Oh man, Distillery — that open-source image gen service. Let me talk about how it all started back in the day.

So I’m always digging into the backstories behind projects for my content, and the Distillery launch post from the FollowFox team was pretty juicy.

Basically, they wanted to build an open-source version of Midjourney that runs on Discord with slash commands for generating images. What’s interesting is the starting point: they found that synthetic images actually performed better than real ones in fine-tuning, so they scraped together tens of thousands of upscaled images from Midjourney as their dataset and trained the Vodka series of models—clever name, since they’re distilling knowledge from big models into small ones, just like distilling spirits. This time they’re using a new model called Bloody Mary, which they say is an upgraded, flavored version of Vodka.

Let me jot down their pricing: free tier gets you 10 requests per day, 2 images each; Early User is $9.99/month for 300 requests, 4 images each; Founders Pass is $24.99 for 1000 requests plus lifetime access to all future features.

What really stuck with me was their promise: as long as the tech allows, they’ll always prioritize open-source small models over big ones hidden behind paywalls. That kind of attitude was pretty rare in the scene back then. Of course, they’re honest that it’s still in alpha, so crashes and delays are normal.

Using MJ upscaled images to train small models was pretty ahead of its time back then.

Vodka to Bloody Mary — that naming’s got some thought behind it.

Honestly, I really hope they stick to that promise about prioritizing open-source small models. Fingers crossed they don’t cave.

Honestly, it’s kinda counterintuitive but yeah — composite images actually fine-tune way better than real photos. Weird but true.

I didn’t believe it at first either—that fine-tuning composite images works better—but after running some tests, it really does, and the noise is actually less.