Stable Diffusion 2

What Stable Diffusion 2 is

Stable Diffusion 2.0 is an open-source text-to-image model release published by Stability AI on 24 November 2022. Like earlier text-to-image systems, it takes a written prompt and produces a picture, but the 2.0 release rebuilt the part of the pipeline that reads the prompt and bundled several additional models with the base generator.^[1]

A new text encoder and a new training set

The text-to-image models in the release use a new text encoder called OpenCLIP, developed by LAION with support from Stability AI.^[1] The text encoder is the component that turns a prompt into the internal representation the image generator is conditioned on, so swapping it changes how the model responds to wording. Training used an aesthetic subset of LAION-5B, a large public image–text dataset; the subset was created by Stability's DeepFloyd team and filtered with LAION's NSFW filter to remove adult content.^[1] Default output resolutions are 512x512 and 768x768 pixels.^[1]

The models shipped alongside the base model

The release was a package rather than a single checkpoint. It included a 4x super-resolution upscaler that produces outputs of 2048x2048 pixels or larger, a depth-guided depth2img model, and a new text-guided inpainting model.^[1]

The depth2img model infers depth from the input image and then generates a new image from text and depth information, which means the spatial layout of an existing picture can be read from the original while the content is regenerated from a prompt.^[1] The text-guided inpainting model, used for filling in or replacing parts of an image, is fine-tuned on the new Stable Diffusion 2.0 base text-to-image model, so it inherits the new base rather than an older one.^[1]

Version 2.1 and the filter adjustment

Stable Diffusion 2.1 followed on 7 December 2022, a few weeks after the 2.0 release, in a 512x512 base variant and a 768x768 variant.^[2] The most discussed change was in content filtering: Stability loosened the NSFW filtering, saying the stricter 2.0 filter had removed too many people from the dataset. The change improved anatomy and hands.^[2] The company describes 2.1 as having a less aggressive NSFW filter that reduces false positives, though it still removes adult content.^[2]

2.1 also improved negative prompting. Stability said negative prompts work even better in 2.1, and its DreamStudio app gained negative prompting.^[2] A negative prompt tells the model what not to generate; in DreamStudio it can be added by appending a negative prompt with weight -1.0 to the prompt.^[2]

Legal aftermath

The models later became part of a trademark dispute. In the Getty Images v Stability AI case, the UK High Court found in November 2025 that some outputs from version 2.1 carrying watermark-like marks infringed the GETTY IMAGES trademark.^[3]

Where it sits

Stable Diffusion 2 marks the point where the project moved to a new text encoder built with LAION and to a curated, filtered training subset, and where the release took the form of a small family of models — base generator, upscaler, depth-guided generator and inpainting model — rather than one network.^[1] The quick 2.1 follow-up shows how directly the training-data filter shapes what the model can produce: tightening it removed people, loosening it brought back better anatomy and hands.^[2] For users, the practical additions were higher default resolutions, the ability to reuse the composition of an existing image through depth, and a way to steer generation away from unwanted content with negative prompts.^[1][2]