.ai

the best open source ai models

Guides

Rankings

FAQ

Upload

How to install a LoRA: folders, file checks and first use in ComfyUI, Forge, Automatic1111 and Draw Things

Q: Loads, but does nothing

Trigger word missing or strength 0. Add the trigger from the model page, set strength 0.8.

Topic: Apps

By Lookout

Published 2025-09-24

Overview

Installing a

LoRA

is copying one file into one folder, but the details trip people up every day: the folder differs per UI, the file must match the base model family, downloads can be corrupted or renamed, and the UI must be told to rescan. This guide is only the installation; what a LoRA is and how weights and trigger words work is covered in the LoRA explainer linked below.

The file you want is a .safetensors file of between 10 and 700 MB. Its model page on this site shows the family (

SDXL

Flux

Wan

and so on), the size, the SHA-256 hash and the links to where the author published it. Download from the author's page; mirrors exist but hashes are how you know a mirror served the real file. The

safetensors

format cannot contain code, which is why every current UI prefers it; old .pt and .ckpt LoRAs should be avoided unless you trust the source.

If you would rather not manage files at all, the same LoRAs are preinstalled on the two cloud partners linked from every model page: on

PirateDiffusion

a LoRA is a tag inside a chat command, on

BitVector

it is a selection in the web app. Many people use the cloud to test a LoRA and only install locally the ones they keep using.

Reference

Name	Type	What it is
ComfyUI	folder	ComfyUI/models/loras/ (subfolders allowed and shown in the picker). Load with the Load LoRA node, or LoraLoaderModelOnly for Flux-style separate loaders.
Forge / reForge	folder	webui/models/Lora/. Appears in the Lora tab under the prompt; clicking inserts <lora:name:1>.
Automatic1111	folder	stable-diffusion-webui/models/Lora/. Same tab and syntax as Forge ; LyCORIS files needed an extension before 1.5.
Fooocus	folder	Fooocus/models/loras/. Picked in the Advanced > Model tab with a weight slider, up to five at once.
SD.Next	folder	models/Lora/ by default; settings allow a shared models directory with other UIs.
Draw Things (Mac, iPhone)	import	Models > LoRA > Import from file or URL; the app converts the file to its own format on import.
SHA-256	check	The hash on the model page. Windows: certutil -hashfile file.safetensors SHA256. Mac and Linux: shasum -a 256 file.safetensors.
Shared folder	setting	ComfyUI extra_model_paths.yaml, Forge --lora-dir, lets several UIs read one LoRA folder instead of three copies.

Skip the install: this LoRA is already in BitVector

BitVector Prism has the LoRAs from this site preinstalled. Pick the model, pick the LoRA, type the prompt. Test it in a minute and only download the ones you keep using.

Try BitVector

Step by step

Open the LoRA on this site and note three things: the family, the trigger words and the hash. Click the source link to the author's page.
Download the .safetensors file. If the page offers several versions (v1, v2, different base models), take the one whose family matches the
checkpoint
you use.
Check the hash: run certutil -hashfile (Windows) or shasum -a 256 (Mac, Linux) on the file and compare the first eight characters with the model page. A mismatch means a corrupted or substituted download; delete it and download again.
Rename if you like (the file name is what you will type in the prompt on Forge and A1111), keep the .safetensors extension, and move the file into the LoRA folder of your UI from the table above.
Refresh: ComfyUI needs the refresh button in the node picker or a page reload; Forge and A1111 have a refresh icon in the Lora tab; Fooocus rescans on tab open; Draw Things imports instead of scanning.
First use in ComfyUI: Load Checkpoint -> Load LoRA (MODEL and CLIP through it) -> your usual CLIP Text Encode and KSampler. Set both strengths to 0.8 and pick the file.
First use in Forge or A1111: click the LoRA card so <lora:filename:1> lands in the prompt, change 1 to 0.8, add the trigger word before your subject.
Generate once with the example prompt from the model page to confirm the LoRA works, then write your own prompts.
Optional: put a preview image next to the file with the same name (name.preview.png in Forge, name.png in ComfyUI managers) so you recognise it in the picker later.

Examples

Checking the hash on Windows and Mac

certutil -hashfile "C:\ComfyUI\models\loras\sdxl\zavy-cinematic-xl.safetensors" SHA256

shasum -a 256 ~/ComfyUI/models/loras/sdxl/zavy-cinematic-xl.safetensors

First use in Forge

zavy cinematic, <lora:zavy-cinematic-xl:0.8> a lighthouse keeper on a stormy cliff, dramatic light, 35mm

Negative: blurry, watermark | 1024x1024 | DPM++ 2M Karras | 28 steps | CFG 6

Same LoRA on the cloud, nothing installed

/render <juggernaut-xl> <zavy-cinematic-xl:0.8> zavy cinematic, a lighthouse keeper on a stormy cliff, 35mm /size:1024x1024

Tips

Keep one LoRA folder per family (loras/sdxl, loras/flux, loras/wan). Every UI shows subfolders and you will stop loading Flux LoRAs on SDXL by mistake.
Civitai files are often named by version number only. Rename them to author-concept-version so the prompt tag means something.
A LoRA larger than 1 GB is usually a full fine-tune mislabelled, or a video LoRA for a 14B model. Check the model page.
On Windows, Defender sometimes quarantines large downloads mid-way; the hash check catches the truncated file.
Flux LoRAs trained with ai-toolkit load fine in ComfyUI but may show "lora key not loaded" warnings in older Forge builds; update Forge.
For Wan and
Hunyuan
video LoRAs, the model page names the exact base (
Wan 2.2
14B high-noise, for example); install only that one.

Point them at one folder with extra_model_paths.yaml (ComfyUI) and --lora-dir (Forge, A1111).

LoRAs as chat tags, nothing to download

On PirateDiffusion a LoRA is a tag in a Telegram command: /render <base> <lora:0.8> your prompt. Thousands are installed, and you can upload your own. Unlimited generation on a fixed monthly price.

Try PirateDiffusion

Questions

Can I install a LoRA on the cloud?

PirateDiffusion and BitVector already have the LoRAs listed on this site installed; a LoRA that is missing can be uploaded on PirateDiffusion by its users. Nothing to install on your side.

Can I use the same file in ComfyUI and Forge?

Yes, it is the same format. Share the folder rather than copying the file.

Does the file name matter?

Only as the tag you type in Forge and A1111. ComfyUI shows a picker. Keep it short and without spaces.

Where do I find older versions?

The model page here lists each version as its own entry with its own hash; the author's page usually keeps the files.

Is a LoRA from an unknown website safe?

A .safetensors file cannot execute code, so the risk is a different or lower-quality model, not malware. Verify the hash against the model page; avoid .pt, .ckpt and .pickle files from strangers.