2025-09-02 · 3 min read

Using LoRA in Automatic1111

Automatic1111 is the most popular Stable Diffusion WebUI. Loading a LoRA is as simple as placing a file in the right directory and referencing it in your prompt.

File placement

After downloading sdxl-watercolor-lora-1.4.0.safetensors from ModelForgeLab:

cp sdxl-watercolor-lora-1.4.0.safetensors ~/stable-diffusion-webui/models/Lora/

Restart the WebUI or refresh the interface. The adapter appears in the "LoRA" dropdown on the Txt2Img tab.

Syntax

Use angle brackets in your prompt:

a mountain landscape, <lora:sdxl-watercolor-lora-1.4.0:0.7>

The format is <lora:filename:weight>. Weight ranges 0.0–1.0; 0.7 is typical starting point.

Weight interpretation

For watercolor adapters, 0.6–0.8 usually works. For subject adapters, 0.4–0.6 is safer.

Stacking multiple LoRAs

Stack two adapters:

a mountain landscape, <lora:sdxl-watercolor-lora-1.4.0:0.7> <lora:sdxl-product-photo-lora:0.3>

Order matters slightly; read left-to-right. The first LoRA is applied, then the second modifies the result.

Practical limit: 2–3 adapters. Beyond that, quality degrades and generation becomes slow.

UNet vs Text Encoder weighting

Advanced syntax separates weights for the model (UNet) and text conditioning:

<lora:sdxl-watercolor-lora:0.8:0.6>

First number (0.8) weights the UNet. Second (0.6) weights the text encoder (CLIP). This is useful when the adapter should strongly affect image generation but subtly affect prompt interpretation.

For most adapters, omit the second number; both default to the same weight.

XY Plot for sweep

Automatic1111 has a built-in XY Plot feature for testing ranges:

1. Go to Script dropdown (bottom left of Txt2Img)
2. Select "X/Y Plot"
3. Set X axis to "LoRA Multiplier Strength" (values: 0.3, 0.5, 0.7, 0.9)
4. Set Y axis to "Steps" (values: 20, 30, 40, 50)
5. Run

The result: a grid of 4×4 images showing how weight and step count affect output. Invaluable for finding the sweet spot.

Embedding vs LoRA

Do not confuse .safetensors LoRA with embeddings/ files (text embeddings). They are different artifacts:

• LoRA: Modifies model weights; strong effect
• Embedding: Teaches new token meanings; subtle effect

Both go in different directories and use different syntax in prompts.

Compatibility matrix

Automatic1111 officially supports:

If you have a FLUX-trained LoRA, use Stable Diffusion WebUI Forge (a maintained fork) instead.

Troubleshooting

LoRA not appearing in dropdown: - Check file extension: must be .safetensors or .ckpt - Restart WebUI or click "Refresh LoRA" button - Verify path: stable-diffusion-webui/models/Lora/

Generation fails or crashes: - LoRA rank too high for available VRAM (try lowering weight) - Architecture mismatch (SD1.5 LoRA loaded with SDXL base) - Corrupted safetensors file (re-download from registry)

Quality poor or artifacts: - Weight too high (reduce to 0.5) - Step count too low (increase to 30+) - Negative prompt may be fighting the LoRA (try empty negative)

Example workflow: watercolor product

Goal: generate a watercolor painting of a coffee mug.

Positive: a ceramic coffee mug on a wooden table, watercolor style, <lora:sdxl-watercolor-lora-1.4.0:0.7>
Negative: photorealistic, sharp, detailed
Steps: 30
CFG: 7.0
Sampler: DPM++ 2M Karras

Run 5 times with different seeds. Pick the best.

Best practices

• Start low: Begin with weight 0.5, increase if needed
• Use XY Plot: Find the right weight for your prompt style
• Combine with negative prompts: Reinforce the LoRA's direction
• Check base model: SDXL LoRA ≠ SD1.5 LoRA
• Save settings: Use Presets to remember good weight/CFG combos

Automatic1111 makes LoRA usage approachable. The WebUI handles all the complexity; users see a simple weight slider.

For production image generation, this is the most deployed pipeline in the world.