ComfyUI Node
MD: ACE-Step XL Latent Processor 🎛️
🎛️ MD: ACE-Step XL Latent Processor — Post-sampling DC correction. Place between KSampler and VAE Decode. latent_shift re-centers the distribution before decode. Outputs corrected LATENT + stats STRING.
MD: ACE-Step XL Latent Processor 🎛️
- samples
- samples
- stats
◄latent_shift-0.15►
◄scale1.00►
◄mean_normalizefalse►
◄passthroughfalse►
CategoryMD_Nodes/Loaders
Inputs (5)
| Name | Type | Default | Description |
|---|---|---|---|
| samples | LATENT | Connect to KSampler LATENT output. | |
| latent_shift | FLOAT | -0.15-2–2 | Additive DC offset applied to all latent channels. Re-centers ACE-Step's non-zero silence_latent. 0.0 = passthrough (disabled) < 0.0 = shift toward center (typical — reduces DC bias) > 0.0 = shift away from center Per-variant recommendations: xl-turbo : -0.12 xl-sft : -0.17 xl-base : -0.15 Tune ±0.05 by ear from the variant default. Too negative = thin/quiet. Too positive = distortion. |
| scale | FLOAT | 1.000.1–3 | Multiplicative output level. Applied after shift. 1.0 = no change (start here). >1.0 = amplify (clipping risk at high values). <1.0 = attenuate (safe headroom). Use sparingly — adjust latent_shift first. |
| mean_normalize | BOOLEAN | false | Subtract per-channel mean before applying latent_shift. More aggressive DC correction than shift alone. Enable if latent_shift doesn't fully resolve DC artifacts. Runs before latent_shift. |
| passthroughopt | BOOLEAN | false | True = bypass all ops. Use for A/B comparison. |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| samples | LATENT | — |
| stats | STRING | — |