You might have noticed we’ve been less frequent with our substack updates in the previous weeks. First of all, our apologies. Secondly, let me tell you why.
Today, we are launching Distillery, our own Generative AI art service (invite link HERE). For the past several weeks, the team has been heads down, focused on developing and testing the system, and we now believe we have something very special to share with you.
Think of Distillery as Midjourney, but open-source — and with its own future trajectory.
We are building it because we believe we can create outstanding generative art by using small AI models like Stable Diffusion, provided we employ the right techniques.
Distillery is currently in an open alpha stage, available exclusively through Discord, and uses our previously unreleased model, Bloody Mary — to which we will dedicate a full post very soon. We found Bloody Mary to be the best cocktail version yet of Vodka, expanding on Vodka’s generality and bringing a unique personality to our generated images.
It’s good to stress that this is the alpha version of Distillery. Bugs, delays and crashes are to be expected as we iron out all the issues and learn how to ramp up the system. We will also be adding a ton of features before moving on to the beta stage. If you are interested in joining (for free!), just join our Discord server here.
It quickly became clear that we could create generalist checkpoints that stand out from the crowd, provided we use the correct finetuning methods and the right dataset. That was when we trained and launched the Vodka models — named as such because its creation process involved the distillation of knowledge from results of the “big AI model” (Midjourney) towards the “small AI model” (SD v1.5).
The results from our Vodka series were quite good, but still not at Midjourney levels of quality. To get the best results, it became clear that we needed to expand our experiments beyond model checkpoint creation and into other areas of quality optimization, including finding the best techniques for pre/postprocessing, and absorbing additional elements such as LoRas, embeddings and hypernetworks into a complete generative workflow.
However, in order to effectively operate this complete workflow, we realized we needed our own user interface. That's how Distillery was born. It’s a full-service generative art service, putting together all of the elements we’ve been working on, and coalescing them into a simple, established UI: a slash command on a Discord server.
As technology progresses and new models emerge, Distillery will serve as our platform to explore them with you. We are eager to tinker and experiment with new models such as Stable Diffusion XL in our quest to discover how to create truly amazing AI art.
A few samples of what our Bloody Mary model can do in Distillery.
Our pledge: Distillery is, and shall forever be, open source
Those who have been reading us for a while know we believe in open source software — we wrote extensively about why we believe the future of AI is open source. Well, we want to be around in the future, and there’s no way better to do this than to work towards building that future.
So, we are very happy to make the following pledge to the community.
All our techniques will be disclosed to the general public via our newsletter. To the extent that the technology allows, we will always prefer working with open source, small models over big models that need enterprise-level GPUs or that live behind a server cluster and a paywall. We will always be transparent with our methods and are eager to discussing them and learning from the community at large.
We look forward to seeing other people using our techniques outside of Distillery — in the comfort of their own homes or in their own generative art studios.
What we are launching today with Distillery
We are launching Distillery today as a generative art service, in Discord, as an alpha release, and under a freemium model. We are offering 3 tiers of service: Free, Early User, and Founder’s Pass.
Free is, well, free. Just join our Discord Server and we will serve you up to 10 free generation requests every day. Each free request provides you with 2 high-resolution images (1024x1024 pixels) using our new Bloody Mary model.
Early User is our basic subscription. For US$ 9.99/month, you get 300 premium requests per month, with no daily usage caps. Each request provides you with 4 images (as opposed to the 2 provided in the free tier). Premium requests also work with up to 4 GPUs running in parallel, as opposed to the Free tier that uses a single GPU per request, nearly doubling the speed of generation in our initial tests. Oh, and this is on top of the free requests, which are still available to use.
Founders’ Pass is our premium subscription, which will be offered for a limited time. For US$ 24.99/month, you get 1,000 premium requests per month (3.33x more than the Early User). Founders’ Pass members will receive lifetime access to all future public tiers and related offerings — we will provide you access to all future functions we release as part of the same subscription. In addition, you get access to an exclusive channel to connect with the Distillery team.
We must stress that this is still our MVP. We have huge plans for Distillery in the very near future, and we greatly look forward to discussing them in upcoming posts.
We look forward to seeing all of you on Discord!
In the meantime, here are a few more Distillery images, with their corresponding prompts. Feel free to test out the prompts in the system, and see for yourself what can be done there.