diffusion steps (typically 50-100 for audio).
• cfg: Classifier-free guidance.
• sampler_name: Dropdown of K-Diffusion samplers.
• scheduler: Dropdown of noise scale schedules.
• denoise: Strength of noise added to latent_image if present. Default 1.0.

Outputs

• samples (LATENT): The raw generated latents, ready to be passed to an Audio VAE Decode node.

---

2. ScromfySamplerSettings

File: nodes/sampler_settings_node.py

A standalone container for advanced sampling parameters. Allows for cleaner workflows by separating model logic from noise schedule tuning.

• Options: omega (ERG intensity), shift (noise schedule bias), custom_timesteps (overrides steps), denoise.
• Outputs: sampler_settings (SCROMFY_SAMPLER_SETTINGS).