Nodes/MD Nodes/MD: ACE-Step XL Loader 🎡
ComfyUI Node

MD: ACE-Step XL Loader 🎡

🎡 MD: ACE-Step 1.5 XL Loader β€” Loads XL variants (hidden_size=2560). Tunable AuraFlow shift baked in β€” no external node needed. Outputs MODEL + sigma plot IMAGE + analytics STRING. Session cache skips 90s reload on re-queue.

By MDMAchineΒ·Created about a year agoΒ·Updated 2 months agoΒ· 14
MD: ACE-Step XL Loader 🎡
    • model
    • sigma_plot
    • analytics
    β—„model_path[place XL model folder path here]β–Ί
    β—„variantxl-turboβ–Ί
    β—„aura_shift3.00β–Ί
    β—„devicecudaβ–Ί
    CategoryMD_Nodes/Loaders

    Inputs (4)

    NameTypeDefaultDescription
    model_pathSTRING[place XL model folder path here]Full path to the XL model directory. Must contain: config.json, modeling_*.py, silence_latent.pt, and weights (.safetensors). Session cache: same path+variant skips the 90s reload on re-queue.
    variantCOMBOxl-turboxl-turbo : DMD2 distilled, 4–8 steps (fastest) xl-sft : 50-step undistilled, broader manifold (best quality) xl-base : Unaligned research weights
    aura_shiftFLOAT3.000.1–10AuraFlow sigma shift β€” controls denoising schedule curvature. Formula: Οƒ(t) = t / (t + shift*(1βˆ’t)), multiplier=1.0 shift = 1.0 β†’ linear schedule shift = 1.73 β†’ AuraFlow default (√3) shift = 3.0 β†’ ACE-Step XL validated (ByteDance default) shift > 1.0 β†’ front-loads denoising (more aggressive early) Per-variant recommendations: xl-turbo : 3.0 (validated) xl-sft : 2.5 (start here, tune down for coherence) xl-base : 1.73 (AuraFlow default as baseline) No external ModelSamplingAuraFlow node needed β€” this loader bakes the shift in directly.
    deviceCOMBOcuda2 options: cuda, cpu

    Outputs (3)

    NameTypeDescription
    modelMODELβ€”
    sigma_plotIMAGEβ€”
    analyticsSTRINGβ€”