We all knew that the quality of Stability Ai models regarding rendering humans, bodies and body parts is bad. But there was not much to complain about because it is a free model to use (SD 1.5 model).
With using so called “Lora`s” you were able to fix it a little bit. So called Low-Rank Adaptation applies tiny changes to standard checkpoint models. Most of the time the best you could do is to use a modified model, which user did train for rendering and made it public for free downloads. For example, the NSFW models of Mklan are amazing regarding fingers and hands. But it is NSFW! When you do use it, you are able to see what would be possible with Stability Ai models. Finally, it is a free to use model with less quality but with the chance with tricks and time-invest to get better quality out of it– so far so good.
When Stability AI did announce their new SD 3.0! model everybody was excited. But very soon all the users had sadly figured out that the quality of rendering pictures regarding humans in general is that bad and a waste of time; it even makes no sense to play around with it. Everybody in the happy community-universe was disappointed about the new model. The result was that not many “Lora`s” and modified models the community did create, and nobody does speak about this model very much. The model is strong when it comes to rendering something without human beings, like landscapes. The idea was to attack the quality of Midjourney picture quality, and it ended in a disaster for Stability AI. Oh, I forgot to mention that the model it is not free to use!
I guess the real disaster was going on behind the scenes. The people who did develop the generator of the Stable Diffusion models at the beginning of Stability AI, like Robin Rombach and Andreas Blattmann who did work as scientist for Stability AI until March 2024, did found their own company! Partrick Esser did join the team with others. They all did study at the past at the “Ludwig-Maximilians” university of Munich taught by the renowned professor Björn Ommer. The name of the new company is BlackForestLabs ! We all already do know this company very well. This is the company behind the new FLUX model which does drive the AI market crazy right now!
The scientist did leave Stability AI when the SD3 model was public and 6 months later these people made a model public, which does have much higher quality regarding physics of rendering of humans. It is not a surprise that there is the impression Stability AI intentionally did not give much care about the quality of the SD3 model. Maybe the company calculated that the community would fix this with self-created models again. It might be the main reason why users don’t use the SD3 model.
The whole situation is like a dilemma for a lot of partner pages of Stability AI. E.g. Civitai.com and others. They are stuck to the SD1.5 model or even the SDXL lightning model while the market did turn forward which huge steps to better quality for rendering pictures. There are rumors right now that Stability AI works on a new model to get close or even better like the quality of the new Flux models. Let’s see if this will happen.
Happy AI generation!
Example of the relationship between the Stability AI model and the Flux model. The same people were and are behind the model generation technics. The two front teeth you do find in any of these models. This example is a Flux model.
