⚙️ Advanced CFG Controller (Expert)

The Expert node is the engine room of the Advanced CFG Controller pack. Where the Easy Mode node hides everything behind presets, this one hands you fifty-plus parameters and trusts you not to break anything. Same underlying machinery - it clones your model, swaps in its own step-by-step CFG function, and adds hooks around the conditional and unconditional passes - but now every dial is exposed, and the sigma windows are the raw values instead of friendly "from step / until step" numbers.

Before you commit an afternoon to it, the same caveat as the Easy version: this pack is built for classic DDPM models (SD 1.5, SDXL, and their fine-tunes) that run real CFG. On a guidance-distilled model that lives at CFG 1 there's no uncond pass to amplify or skip, so the whole premise collapses. It's also a fresh, single-commit pack from FlareAI-Studios - no community track record, so treat behavior as subject to change.

The sections, grouped

The inputs sort into families, and knowing the family beats memorizing names:

• Adaptive CFG - adaptive_cfg (Disabled / Soft Intensity / Hard Intensity / Hard Squared / Range Based / Spectral Analysis), adaptive_cfg_topk, and adaptive_cfg_ref. This measures the latent's intensity each step and steers the effective scale toward adaptive_cfg_ref (default 8). Lower adaptive_cfg_topk = finer measurement, slower.
• Skip the negative - skip_uncond plus uncond_sigma_start/end controls the window where the unconditional pass is skipped (your speed boost). fake_uncond_start and the fake_uncond_exp* family decide what substitutes for the real negative: fake_uncond_exp_method offers a huge menu (Use Positive, Zero Tensor, Sine Wave, Amplify, even "The Darkness"), fake_uncond_exp_value its strength, fake_uncond_multiplier flips it to -1, 0, or 1.
• Blending - lerp_uncond blends positive and negative predictions together within a sigma window; lerp_uncond_strength is how hard.
• Post-CFG color - subtract_latent_mean fixes color drift by centering latent values, latent_intensity_rescale rescales intensity against latent_intensity_rescale_cfg (8 = the reference). Both have sigma start/end windows.
• Prompt shaping - cond_exp / uncond_exp apply math transforms (amplify, root power, error function, sine…) to the positive or negative conditioning, each with its own method dropdown and value.
• Wildcard stuff - disable_cond (turn the positive prompt off inside a window), adaptive_noise_strength, spectral_norm_target, frequency_filter_cutoff, dynamic_range_compression, enable_performance_monitoring.

The attention modifier inputs

This is what makes the Expert node the center of the pack. Four optional inputs - attention_modifiers_positive, attention_modifiers_negative, attention_modifiers_fake_negative, and attention_modifiers_global - accept the pack's custom ATTNMOD wire type, produced by its Attention Modifier nodes. Those plug in as Python-expression attention replacements at specific UNet blocks, active only within sigma ranges. This is genuinely fiddly, experimental territory; start with the positive slot and one modifier, not all four.

Outputs are the patched MODEL and a STRING - a dump of every parameter you passed. Because the STRING output carries the full settings, you can wire it into a text node to save a "recipe" of a working config. args_filter (optional, multiline) trims which parameters appear in that string.

Install and where it lives

Same pack, same install: Manager → search Advanced CFG Controller, or clone into custom_nodes and restart:

cd ComfyUI/custom_nodes
git clone https://github.com/FlareAI-Studios/advanced-cfg-controller

No model downloads, no heavy deps (torch/numpy/colorama only). It lives under model_patches/Advanced_CFG_Controller, with the attention modifiers and utilities in subfolders.

Practical advice

The sigma defaults (start 1000, end 1) are sized for a full-range SDXL sampler: the 1000 ceiling never actually triggers, and the 1 floor means the negative pass is skipped only in the final fine-detail steps (sigma < 1) - a gentle default, not the drastic speed-up the toggles promise. If you want the real speed boost, widen that window toward your sampler's middle sigmas. Don't touch everything at once - the pack's own monkeypatch of ComfyUI's global sampling function means mistakes can bleed into other nodes, and a restart clears it. My honest recommendation: get comfortable with adaptive_cfg, skip_uncond, and the two color-fix toggles first. If you still need fake_uncond_exp_method at "The Darkness," you've gone off the deep end - and that's fine, just save the settings string so you can come back.