Nodes/FunUI/FunUI Load Modifiers
ComfyUI Node

FunUI Load Modifiers

Turning On FunUI's Sampler Tricks Without Changing Your Sampler

By digital-garbage·Created about a year ago·Updated a day ago· 21
FunUI Load Modifiers
  • model
  • settings
  • model
  • status

FunUI ships a pile of sampling techniques - the papers cited in the README are a reading list of 2025–26 tricks: ALG, FreeSliders, Decisiveness, momentum, REINS, STAS. Here's the design decision that makes them usable: none of them are inputs on the sampler. They attach to the model, and this node is where they attach.

Which means something slightly rude and quite useful: you can wire this node's model output into ComfyUI's own SamplerCustomAdvanced and every per-step trick still runs. FunUI's sampler is not required, so no sampler can own the features.

What it actually does

It clones the model, strips any wrappers this node installed previously, then asks the registry which installed modules offer a "modifier" capability and whether this particular model can accept them - the check reads traits off the model itself rather than asking you which architecture you're on. Compatible ones install in dependency order through ComfyUI's own patcher_extension wrappers on the clone.

Two details from the source earn their keep. First, installing on a clone means the model everyone else is holding is untouched; ModelPatcher.clone() copies wrapper lists per key. Second, every install is preceded by a strip of the node's own keys, because add_wrapper_with_key appends - running a model through this node twice (two passes, a flattened subgraph, anything chained) would otherwise install the same wrapper twice and run it at double strength per step while reporting "1 modifier applied".

Modifiers run guarded: a hook that raises at step 3 is dropped for the rest of the run and noted, rather than killing a rental that's already burned the GPU time. There's an unguarded mode too - "full control" - which you should only reach for if you're debugging a modifier and want its exception, not its silence.

Inputs and outputs

  • model - whatever your checkpoint or diffusion-model loader gave you. Wire this after LoRAs, since the modifiers want to see the patched model.
  • settings - optional, a FUNUI_SETTINGS object from FunUI Modifier Settings. This is where your chosen values and the on/off switches travel as one JSON blob. Leave it unwired and every compatible modifier installs with its declared defaults and nothing is switched off; that's a perfectly good way to find out what a modifier even does, and a terrible way to ship a workflow you can't reproduce.

Outputs:

  • model → your sampler (FunUI's or core's), then decode.
  • status → what got applied, what got dropped, and why. Multi-line, blunt. If a module was quarantined after an earlier failure, this says so and tells you to turn it back on in Settings ▸ Modules or restart.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/digital-garbage/FunUI
cd FunUI && pip install -r requirements.txt

Manager: search FunUI. Restart. (Pack page says olivv-cs/ComfyUI-FunPack; repo is now digital-garbage/FunUI. Same project.)

Do look at the requirements file if you plan to use the fancier attention tricks: comfy-kitchen>=0.2.32 is a floor for a compiled kernel, and the pack deliberately reports a low version rather than upgrading it for you, because it's built against your torch. Installing a new comfy-kitchen yourself can mean a rebuild, not a pip line.

Where people get burned

Assuming the node names the features. It doesn't - that's the architecture. Which modifiers exist depends on what's installed and enabled in Settings ▸ Modules, and which apply depends on your model. Read status before assuming a technique is on.

A modifier that "worked at step 0". The guarded path exists because failures partway through sampling are the expensive kind. If output suddenly degrades after enabling something, turn it off in Settings and re-run - the log says which module was dropped and when.

Double wiring. Two Load Modifiers nodes in one graph is fine (each strips before it installs), but stacking them is also pointless. One per branch.

Expecting settings to travel. A FUNUI_SETTINGS object from one project's Modifier Settings node, reused in a different workflow, may reference modules that aren't loaded - the pack logs a warning and ignores those rather than failing. Harmless, but it's a hint your saved workflow assumed more modules than your install has.

Gating on speed. Every one of these is per-step work on the model. "Free" is a word people use about these techniques; the wall clock disagrees. Benchmark one modifier at a time on a short clip.

CategoryFunUI/Sampling

Inputs (2)

NameTypeDefaultDescription
modelMODEL—
settingsoptFUNUI_SETTINGS—

Outputs (2)

NameTypeDescription
modelMODEL—
statusSTRING—