Wrong family, missing trigger word, or weight 0. Check the family on the model page, add the trigger word, set weight 0.8.

Trends

.ai

the best open source ai models

What is a LoRA and how to use one in ComfyUI, Forge, Automatic1111 and on the cloud

Q: Faces or textures look fried

Weight too high or several LoRAs stacked at full strength. Lower to 0.6, lower the others, drop the CFG by one.

Q: Style LoRA changes the character too

The CLIP part of the LoRA dominates the prompt. In ComfyUI lower strength_clip to 0.3 to 0.5; in A1111 use the lbw extension or lower the weight.

Topic: Models

By Captain

Published 2026-05-20

LoRAs explained for people who want results, not maths: what a LoRA file changes, why it must match the base model's family, trigger words, weights, stacking several LoRAs, the folder and syntax for each UI, and how to try a LoRA on the cloud before downloading anything.

Overview

A LoRA (Low-Rank Adaptation) is a small file, usually 20 to 400 MB, that teaches an existing model something new without retraining it: a face, a character, an art style, a piece of clothing, a camera look, a motion for video models. It works by adding a small correction to a subset of the model's weights at run time, so the same base

checkpoint

can become a hundred different specialists by loading a hundred different LoRAs. Most of the models on ModelTrends.ai are LoRAs, and the family label on each one names the base model it was trained for.

The two rules that matter: a LoRA only works with checkpoints of its own family (an

LoRA on SDXL, a

LoRA on

, a

LoRA on Wan 2.2 of the same size), and most LoRAs have a trigger word that must appear in the prompt to switch the learned concept on. The model page lists the family, the trigger words and the creator's recommended weight, taken from the source where the author published them.

You do not have to install anything to try one. The Run box on every model page opens the LoRA on

PirateDiffusion

, where it is used as <lora-name> inside a /render or workflow prompt, or on

BitVector

, where you pick it in the web app. Locally, every UI has a loader or a prompt syntax described below.

Reference

Name	Type	What it is
Family	label	The base model the LoRA was trained on: SD 1.5 , SDXL (incl. Pony , Illustrious, NoobAI), Flux, FLUX.2 , Qwen , Z-Image , Wan , Hunyuan . Must match the checkpoint.
Trigger word	prompt	A token the author trained the concept under (ohwx woman, zavy-style, ral-chrome). Put it in the prompt; without it many LoRAs do little.
Weight / strength	number	0.6 to 1.0 is normal. Above 1.2 artefacts appear; negative weights invert some sliders. Flux LoRAs saturate sooner than SDXL ones.
Model strength vs CLIP strength	ComfyUI	The Load LoRA node has two values: the Unet part (what it draws) and the text-encoder part (how it reads the prompt). Keep them equal unless a style bleeds into everything.
<lora:filename:0.8>	A1111 / Forge	Prompt syntax; the file name without extension. The Extra Networks tab inserts it for you.
Load LoRA node	ComfyUI	Sits between the checkpoint loader and the sampler / CLIP encode. Chain several nodes to stack LoRAs.
<lora-name:0.8>	PirateDiffusion	The trigger shown in the Run box, in angle brackets with an optional weight, inside /render or a workflow prompt.
Pony / Illustrious score tags	SDXL variants	LoRAs for Pony need score_9, score_8_up...; Illustrious / NoobAI LoRAs use masterpiece, best quality and artist tags . The family page explains.
LyCORIS / LoCon / DoRA	formats	Variants with different internal structure, loaded the same way in modern UIs. Older A1111 builds needed an extension.

No install required - run AI models on the cloud

BitVector Prism is the easiest way to start: pick a model, type a prompt and generate in a clean web app, with nothing to configure. BitVector is also available on Discord and on the web (SpyGlass).

Try BitVector

Step by step

Pick a LoRA on this site and read the model page: family, trigger words, recommended weight, example prompts under the renders, and the creator notes.
Make sure you have a checkpoint of the same family. The family page (linked in the breadcrumb) lists the base models and the hottest LoRAs together.
Download the .safetensors file into the LoRA folder: models/loras (
ComfyUI
), models/Lora (A1111 and
Forge
). Refresh the UI's lists.
ComfyUI: add Load LoRA after Load Checkpoint, wire MODEL and CLIP through it, pick the file, set strength 0.8 for both. For Flux with separate loaders use LoraLoaderModelOnly.
A1111 / Forge: open the Lora tab under the prompt, click the card, and the tag <lora:name:1> appears in the prompt; change the number to 0.8 and add the trigger word.
Write the prompt as the examples do: trigger word first, then the subject, then style words. Generate at the family's native resolution with the sampler and steps from the example renders.
Adjust: too weak, raise the weight to 1.0 or repeat the trigger word; too strong or deformed, lower to 0.6 and remove style words that fight the LoRA.
Stack: one character LoRA plus one style LoRA plus one detail LoRA is a normal limit; keep the sum of weights near 2.0 and lower each when adding more.
No GPU or no install: press the PirateDiffusion button in the Run box and send /render <trigger> your prompt, or open BitVector and select the LoRA in the app.

Examples

A1111 / Forge prompt with two SDXL LoRAs

<lora:zavy-cinematic-xl:0.7> <lora:detail-tweaker-xl:0.5> zavy cinematic, a lighthouse keeper on a stormy cliff, dramatic light, 35mm

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

ComfyUI Flux chain

Load Diffusion Model (flux1-dev-fp8) -> LoraLoaderModelOnly (flux-realism.safetensors, 0.8) -> ModelSamplingFlux -> KSampler

DualCLIPLoader (clip_l + t5xxl_fp8) -> CLIP Text Encode: "r3al photo of a street market at dusk, soft light, film grain"

PirateDiffusion (Telegram)

/render <realvis4-xl> <zavy-cinematic-xl:0.7> zavy cinematic, a lighthouse keeper on a stormy cliff, 35mm [[blurry, watermark]] /size:1024x1024 /steps:28

Tips

Check the hash: the model page shows the file SHA256 and AutoV2 so you can confirm a download matches the published file.
Example renders on the model page carry their full prompt and settings; copy one before you write your own.
Character LoRAs on SDXL and Flux work best at 0.7 to 0.9 with the character's name as trigger and no competing celebrity or style words.
Slider LoRAs (age, weight, detail) are meant to run at negative and positive weights; the model page notes the range the author intended.
Video LoRAs for Wan and Hunyuan are matched to a model size (1.3B, 5B, 14B) and often to the high-noise or low-noise expert; the page names it.
A LoRA that produces nothing usually has the wrong family or a missing trigger word; a LoRA that produces garbage usually has too high a weight.
Civitai LoRA pages sometimes list several versions; the catalog here lists each version separately with its own heat score.

models/loras (ComfyUI) or models/Lora (A1111, Forge), then the refresh button.

No install required - run AI models on the cloud

PirateDiffusion is Telegram only and built for pros: thousands of models, LoRAs and workflows driven by chat commands, with unlimited generation on a fixed price plan.

Try PirateDiffusion

Questions

Can I use a LoRA without downloading a checkpoint?

On the cloud, yes: PirateDiffusion and BitVector have the base checkpoints installed, you only name the LoRA. Locally you need the base model.

How many LoRAs at once?

Two or three with total weight around 2 is comfortable; more works on SDXL with lower weights, less well on Flux.

Are LoRAs safe to download?

The .safetensors format cannot contain executable code. Avoid old .ckpt or .pt LoRAs from unknown sources; compare hashes with the model page.

What is the difference between a LoRA and an embedding?

An embedding (textual inversion) is a few kilobytes that teaches the text encoder a new word; a LoRA changes the image model itself and can learn far more. Embeddings are mostly an SD 1.5 and SDXL thing (negative embeddings especially).

Can I train my own?

Yes, with kohya_ss, OneTrainer or AI Toolkit for Flux; 15 to 30 images and an hour on a 12 GB card for SDXL. Video LoRAs need more. This site lists what others trained so you can start from a finished one.