FunUI LoRA Loader
One LoRA Per Node, and It Tells You When One Did Nothing
- model
- clip
- model
- clip
- status
The most useful thing about FunUI's LoRA loader isn't that it loads LoRAs. It's that it complains. Load a Flux LoRA onto an H3 model with any other loader and you get a picture with no LoRA applied and no error, then twenty minutes of wondering why the style didn't land. This one checks whether the weights match anything in the model in front of it, and says so out loud.
What it actually does
One LoRA per node. That's a deliberate regression from FunUI 4, where N LoRAs lived in a single node behind a JSON widget; the source argues the case and it's hard to disagree - chaining is how ComfyUI's own loader works, every strength is a real socket, and the graph shows you how many LoRAs are actually in play instead of hiding the count in a string.
Mechanically: it loads the file, matches its keys against the model (and the CLIP if you wired one), and applies the resulting patches to a clone. Three worth-knowing behaviours come straight from the code:
- Picking None returns your model and clip untouched - the node stays in the pipeline so a LoRA is one dropdown away.
- Both strengths at 0 is a valid request, so it does nothing and raises a log alert, because "applied a LoRA at 0" and "forgot to set the strength" look identical in the output image.
- If nothing matched, or some weights didn't fit the model's shapes, you get an alert with counts - "matched NOTHING in this model" or "17 of 240 weights do not fit this model's shapes". That second one is the classic symptom of grabbing the variant-trained LoRA for a differently-shaped base.
Inputs and outputs
Set these:
- model - what you're patching.
- lora_name - a combo from
ComfyUI/models/loras, withNoneas the default. - strength_model - −100 to 100, default 1.0. Negative inverts the LoRA, which is how you subtract a style without needing a "negative" file.
Optional, and worth understanding:
- clip - wire it when the LoRA also moves the text encoder. Leave it unwired for a model-only LoRA; you don't have to fake a connection.
- strength_clip - only used when a clip is wired. Often this wants to be lower than the model strength.
Outputs are model, clip and status. The model goes to your sampler or the next loader in the chain; the clip output is the patched encoder (and is None if you wired no clip, so don't feed it into an encoder node in that case). The status string carries the same detail as the log: name, both strengths, how many weights matched, how many were dropped.
Installing it
Part of the FunUI pack:
cd ComfyUI/custom_nodes
git clone https://github.com/digital-garbage/FunUI
cd FunUI && pip install -r requirements.txt
Or ComfyUI Manager → search FunUI. Restart, then drop .safetensors LoRAs in models/loras. Requirements are light (comfy-kitchen, gguf); don't let pip touch torch.
The pack page still reads olivv-cs/ComfyUI-FunPack while the repo is now digital-garbage/FunUI - old name, same project.
Where people get burned
Strength 1.0 on everything. The default is usually too much. Try 0.5–0.8; if the look is right but the image is crunchy, that's your number, not the LoRA.
Missing trigger words. No loader can fix this. If the LoRA page says it needs a trigger phrase, put it in the prompt - or use a shortcut/lorebook in FunUI's Composer if you're working inside the studio.
Wrong family, silently. A Pony or Illustrious LoRA is selectable on an SDXL model and often just doesn't work, because they share architecture but not training. FunUI's shape-mismatch alert catches the crude version of this; it can't catch a soft no-op.
Chain order is not cosmetic. Patches stack in order, and the common advice - heavier, structure-changing LoRAs first, lighter style ones later - still holds. Chaining five 1.0s also argues for lowering each one.
A subtle one the source calls out: at strength 0 the node returns the same objects it was handed, not clones. That's deliberate, so a later loader in a chain doesn't lose patches. It also means a node you think is inert might not be where the real blame lies if something downstream looks wrong - read the status line, it's specific about what it did.
Inputs (5)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| lora_name | COMBO | None | None leaves the model untouched. |
| strength_model | FLOAT | 1.00-100–100 | How strongly the LoRA changes the model. 0 = not at all, negative = the opposite. |
| clipopt | CLIP | Leave unwired for a model-only LoRA. | |
| strength_clipopt | FLOAT | 1.00-100–100 | How strongly the LoRA changes the text encoder. Only used when a clip is wired. |
Outputs (3)
| Name | Type | Description |
|---|---|---|
| model | MODEL | — |
| clip | CLIP | — |
| status | STRING | — |