logoalt Hacker News

echelonyesterday at 11:07 PM0 repliesview on HN

It's funny to see as a joke, but you can go the other way with this too. Image editing models and LoRAs for "previz-to-render upscaling" workflows are actually incredibly useful.

I was just writing about this (scroll about halfway down to the images of Sam Altman - though if you like that, do watch the second video):

https://getartcraft.com/news/world-models-for-film

The best model I've found for this, that almost bakes in full ControlNet capability, is oddly gpt-image-1.5. It's absolutely OP at understanding how to turn low-fidelity renders into final draft upscales.

Here are some older experiments:

https://imgur.com/a/previz-to-image-gpt-image-1-5-3fq042U

https://imgur.com/gallery/previz-to-image-gpt-image-1-x8t1ij...

https://imgur.com/aOliGY4

I just wish it didn't require invoking such heavy-weight, slow, and expensive models to do this. I'm sure open models will do this work soon, though.