alt Hacker NewsIf we think of GenAI models as a compression implementation. Generally, text compresses extremely well. Images and video do not. Yet state-of-the-art text-to-image and text-to-video models are often much smaller (in parameter count) than large language models like Llama-3. Maybe vision models are small because we’re not actually compressing very much of the visual world. The training data covers a narrow, human-biased manifold of common scenes, objects, and styles. The combinatorial space of visual reality remains largely unexplored. I am looking towards what else is out there outside of the human-biased manifold.
Replies
I find it likely that we are still missing a few major efficiency tricks with LLMs. But I would also not underestimate the amount of implicit knowledge and skill an LLM is expected to carry on a meta level.
Images and video compress vastly better than text. You're lucky to get 4:1 to 6:1 compression of text (1), while the best perceptual codecs for static images are typically visually lossless at 10:1 and still look great at 20:1 or higher. Video compression is much better still due to temporal coherence.
1: Although it looks like the current Hutter competition leader is closer to 9:1, which I didn't realize. Pretty awesome by historical standards.
> Generally, text compresses extremely well. Images and video do not.
Is that actually true? I'm not sure it's fair to compare lossless compression ratios of text (abstract, noiseless) to images and video that innately have random sampling noise. If you look at humanly indistinguishable compression, I'd expect that you'd see far better compression ratios for lossy image and video compression than lossless text.