Today's news is the emergence of an initiative Open Model Initiative, which has made an application to create open, including from a licensing point of view, generative AI models. The scope is impressive: plans have been announced to create models for generating pictures, video and audio. The immediate plans are more modest: to create a working group. But what matters most is the context of this news.
Let's rewind a few months. Stability AI, a company known for its Stable Diffusion models, announced the release of the revolutionary Stable Diffusion 3. This is a set of models that was supposed to replace both the current SDXL and the old SD 1.5, which is still used where speed is important or where there are simply not enough computing resources. The announcement strongly leaned towards the “security” of the new model, which caused legitimate suspicions: the failure of SD 2.0, which, due to total censorship, had great difficulty generating ordinary earthly people, is still on everyone's lips.
Advertisement
Two weeks ago, the Stable Diffusion 3 Medium model was finally made available for download and local use. The model turned out to be even more “safe” than the failed SD 2.0; the request “woman lying on grass” managed to become a local meme.
Stable Diffusion 3 turned out to be emasculated not only in terms of generating regular people; compared to SDXL, almost all pictorial styles disappeared from the model; images mentioning the names of artists or famous people give out something average in the style of stock photos. The model turned out to be practically useless, but even in this form, third-party developers could probably fix it. However, the license for the new model turned out to be so strict that no one wanted to invest in training derivative models, and the owners of the Civit.Ai website, after consulting with a lawyer, completely deleted.
Thus, enthusiasts who received the long-awaited Stable Diffusion 3 model were not delighted with its quality, and the developers did not want to get involved with it because of the license.
The king is dead.
Advertisement
Let's return to our news. Having lost faith in the future of Stable Diffusion, the developer community decided to coordinate to create a new, improved model. Major participants in the project included application developers Invoke and Comfy, the Civitai portal and the LAION company. The developers of Invoke and Comfy (by the way, a former employee of Stability AI) have the experience necessary to create the model, LAION has a giant database of annotated images on which the model will be trained, and the Civitai portal has the computing resources necessary to train the model.
This composition of the team inspires some hope for possible success, but in any case one should not expect the model to appear soon: for now, a management structure and working groups are being created that will coordinate the development of the community.