ACEStep XL Turbo and the Oobleck VAE produce a characteristic harmonic artifact present in all generations. The artifact has two independent origins:
VAE decoder bias — static reconstruction bias baked into the Oobleck decoder during its original training. Present at consistent amplitude across all generations, independent of length. Characterized via Ableton EQ Eight direct measurement:
Frequency
Depth
Character
74.4 Hz
-1.43 dB
Sub fundamental, BPM-dependent
94.0 Hz
-1.86 dB
Upper sub resonance
654–669 Hz
-12 to -15 dB
Primary hum, BPM shift ~2Hz/octave
211 Hz
~-7 dB
3rd harmonic
1090 Hz
~-7 dB
5th harmonic
DiT body resonance — accumulates over the denoising trajectory, perceptible after ~60 seconds. Hypothesized to arise from ODE manifold shearing in early denoising steps creating resonant modes in FFN activations of the decoder layers.
Both components require independent treatment. Post-processing filters (like the hum notch in MD Audio VAE Tiled) address the symptom. This regrind addresses the cause at the weight level.
Start with vae/acestep_1.5_vae_Regrind_V10b_blend50-BF16.gguf (or .safetensors for ComfyUI) — 50% blend between V10b regrind and the base decoder. Conservative, closest to the base model's original frequency response. Recommended for most material, especially anything spectrally dense or noisy.
Use vae/acestep_1.5_vae_Regrind_V10b-BF16.gguf for full-strength V10b suppression when the blend isn't aggressive enough.
V9b variants are the previous generation. Still valid if V10b causes any issues on your material.
V7 safetensors is the oldest format, kept for legacy loaders.
On raw ACE-Step XL Turbo without the regrinded base: load the LoRA at 0.50–1.00. No trigger words. Standard ACE-Step prompting.
Pair with MD Audio VAE Tiled
For long-form generation (>30s), pair with the MD Audio VAE Tiled decoder node for ComfyUI. Handles BPM-synced tiling, dual-pass seam elimination, and post-decode spectral cleanup (Wiener denoiser, harmonic preservation crossfade, hum notch stack).
The regrind addresses artifact suppression at the weight level. MD Audio VAE Tiled addresses the decode mechanics. Both together give the cleanest output.
These are derivative works of ACE-Step XL Turbo (Apache 2.0, ACE Studio / StepFun). Free for personal and research use. Commercial use requires a separate license — contact A&E Concepts.
Source code for associated ComfyUI nodes, C++ headers, and Lua plugins is licensed separately under GPL v3 in the repositories above.
Citation / acknowledgment
If you use these weights in your work, please credit:
