What is Stable Diffusion?
The open-weight model family behind most image tools we review: versions, what each license allows, how its safety checker works and where undress apps fit.
What is Stable Diffusion in one paragraph?
Stable Diffusion is a family of open-weight latent diffusion models for generating and editing images from text, released by CompVis, Stability AI and Runway on August 22, 2022. Version 1 has an 860 million-parameter UNet at 512×512; later releases reached 8 billion parameters. Its open weights and permissive licenses made it the base of most undress and NSFW generators.
| Version | Date | License | Safety filter |
|---|---|---|---|
| 1.4 / 1.5 | August and October 2022 | CreativeML OpenRAIL-M | Optional CLIP-based safety checker in the Diffusers pipeline |
| 2.0 / 2.1 | November and December 2022 | CreativeML OpenRAIL++-M | Training data filtered with a NSFW score threshold; checker optional |
| SDXL 1.0 | July 2023 | CreativeML OpenRAIL++-M | Same optional checker; 1024×1024 base |
| 3 Medium / 3.5 | June and October 2024 | Stability Community License (free under $1 million revenue) | Data filtering plus usage restrictions in the license |
How does the model work?
The 2021 latent diffusion paper is the blueprint. A variational autoencoder compresses an image into a small latent grid; a UNet learns to remove noise from that grid step by step, guided by a text encoder (CLIP ViT-L/14 in version 1, OpenCLIP-ViT/H in version 2).
Generation starts from pure noise and denoises for 20–50 steps; img2img and inpainting start from a noised copy of an existing image instead.
By decomposing the image formation process into a sequential application of denoising autoencoders, diffusion models (DMs) achieve state-of-the-art synthesis results on image data and beyond.
— Rombach et al., latent diffusion paper (arXiv:2112.10752, 2021), accessed September 3, 2026
What do the licenses allow?
Versions 1 and 2 ship under CreativeML OpenRAIL licenses: weights may be used commercially, but the license lists prohibited uses, including generating sexual content without consent of the people depicted and content that exploits minors.
SDXL uses the OpenRAIL++-M variant. Version 3.5 moved to the Stability Community License, free for individuals and organizations under $1 million in annual revenue and paid above that, with the same use restrictions.
How does the safety checker work?
The Diffusers pipeline includes an optional safety checker: a CLIP-based NSFW classifier that compares the output embedding with a set of concept embeddings and blacks out images above a threshold.
Because the weights are open, anyone can disable it with one argument, which is what NSFW generators do. Version 2 also filtered its training set with the LAION NSFW detector, which is why the 1.5 lineage remains the most common base for adult fine-tunes.
Why do undress apps build on it?
Open weights, a working inpainting checkpoint, thousands of community fine-tunes and LoRA adapters, and hardware needs that fit a single consumer GPU.
Tools reviewed in the NSFW generator category rarely name their base model, but native 512 px output, characteristic artifacts and support for LoRA files point to the 1.5 lineage; SDXL-based services show 1024 px output. FLUX competes on quality but carries stricter licenses.
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
- Rombach et al., latent diffusion (2021) — arXiv:2112.10752 via Hugging Face, accessed September 3, 2026.
- CompVis/stable-diffusion repository — GitHub, accessed September 3, 2026.
- Model card: CompVis/stable-diffusion-v1-4 — Hugging Face, accessed September 3, 2026.
- Model card: stabilityai/stable-diffusion-xl-base-1.0 — Hugging Face, accessed September 3, 2026.
- Model card: stabilityai/stable-diffusion-3.5-large — Hugging Face, accessed September 3, 2026.
- LAION-AI/CLIP-based-NSFW-Detector — GitHub, accessed September 3, 2026.
FAQ
Is Stable Diffusion free to use?
Versions 1, 2 and SDXL are free under CreativeML OpenRAIL licenses, including commercial use, subject to a list of prohibited uses. Stable Diffusion 3.5 is free under the Stability Community License below $1 million in annual revenue; larger companies need an enterprise license.
Who created Stable Diffusion?
The CompVis group at Ludwig Maximilian University of Munich developed the latent diffusion method with Runway; Stability AI funded the compute for the August 2022 release. Later versions, including SDXL and 3.5, were released by Stability AI. Training data came from LAION-5B subsets.
Does Stable Diffusion have a built-in NSFW filter?
The reference pipeline includes an optional safety checker that blacks out images flagged by a CLIP-based classifier, and version 2 removed most explicit images from training. Because the weights are open, the checker can be disabled and models fine-tuned on adult data, which is how NSFW generators operate.
Which version do undress apps use?
Most do not say. Test behavior points to fine-tuned checkpoints from the 1.5 lineage, which produces 512 px output and has the largest supply of adult fine-tunes and LoRA adapters, or to SDXL derivatives with 1024 px output. Version 2 is rare in this category.
Can I run Stable Diffusion on my own computer?
Yes. Version 1.5 runs on a GPU with 4–6 GB of memory; SDXL needs about 8 GB and 3.5 Large around 16 GB without quantization. Local use keeps photos off third-party servers and removes the retention risk behind the 30-point privacy component in our reviews; the same consent rules apply.
Does the license forbid sexual content?
It forbids sexual content produced without the consent of the people depicted and any content sexualizing minors, among other listed uses. Consensual adult content is not excluded by the OpenRAIL text. Fine-tuning the weights to undress non-consenting people breaches the license as well as federal law.