Automatic CFG - Preset Loader

If the plain Automatic CFG node is the one dial you actually touch, the Preset Loader is the whole control room. It loads one of 26 shipped .json presets - each a full configuration for the same underlying patch engine - so you get attention tweaks, negative handling, and automatic CFG rescaling all from a single dropdown instead of a wall of toggles. Most are faster than the original node, some are slower, and a few are genuinely weird. The default, Excellent_attention, is the same preset the standalone Excellent attention node applies.

The honest framing: this is the node to reach for once the plain Automatic CFG has sold you on the idea and you want the curated stuff - attention-layer mutes, custom attention modifications, boosted negatives - without opening the Advanced node.

How it works

A preset is just a JSON file of parameters, saved into the pack's presets/ folder (that's also where the Advanced node's "save as preset" writes). The loader reads one, lets you override a couple of top-level things, and hands the whole dict to the same advancedDynamicCFG.patch() engine the other nodes wrap. The two STRING outputs let you see exactly what got loaded: Preset name and a Parameters as string you can dump into a text-display node for debugging.

The inputs that matter

• preset - the dropdown. 26 shipped choices. Highlights from the README's own notes: Crossed conds customized 3 is "amazing" on SD 1.5; Reinforced_style gives your negative prompt much more weight and works at 12 steps (slightly slower); The red riding latent is SDXL-only, an almost nonsensical mix that tends to dress characters in red; Potato Attention Guidance is great for portraits of happy people. Most of these were only really tested on SDXL.
• automatic_cfg - From preset (default), or override to None / soft / hard / hard_squared / range. Handy when a preset is all about attention and you want to switch its CFG-rescaling behavior.
• uncond_sigma_end with use_uncond_sigma_end_from_preset - override where in the sigma range the negative pass turns off. Leave the toggle on unless the preset's own value is misbehaving.

Optional join_global_parameters takes an ATTNMOD from the attention-modifier nodes and layers it on top of whatever the preset does - the main way you'd extend a preset without editing JSON.

The eval() warning is not a formality

Presets are JSON, but the strings inside get interpreted with Python's eval() at sampling time. The README is explicit: only load presets you trust, because a shared preset can execute arbitrary code. The author mentions he almost shipped a joke preset that would shut your computer down in 60 seconds to make the point. Don't grab random preset files from strangers; if you must, read the JSON first.

Also: don't delete Excellent_attention.json (or do_not_delete.json) - the loader lists files from the folder at startup, and several nodes assume those exist. If the preset list looks short or a node errors, check the folder is intact.

Install and use

Same pack as everything else here - ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Extraltodeus/ComfyUI-AutomaticCFG

Restart ComfyUI. Wire the Model output into your KSampler, set CFG to 8 on the sampler (the reference the presets assume), and start swapping presets with a fixed seed if you want honest before/afters. The loader is the best on-ramp to this pack's attention experiments - just keep the eval warning in mind and treat strangers' presets like executable files.