What is a LoRA?
The adapter format that turned model customization into a file download: where it comes from, how it works inside a diffusion model, and where consent enters.
What is a LoRA in one paragraph?
A LoRA (Low-Rank Adaptation) is a small set of trainable matrices injected into a frozen base model so that a new style, subject or concept can be learned without retraining the network. Proposed by Microsoft researchers on June 17, 2021, it cut trainable parameters by up to 10,000 times and became the standard add-on format for Stable Diffusion.
| Method | What changes | Typical size | Limits |
|---|---|---|---|
| LoRA adapter | Low-rank matrices added to attention layers; base weights frozen | A few to a few hundred MB | Quality depends on rank and data; stacks can conflict |
| Full fine-tune (checkpoint) | All model weights | 2–7 GB | Expensive to train and store; hard to combine |
| Textual inversion (embedding) | One new token vector in the text encoder | Tens of KB | Weak for complex styles or identities |
| ControlNet | A parallel network conditioning on pose, depth or edges | 0.7–1.4 GB | Controls structure, not style or identity |
How does a LoRA work?
Fine-tuning normally updates every weight matrix W in a network. LoRA freezes W and learns two thin matrices A and B whose product has a low rank r, typically 4 to 128, so the update is W plus A times B. Only A and B are trained and stored.
At inference the product is added to the frozen weights with a scale factor that users can dial from 0 to 1. The paper reports up to 10,000 times fewer trainable parameters than full fine-tuning of GPT-3.
We propose Low-Rank Adaptation, or LoRA, which freezes the pre-trained model weights and injects trainable rank decomposition matrices into each layer of the Transformer architecture, greatly reducing the number of trainable parameters for downstream tasks.
— Hu et al., “LoRA: Low-Rank Adaptation of Large Language Models” (arXiv:2106.09685, 2021), accessed September 3, 2026
How are LoRAs used with Stable Diffusion?
In Stable Diffusion the adapters attach to the attention layers of the UNet and often to the text encoder. Users load one or several adapters on top of a base checkpoint and set a weight for each. A “style LoRA” encodes a look, a “character LoRA” a face or body, a “concept LoRA” a pose or object.
Why do LoRAs matter for undress and NSFW tools?
They are the cheap route to specialization. An NSFW generator can pair a general base model with adapters for anatomy, lighting or a drawing style, which is how tools in the art gallery produce consistent looks; OpenArt AI exposes a public library of community styles.
The consent problem is the character LoRA: a few dozen photos suffice to train an adapter that reproduces a real person’s face on demand.
What are the legal and consent limits?
Training an adapter on photos of a real person without consent and using it to generate intimate images produces digital forgeries under the TAKE IT DOWN Act, exposes the maker to 15 U.S.C. §6851 claims and, for minors, to 18 U.S.C. §2256. Reviewed tools that accept custom adapters are scored on whether they screen training images and handle removal requests for identity-based adapters.
Consent Note
This site is a publisher: it reviews third-party tools and does not create, edit, accept or store images. Docs entries are general information, not legal advice.
Sources
- Hu et al., LoRA (2021) — arXiv:2106.09685 via Hugging Face, accessed September 3, 2026.
- microsoft/LoRA repository — GitHub, accessed September 3, 2026.
- Diffusers docs: LoRA training — Hugging Face, accessed September 3, 2026.
- CompVis/stable-diffusion repository — GitHub, accessed September 3, 2026.
- TAKE IT DOWN Act, Public Law 119-12 — GovInfo, accessed September 3, 2026.
- 15 U.S.C. §6851 — Cornell LII, accessed September 3, 2026.
FAQ
Is a LoRA a model?
Not on its own. It is a set of small weight deltas that only work on top of the base model it was trained for; a Stable Diffusion 1.5 adapter does nothing useful on SDXL or FLUX. Loading it modifies the base model’s behavior for one style, subject or concept.
How big is a LoRA file?
Usually between a few megabytes and a few hundred megabytes, depending on rank and on how many layers it covers. A full checkpoint of the same base model is 2–7 GB. This size difference is why adapters spread as downloads and one base model hosts thousands of styles.
Can I train a LoRA on my own photos?
Yes. Training on your own images, or on images of adults who have consented in writing, is lawful and needs roughly 20–50 photos, a consumer GPU and an hour or two. Training on someone else’s photos for intimate output is unlawful in the United States.
What is a “character LoRA”?
An adapter trained on many photos of one person or fictional character so that the model reproduces that identity from a trigger word. For fictional characters it is a style question; for real people it requires consent, because combining it with an NSFW base model produces digital forgeries.
Do LoRAs work with FLUX or SD 3.5?
Yes, but each base needs its own adapters, and the newer models’ transformer architecture changes which layers are targeted. FLUX and SD 3.5 adapters are typically larger and trained at 1024 px. Licenses differ: FLUX.1-dev restricts commercial use of derived adapters.
Why does the site mention LoRAs in reviews?
Because support for adapters explains quality differences and consent risks between tools. A review notes whether a service lets users upload custom adapters, screens the training images, and offers a removal route for identity-based adapters. Those answers feed the privacy and support components of the score.