So I make lots of models. It’s sort of obsessive. Like truly. Collecting, cleaning, organizing. It’s definitely on some sort of hoarding spectrum. Then at the end of the day I’ll never use the model (much). I’ll come up with the idea for another model- (like the Julia Child one which there is exactly zero use case for). Anyway, here’s a model I built a few years ago and ran some video clips through it recently. This it the vector graphics game LoRA I made from SDXL (!). I intend to re-build it for Flux/Zimage or something else at some point. Apparently Zimage is a super good choice for building LoRA’s on.

It’s clear on this last video you really need a higher resolution to capture the detail. So although I like this model for certain things, I think retraining the content on Flux 1 and genering synthetics from that to create video will look a little less…mushy. Although the look can be pretty cool.

A Wintry kinda thing.

02/17/2026

These are upscaled to 4k and slight interpolation added. I think I’ve pretty much hit the wall with this model as far as quality without re-training on a higher res dataset. There are hacks for longer clips but it really does put a hurtin’ on my server at home- even with 48GB @ 848x there’s a tradeoff between rendering at higher res (running out of memory in a shorter time span), or rendering at a lower res and upscaling later (lower quality output usually). At some point I’d like to make minutes-long clips but it requires Runpod and a lot more time than I currently have. These demos are fun though.

---