Skip to content
,

ComfyUI LoRA Not Working? Every Fix (Incl. “lora key not loaded”)

9–13 minutes
Woman artist holding a card reading LORA, hook LoRA not working fixed

You downloaded a LoRA, added it to your ComfyUI workflow, hit Queue, and the image looks exactly the same as before. Or the console fills up with hundreds of lines that say lora key not loaded. Both problems are very common, and almost always come down to a handful of causes.

This guide walks through every reason a LoRA does nothing (or breaks the image) in ComfyUI, in the order you should check them: the file, the wiring, the base model, strength, trigger words and the warnings in the console. Each fix takes a minute or two.

TL;DR

  • Not in the list? Put the file in ComfyUI/models/loras/ and press R (refresh) in the browser.
  • No effect? The LoRA node’s MODEL and CLIP outputs must go to the KSampler and the text encode nodes. An unconnected LoRA node never runs.
  • Hundreds of “lora key not loaded” lines? The LoRA was made for a different base model (e.g. SDXL LoRA on SD 1.5 or Flux). Use a matching LoRA.
  • <lora:name:1> in the prompt does nothing in core ComfyUI. Use a Load LoRA node instead.
  • Still weak? Add the trigger word from the model page and try strength 0.8-1.2.

New to ComfyUI? My ComfyUI beginner guide covers install, models, nodes and fixes in one place.

Quick Checklist

Symptom Most likely cause Fix
LoRA not in the dropdown Wrong folder, or page not refreshed Move to models/loras, press R
Image unchanged, no warnings Node not connected, or strength 0 Rewire MODEL and CLIP through the LoRA
Image unchanged, many “lora key not loaded” LoRA for another base model Use a LoRA made for your checkpoint
A few “lora key not loaded” with model-only loader Text encoder part of the LoRA skipped Normal; use the full Load LoRA node if you want it
Effect is weak No trigger word, low strength Add trigger word, raise strength
Image is burnt, noisy or distorted Strength too high, or too many LoRAs Lower strength, remove LoRAs one by one
“WARNING SHAPE MISMATCH … WEIGHT NOT MERGED” Base model mismatch Match the LoRA to the model family

Fix 1: LoRA Missing from the List

ComfyUI only lists files from its LoRA folders:

  • Portable or manual install: ComfyUI/models/loras/
  • ComfyUI Desktop: the models/loras folder inside the base path you picked at install time
  • Any extra path set in extra_model_paths.yaml under loras:

Subfolders are fine; they show up as subfolder/name.safetensors in the dropdown.

After copying a new file, press R in the ComfyUI page (or use the refresh button) so the node lists reload. A full restart is not needed for LoRAs.

Also check the file itself:

  • It should end in .safetensors. A file that only half downloaded is often a few KB or has an odd extension like .part.
  • A real LoRA is usually 10 MB to 1 GB. A 6 GB file in your loras folder is probably a full checkpoint, which belongs in models/checkpoints.

The full folder map is in my ComfyUI model folders cheat sheet.

Fix 2: The LoRA Node Is Not Wired In

This is the number one cause when there are no warnings at all. ComfyUI only runs nodes that lead to an output. If the LoRA node sits on the canvas but its outputs go nowhere, it is silently skipped.

The correct chain for a normal checkpoint workflow:

  1. Load Checkpoint MODEL and CLIP go into Load LoRA.
  2. Load LoRA MODEL output goes to the KSampler model input.
  3. Load LoRA CLIP output goes to both CLIP Text Encode nodes (positive and negative).

Common mistakes:

  • KSampler still connected straight to the checkpoint’s MODEL, bypassing the LoRA.
  • The text encoders still take CLIP from the checkpoint, so the LoRA’s text side (and trigger word) is ignored.
  • The LoRA node is bypassed (purple) or muted (grey). Select it and press Ctrl+B or Ctrl+M to toggle back.
  • For models loaded with separate loaders (UNET/diffusion model loader plus a CLIP loader), the LoRA must sit between the diffusion model loader and the sampler, or the model sampling node if your workflow has one.

Fix 3: Wrong Base Model

A LoRA is a small patch for one specific model family. It only works on the family it was trained for:

LoRA made for Works with Does not work with
SD 1.5 SD 1.5 checkpoints and fine-tunes SDXL, Flux, SD3
SDXL SDXL, and usually Pony/Illustrious-style SDXL fine-tunes (results vary) SD 1.5, Flux
Flux.1 Dev Flux.1 Dev and most Dev fine-tunes, Schnell with mixed results SDXL, Flux.2, SD 3.5
Qwen-Image / edit models The same model and version Other image families
Wan, Hunyuan, LTX (video) The exact model and size it was trained on Other video models or sizes

The model page on Civitai or Hugging Face always lists the “Base Model”. If it does not match your checkpoint, the LoRA will either do nothing or flood the console with warnings. Check which model you actually loaded: the file name is shown in the Load Checkpoint or diffusion model loader node.

What “lora key not loaded” Really Means

Inside ComfyUI, every LoRA is a list of named weights (keys). ComfyUI builds a map of the keys your model has and loads each LoRA key that matches. Any key it cannot place is printed as:

lora key not loaded: lora_unet_input_blocks_4_1_proj_in.lora_down.weight

How to read it:

  • Hundreds or thousands of lines, covering unet, transformer or diffusion_model keys: the LoRA does not match your model at all. It is doing nothing. Fix the base model.
  • Only keys starting with lora_te, text_encoder or te_, and you use LoraLoaderModelOnly: that node does not touch CLIP, so the text encoder part of the LoRA is skipped and reported. The main effect still applies. Switch to the full Load LoRA node if you want the text part too.
  • Only text encoder keys with the full Load LoRA node: the LoRA was trained against a different text encoder (common with SDXL fine-tunes or Flux T5 parts). Usually the image side still works.
  • A handful of odd keys such as extra metadata: harmless.

A related warning, WARNING SHAPE MISMATCH ... WEIGHT NOT MERGED, means a key name matched but the sizes did not. That also points to a LoRA from a different model or model size.

Also update ComfyUI. Newer versions add key conversions for more LoRA formats (Kohya, Diffusers, XLabs, and newer model families), so a LoRA that failed on an old build may load fine after updating ComfyUI.

Fix 4: Prompt Syntax Does Not Load LoRAs

If you come from Automatic1111 or Forge, you may type this in your prompt:

a portrait of a woman, <lora:filmGrain:0.8>

Core ComfyUI does not read that syntax. The text is simply sent to the text encoder as words, and no LoRA is loaded. Use a Load LoRA node instead.

Some custom node packs add prompt-tag loading, but the built-in nodes do not. If you prefer that style, my Forge WebUI guide covers a tool that supports it natively.

Fix 5: Missing Trigger Word

Many character, style and concept LoRAs were trained with a special word, such as ohwx or pixelart style. Without it in the positive prompt, the effect is weak or missing.

  • Find it on the model page under “Trigger Words” or in the description.
  • Put it near the start of your positive prompt.
  • For Flux and other models with a T5 or LLM text encoder, a short natural sentence using the trigger word often works better than a list of tags.

The trigger word only works if CLIP is routed through the LoRA node (see Fix 2).

Fix 6: Strength Too Low or Too High

The Load LoRA node has two sliders:

  • strength_model: how much the LoRA changes the image model. This is the main one.
  • strength_clip: how much it changes the text encoder (helps with trigger words).

Both default to 1.0, and if both are 0 the node returns the model unchanged. Values can be negative, which pushes the image away from the LoRA’s style (handy for “detail” LoRAs).

Strength Typical result
0.3-0.6 Subtle style hint
0.7-1.0 Normal use for most LoRAs
1.0-1.5 Strong effect, risk of artifacts
Above 1.5 Usually burnt colours, noise or melted faces

Always follow the strength the creator recommends first. Speed LoRAs (Lightning, Turbo, Hyper and similar) also need the step count and CFG they were made for, often 4-8 steps and CFG around 1, or the image turns grey or noisy.

Fix 7: Stacking Several LoRAs

To use more than one LoRA, chain the nodes: checkpoint into LoRA 1, LoRA 1 outputs into LoRA 2, and LoRA 2 outputs into the sampler and text encoders. Every node in the chain must be connected.

Tips for stacks:

  • Lower each strength when stacking. Two LoRAs at 1.0 often fight each other; try 0.6-0.8 each.
  • Two style LoRAs, or two character LoRAs, rarely mix well.
  • If the result breaks, bypass LoRAs one at a time (Ctrl+B) to find the culprit.
  • Group the chain to keep the canvas tidy; my ComfyUI groups guide shows how.

Special Cases: GGUF, Flux, Wan and Video Models

  • GGUF models: the standard Load LoRA or LoraLoaderModelOnly node works with models loaded by the GGUF loader. Connect it after the GGUF UNET loader. See my ComfyUI GGUF guide.
  • Flux and other models with separate loaders: use LoraLoaderModelOnly after the diffusion model loader if the LoRA has no text encoder part. Expect some harmless text-encoder key messages if it does.
  • Wan 2.2 (high-noise and low-noise models): many Wan 2.2 LoRAs ship as a pair. Load the high-noise LoRA on the high-noise model branch and the low-noise LoRA on the low-noise branch. Putting one LoRA on only one branch gives a weak or uneven effect. My Wan 2.2 LoRA training guide explains the pair.
  • Custom speed loaders (for example Nunchaku or TensorRT engines): these often need their own LoRA loader node from the same pack. The core node may not affect them.
  • ControlNet workflows: the LoRA goes on the model, not the ControlNet. See my ControlNet guide.

Fix 8: The LoRA Breaks the Image

If the LoRA clearly works but ruins the picture:

  1. Lower strength_model to 0.6 and try again.
  2. Check the recommended sampler, steps, CFG and resolution on the model page. SDXL LoRAs at 512×512, or SD 1.5 LoRAs at 1024×1024, look bad.
  3. Remove other LoRAs to rule out a clash.
  4. If you get a fully black image, see my ComfyUI black image fix, which is usually a VAE or precision issue rather than the LoRA.
  5. Out of memory after adding LoRAs? Large LoRAs on top of big models can tip VRAM over. My CUDA out of memory fix and speed-up guide help.

Other errors are collected in my ComfyUI errors and fixes list.

FAQ

How do I know if my LoRA is actually working?

Keep the same seed, then queue once with the LoRA bypassed and once with it active. If both images are identical, it is not applied. Also check the console for “lora key not loaded” lines right after you queue.

Is “lora key not loaded” an error?

It is a warning, so the workflow keeps running. A few text-encoder lines are often fine. Hundreds of model lines mean the LoRA is not doing anything.

Can I use an SD 1.5 LoRA with SDXL or Flux?

No. The layer shapes are different. You need a version of the LoRA trained for your model.

What is the difference between Load LoRA and LoraLoaderModelOnly?

Load LoRA patches both the image model and CLIP. LoraLoaderModelOnly patches only the image model, which suits models whose LoRAs have no text encoder part, such as many Flux and video LoRAs.

Do I need to restart ComfyUI after adding a LoRA?

No. Press R to refresh the lists. Restart only after installing custom nodes.

Why does the same LoRA look different than on Civitai?

Different checkpoint, sampler, steps, CFG or resolution. Copy the example image’s settings, which are often listed under the image on the model page.

Sources

Free weekly email

Get new AI guides every week

One short weekly email with our newest local AI, ComfyUI and AI tool guides. No spam, no pop-ups, just the good stuff.

Free. Unsubscribe anytime in one click. We only use your email to send this newsletter. See our privacy policy.

Woman artist holding a card reading LORA, hook LoRA not working fixed