LocalForge AILocalForge AI
LibraryBlogFAQ

How to Install and Use LoRAs

Here's the answer you came for: LoRA files go in models/Lora (A1111 and Forge) or ComfyUI/models/loras, and in A1111/Forge they only fire when your prompt contains the <lora:filename:weight> tag — the trigger word alone does nothing. The rest of this guide covers the full download-to-verified workflow plus the three silent failures that make LoRAs "not work." First-time setup takes 10–15 minutes.

What You Need

  • A working UI: Forge, ComfyUI, or A1111 — LoRA support is built into all three, no extension needed.
  • A base checkpoint that matches the LoRA's base model. This matters more than anything else on this page.
  • GPU: whatever runs your checkpoint — LoRAs add almost no VRAM on SD 1.5/SDXL. Flux LoRAs want 12 GB+, and Forge's LoRA patching is a known pain on 4–8 GB cards.
  • Disk: LoRA files are small — typically 10–150 MB, negligible next to multi-GB checkpoints.

Step 1 — Check the Base Model, Then Download

On the Civitai model card, look at the Base Model field in the right sidebar — "SD 1.5", "SDXL 1.0", "Pony", "Illustrious", "Flux.1 D" — before you click Download. It must match your checkpoint's family. A mismatched LoRA either errors out or, worse, loads fine and silently does nothing.

You'll get a .safetensors file in your Downloads folder.

Step 2 — Copy the Trigger Words and Recommended Weight

Same page: copy the Trigger Words shown on the card and skim the description for the creator's recommended weight. Character and concept LoRAs barely work without their triggers; style LoRAs usually don't need one. If the page lists no weight, 0.6–0.8 is the standard starting range per Civitai's own FAQ.

Step 3 — Move the File to the Right Folder

  • A1111: stable-diffusion-webui/models/Lora.
  • Forge: webui/models/Lora — fresh installs sometimes don't have the folder; just create Lora inside models. You can also point Forge at an existing A1111 library with --lora-dir in webui-user.bat.
  • ComfyUI: ComfyUI/models/loras.

Subfolders (like Lora/SDXL/ or Lora/Pony/) work in all three and show up as filter sections. The classic mistake: dropping the file in models/Stable-diffusion — that's the checkpoint folder, and the LoRA won't appear.

Step 4 — Refresh the UI

A1111/Forge: open the Lora tab under the prompt box and click the refresh icon. ComfyUI: click Refresh or restart so the file appears in the node dropdown.

Don't see it? A1111 and Forge hide LoRAs they think are incompatible with the loaded checkpoint. Either load a matching checkpoint, or enable Settings > Extra Networks > Always show all networks on the Lora page, apply, and refresh again.

Step 5 — Add It to Your Prompt (A1111/Forge)

Click the LoRA's card — it inserts the tag for you — or type it manually:

a samurai, <lora:my_style:0.8>, cinematic

Filename without .safetensors, weight after the colon. The tag is what loads the LoRA; the trigger word activates the concept. Typing only the trigger word without <lora:> loads nothing — it just gets tokenized as a regular prompt. That's the #1 "my LoRA doesn't work" cause. One more wrinkle: the tag name can be the file's internal alias rather than the filename, so clicking the card beats typing.

Step 6 — Wire It in ComfyUI

Add a Load LoRA (LoraLoader) node between Load Checkpoint and the KSampler:

  • Checkpoint MODEL → LoraLoader model → KSampler.
  • Checkpoint CLIP → LoraLoader clip → both CLIP Text Encode nodes.

Pick the file in lora_name, set strength_model and strength_clip (defaults 1.0). Trigger words still go in the prompt node. Stacking: chain LoraLoader nodes in series, or install rgthree's Power Lora Loader — one node, any number of LoRAs, per-row toggles and strengths.

Step 7 — Set the Weight and Add Triggers

Start at the creator's recommended weight. The default :1 is "usually too high" per Civitai's own wiki — most LoRAs behave between 0.5 and 1.0. Stacking rules:

  • Give the star the highest weight: <lora:style_a:0.6>, <lora:character_b:0.7>, <lora:concept_c:0.4>
  • Drop every weight when stacking: full-strength LoRAs fight each other and fry the image — 0.5–0.7 each, added one at a time
  • Negative weights invert the effect: slider LoRAs are built for this — Detail Tweaker runs from -2 to 2, with negative weight reducing detail

Verify It Works

Fix the seed, generate once with the LoRA tag and once without, and compare. In A1111/Forge, check the generation info under the image: a loaded LoRA shows its name and hash in the metadata. Lora not found: <name> means wrong folder or wrong name. In ComfyUI, bypass the LoraLoader node for the A/B test. No visible difference with no error? Almost certainly a base-model mismatch.

Troubleshooting

  • LoRA has no effect: your prompt has the trigger word but no <lora:> tag — the LoRA never loaded. Second suspect: base-model mismatch. The four SDXL lineages (base SDXL, Animagine, Pony, Illustrious) cross-load without errors but the effect silently fails.
  • LoRA missing from the Lora tab: the compatibility filter is hiding it — enable "Always show all networks" in Settings > Extra Networks. Duplicate filenames across subfolders also make entries vanish; rename so every file is unique.
  • Pony output looks plastic or muddy: clip skip must be 2 on Pony V6. A1111/Forge: set CLIP_stop_at_last_layers to 2. ComfyUI: add a CLIPSetLastLayer node at -2 feeding both text encoders.
  • Forge hangs on "Patching LoRAs" (Flux, low VRAM): known issue on 4–8 GB cards. Set Diffusion in Low Bits to Automatic (fp16 LoRA) to skip the heavy patching pass — confirmed working on 6 GB, at some quality cost with NF4 checkpoints.
  • LoRAs broke after an A1111 update (TypeError at startup): lora_in_memory_limit got set to null in config.json — change it to 0 or delete config.json and let the UI regenerate it.

What to Do Next

  • Need LoRAs worth installing? Best NSFW LoRA models — the cross-architecture hub, organized by base model.
  • On an SDXL stack? Best SDXL NSFW LoRAs — triggers and weights included for every pick.
  • Skip the setup entirely: LocalForge AI ships Forge pre-configured — drop .safetensors files in and they show up, no folder hunting.

FAQ

Where do I put LoRA files in Stable Diffusion? +
A1111 and Forge: models/Lora. ComfyUI: ComfyUI/models/loras. Subfolders work in all three. If Forge has no Lora folder, create it manually inside models.
Why is my LoRA not doing anything? +
Two causes cover almost every case: the prompt has the trigger word but no <lora:name:weight> tag (the LoRA never loads), or the LoRA's base model doesn't match your checkpoint — mismatches fail silently with no error.
Do I need the trigger word AND the lora tag? +
In A1111/Forge, yes for most character and concept LoRAs: the <lora:> tag loads the weights, the trigger word activates the concept. Style LoRAs often work without triggers.
What LoRA weight should I use? +
Start at the creator's recommendation, or 0.6-0.8 if none is given. The default :1 is usually too high. When stacking, drop each LoRA to 0.5-0.7.
Can I use an SD 1.5 LoRA on SDXL or Flux? +
No — SD 1.5, SDXL, and Flux LoRAs are architecturally incompatible. Within SDXL, the Pony and Illustrious branches don't cross-load cleanly either.
How do I know the LoRA is actually loaded? +
Generate with a fixed seed, then check the generation info under the image — a loaded LoRA shows its hash in the metadata. 'Lora not found' means wrong folder or wrong name.