Hacker News Reader

Comment by pwinnski | Hacker News Reader

Let's try "chinese girl making heart with fingers --v 5", a prompt I had used before under v4. Result: https://i.imgur.com/jPgBqsX.png

In image 1, the subject appears to be missing a digit on each hand.

In image 2, the missing digits issue expands, combined with what appears to be a merging of the thumbs.

Image 3 introduces a supernumerary digit on each hand, and has extra "parts" disappear into and out of other fingers.

Image 4... isn't right in a number of ways, but still seems to have fewer than the expected number of digits.

I don't think these results are any better than the v4 model was, but decide for yourself. This is what I got with the same prompt using -v 4 on March 1. https://i.imgur.com/hQA2K7m.png I ended up just taking a photo of my daughter and blurring the background.

View on HN

Wow, imgur.com ate my back button.

Edit: I accidentally clicked share before trying to leave. Seems to persist even after closing the share dialog. Less of a dark pattern and more likely a bug, I suppose.

Someone's A/B test probably went great. Engagement and time spent goes way up if nobody can ever leave the page :)

Weird, I linked directly to the image as a PNG file, and checked it on my computer, it is just the image, nothing else.

Based on your comment, I just tried clicking on my phone instead, and sure enough, imgur intercepts and redirects to imgur.io on mobile. But it still didn't break my back button on iOS, so I'm not sure what's different for you.

same

V5 is vastly better. V4 mutations are outright creepy.

Neither of them produced a result I could use. One or the other might be more aesthetically pleasing to you depending on what you focus on, but both are very badly broken in fundamental ways.

If anything, I'd say v5 is more confident about its wrongness. It is as if humans have always had four fingers, how could you possible think otherwise? It that sense it seems more like the text-based LLMs: confidently incorrect.

V5 looks like photos with filters on. V4 looks like a painting.

they're both pretty flawed but v5 is better

it still doesn't seem to respect anatomy when it comes to hands particularly, perhaps an artefact of not paying attention to hard constraints that are not visually self-evident

These systems can't count in general. Doesn't matter if its fingers, n clocks on a wall (for all 1 digit vals of n), some variation of n items on an x, or even "a picture within a picture within..." k times. The machine can't count.

>I ended up just taking a photo of my daughter and blurring the background.

Ah, the diffusion process.

I suspect inability to count is a thing that can be trained for, although whether that will come at the expense of something else, I'm not sure.

I appreciate the humor. :)

Why the optimism? If it hasn't picked up on the gist of the meaning of 1-10 after 10 billion examples, why expect it to work after 20 billion examples? These models have worked unreasonably well so far, but there has to be a limit to how much we can substitute more data for our lack of conceptual understanding of cognitive processes.

Because I'm not sure counting is something that has been explicitly tagged or trained for. Clearly always-accurate counts are not an emergent feature, but if you tagged a large set of images with tags we think are too obvious to tag as humans, like "ten fingers" and so on, if that was an actual goal to be rewarded, I think it could improve results.

I'm overall in the skeptic camp, but it seems like these models can generally deliver what they're trained for. It doesn't appear any have been trained with counting as a primary goal.

Ten fingers surely wouldn't be tagged, but the thing about numbers is that they work as well for 10 apples as they do for 10 coins and 10 cups and 10 whatever. It never made the leap of abstraction to learn in the latent space the meaning of numbers 1-10 independent of the particular object. This lack of extrapolation lends it to a very rote learned style.

I already have a way around the specific numerical vocabulary/training problem. When I ask for "a picture in a frame", "a picture of a picture in a frame", "a picture of a picture of a picture in a frame" etc, I'm trying to use linguistic recursion to make numeracy emerge. But even that form of counting without predefined numbers fails. There's no reliable way to make a prompt the produces k of something. That's a deeper issue than it not deducing the specific meaning of characters 0-9 by example.

It can't learn that by example as there really isn't training data with more than 2 nested pictures of pictures, and by itself it will never realize it can just fill in the nested painting by prompting itself with the nested statement. It lacks thought loops.

Wrong. Look at eg Parti. Solely a matter of scaling the text encoder and then not screwing it up with unCLIP etc.

Its not wrong its an experimental observation based on systematic tests I've done. Maybe, with great strain either in quantity of training or ad hoc tweaking, this particular low bar can be hurdled. But it is still the case in general that numeracy is not prone to naturally and unexpectedly emerge from this type of system. The extent to which that matters depends on what your goal is. If you care about results in the here and now, then random patches optimized to use cases is fine. If you care about figuring out where to look for things we can do cognitively that are not very well suited to these generator models, the surprising lack of numeracy so far is a great starting point. The next obvious question is why it takes billions to notice the numbers 1 - 10? What models wouldn't suck at that?

> Its not wrong its an experimental observation based on systematic tests I've done.

You've done systematic tests on Parti?

How quickly the bar was raised. I can remember a time when we nitpicked Dall.E for creepy eyes.

Surprisingly some things we consider easy seem to be hard computationally.

I can count eleven mistakes in your comment on one hand. (Born near Chernobyl, ymmv)

Reply on news.ycombinator.com