Dummy LoRA Generator
Create a valid SDXL LoRA .safetensors file with custom trigger tags — no training, no GPU, no model needed. Runs entirely in your browser.
What is a dummy LoRA and why would you use one?
A LoRA (Low-Rank Adaptation) file normally contains trained weight deltas that push a Stable Diffusion model toward a new style or concept. A dummy LoRA is different — its weights are near-zero (1e-6), so it barely changes the output. What it does carry is metadata: specifically the trigger tags you embed when you create it.
This is useful when a base model like Illustrious XL or Pony Diffusion already understands a concept — say, a specific character or art style — but the tags that activate it aren't widely known. Instead of training a full LoRA, you generate a tiny tag-carrier file that tells users exactly which tags to use. They load it in Forge or A1111, and the trigger tags auto-populate.
How it works
Enter a filename
Give your LoRA a name. It becomes the output filename and the modelspec title in the metadata.
Pick your base model
Select Illustrious XL, Pony, SDXL, or NoobAI. This gets written into the metadata so WebUIs know which model it targets.
Add trigger tags
Type each tag and press Enter. Tags are stored in the LoRA metadata. When loaded in Forge or A1111, they appear in the trigger word field automatically.
Download
Click Generate & Download to get a valid .safetensors file built from the SDXL LoRA architecture. Drop it into your Lora folder and use it immediately.
Use cases
Hidden tag discovery
Base models are trained on millions of images with tags that aren't widely documented. If you've discovered that a specific set of tags reliably activates a high-quality output — a character, style, or concept — package that knowledge as a dummy LoRA so others can benefit without needing to rediscover it.
Trigger word distributor
If you're sharing a workflow that depends on specific trigger tags, a dummy LoRA is a clean delivery mechanism. Users drag it into their Lora folder, load it at low strength (0.1-0.3), and the tags appear automatically in the trigger word field — no copy-pasting required.
Metadata-rich placeholder
Use the optional training metadata fields to create a LoRA that looks authentic to metadata readers and CivitAI upload forms. Fill in image count, steps, epochs, and optimizer to complete the picture before you eventually replace it with a real trained version.
Technical details
The generated file is a valid .safetensors LoRA targeting the SDXL text encoders (te1 = CLIP-L at 768 dim, te2 = CLIP-G at 1280 dim). It uses rank 1, the smallest possible LoRA rank, with weights initialized to 1e-6 — effectively zero effect on generation.
The safetensors format is built from scratch in JavaScript: an 8-byte little-endian header size, a JSON tensor metadata block, and packed float16 tensor data. No libraries, no WebAssembly, no network requests.
File structure
Architecture: SDXL LoRA (te1 + te2)
Rank: 1 | Alpha: 1
Layers: 36 tensors (12 per encoder × 3)
Dtype: float16
Weight value: 1e-6 (near-zero)
File size: ~120 KB
Format: safetensors v1
Frequently asked questions
Will this LoRA affect my image generation?▾
Barely. The weights are set to 1e-6, which is effectively zero. At the default LoRA strength of 1.0, the visual impact is negligible. If you want zero effect, use it at strength 0.1 or lower.
Can I use this on Pony Diffusion if I select Illustrious as the base model?▾
Yes. The SDXL LoRA architecture is identical across all SDXL derivatives. The base_model_hint in the metadata is informational only — it doesn't change the tensor structure. The file will load on any SDXL-based model.
What does the optional training metadata do?▾
Those fields (image count, steps, epochs, LR, etc.) write kohya_ss-compatible ss_ fields into the metadata. Tools like CivitAI model pages and LoRA metadata viewers read these to display training info. They have no effect on image generation.
Why is the file only ~120 KB when real LoRAs are hundreds of MB?▾
Real LoRAs use rank 32-128 and target dozens of UNet layers in addition to text encoders. At rank 1 targeting only text encoders, there are very few parameters. The file is small because there's genuinely very little in it — which is exactly the point.
Is my data private?▾
Yes. Everything runs in your browser. Your tags, filename, and metadata are never sent to any server. The .safetensors file is generated locally and downloaded directly to your device.