Stable Diffusion 3.5

What Stable Diffusion 3.5 is

Stable Diffusion 3.5 is a family of text-to-image models from Stability AI. The company announced the family on 22 October 2024 and released two of its members, Large and Large Turbo, that same day; a third, Medium, followed on 29 October 2024.^[1]

Like other text-to-image diffusion models, it works by starting from random noise and removing a little of it step by step, guided by a text prompt, until a picture forms. The prompt is the user's written description of the image, and how closely the result follows that description is what people usually call prompt adherence.

Three models at two sizes

Stable Diffusion 3.5 Large has 8.1 billion parameters and is aimed at professional use at 1-megapixel resolution.^[1] Stability says the model offers superior quality and prompt adherence, meaning it is meant to render both a clean image and the details a prompt asks for.^[1]

Stable Diffusion 3.5 Large Turbo is a distilled version of Large.^[1] Distillation means training a second model to reproduce much of the behaviour of a large one at lower cost. In this case, the result is a model that generates images in 4 steps, far fewer than a full sampling run would normally need, which makes generation markedly quicker.^[1]

Stable Diffusion 3.5 Medium, released a week later, has 2.5 billion parameters and is the smaller, more flexible sibling. It generates images between 0.25 and 2 megapixels, and needs 9.9 GB of VRAM excluding text encoders.^[1] The memory figure matters in practice: it sets the class of graphics card on which the model can be run locally, and it is why smaller variants of image models exist at all alongside the large ones.

Inside Medium

Medium uses an improved MMDiT-X architecture, and Stability highlights Query-Key Normalization in the transformer blocks as a change that stabilises training and simplifies fine-tuning.^[1] In a modern diffusion model, transformer blocks are the parts that let the image being drawn and the text prompt exchange information at every step. Attention is the mechanism that does the exchanging: it compares a query vector, standing for what a position is looking for, against key vectors, standing for what other positions offer, and uses the match scores to decide what to pull in. Normalizing queries and keys keeps those scores in a well-behaved range, which is what makes training steadier and fine-tuning easier.

Stability says Medium pairs the architecture with improved training methods.^[1] The practical consequence of a model this size is that a person or a small team can adapt it with their own images on ordinary hardware, which is harder to do with an 8.1-billion-parameter model.

Licence and where to get it

The models are released under the Stability AI Community License. That licence makes them free for non-commercial use, and free for commercial use by organisations with under $1 million in annual revenue; larger organisations need an Enterprise License.^[1] Users keep ownership of the outputs they generate.^[1] This is a middle path between a fully open licence and a closed service: small users and researchers get the weights at no cost, while larger commercial users are asked to pay.

The weights are published on Hugging Face and the inference code on GitHub, so the models can be downloaded and run directly.^[1] For people who would rather not manage hardware, there is hosted access through the Stability AI API as well as through Replicate, Fireworks AI, DeepInfra and ComfyUI.^[1]

Where it sits

The release is organised around choice rather than a single checkpoint. Large targets quality at a fixed, professional working resolution; Large Turbo trades some of the sampling cost for speed by cutting generation to 4 steps; Medium targets lower memory and a wide range of output sizes.^[1] Distributing the weights through several hosts and a permissive-for-small-users licence puts the models in front of both individual users and product teams, which is where a model family of this kind is usually judged: not only on benchmarks, but on whether people can actually run it.