Hacker News Reader

Comment by Flux159 | Hacker News Reader

This kind of research is great for reducing training costs as well as enabling more people to experiment with training large models. Hopefully in 5-10 years we'll be able to train a model on par with SD 1.5 with consumer gpus since that would be great for teaching model development.

View on HN

I'm pretty sure we are looking at something like 12 months. Not 5-10 years.

Pixart and this paper are good data points. Another even just 50x reduction in cost will make it possible on consumer hardware easily. This paper already claims over 100x reduction

I hope that somewhere around that period of time we will have AI-based "game" engines working at 30-40fps at 4K (of course with upscaling). I mean it might not be game engines per se, but universal, interactive pipelines for audio-visual content creation/consumption. Because right now I do not see any hope of such engines due to the number of models involved and the latencies this implies.

i'm really hoping for this as well! i'm a big believer that neural rendering pipelines will overtake traditional push-tris-to-gpus that we've essentially been using since Descent

Getting parity with SD 1.5 should require a similarly comprehensive data set, which seems a lot harder to source than a computer GPU. Especially now that we've got the A I-equivalent of pre/post nuclear steel.

Given how little artistic data humans need, there are probably breakthroughs coming that will reduce the size of the data set needed. Or make it so that a lot of the data required is more generic (like how a human artist needs vast amounts of audio-visual data from walking around every day, but maybe as little as a few megabytes to go from nothing to copying a new style and subject - then we can have a curated open source "highlights of the first 20 years of life" data set that everyone uses for basic training).

> Getting parity with SD 1.5 should require a similarly comprehensive data set, which seems a lot harder to source

Wasn't SD1.5 trained on LAION? So we know what it was and you could recreate it.

Although I thought LAION was why SD1.5 is kinda ugly at base settings because LAION is just random images both good and bad content and quality not aesthetic and high quality images.

Low-quality, non-aesthetic (but still natural) images are still pretty important for training a high quality, diverse model.

Also unless someone saved the actual data somewhere and will make it available, LAION-5B is rotted away.

The LAION datasets don't contain actual images, but URLs pointing to images. Due to link rot and deliberate scraper-blocking it may be difficult to download LAION images to retrain a model to match SD 1.5.

Reply on news.ycombinator.com