HN Explorer

Things ML cannot draw or produce so far

3 points by aszantu ↗ HN
I do artsy things, and drawing and painting is one of them. You can see some of my stuff on https://www.deviantart.com/aszantu (forgot what email I registered it to and cant access DA anymore) and https://ko-fi.com/zantu2019

Dall-E and stable diffusion felt exciting for a while and I used up my Stable diffusion contingent quickly, could not manage to set it up on my linux, something always goes wrong in my setup and then it does not run.

So things I found the models could not draw on prompts. This one was used https://stablediffusionweb.com/#demo

An angel sitting under a mushroom

An angel with wings made from money

Feet

A cow with tiger stripes

a cat with wings

snake dragon winding around a key

Interesting results:

dog-faced cat, made the eyes look weird

milipede cat, I liked the patterns

maybe its me and im just not able to do the right prompts, if artists want to survive this they have to come up with some unique ideas. I gave up trying to survive as an artist long ago and went into IT

Edit: Formatting the list

4 comments

[ 3.0 ms ] story [ 19.2 ms ] thread
PaulHoule
Current large language models can't really extract semantic relationships from the syntax of language. That is, you might say

   "an fairy with a red left eye and a blue right eye sitting on top of a mushroom"
the model has no understanding that there are discrete entities being talked about that are to the left, right or are on the bottom or top of something else.

The model might do OK with the "fairy sitting on top of a mushroom" bit because it is trying to create an image that has various parts that match bits and pieces of the prompt, and the area around the bottom of the fairy and the top of the mushroom might match "sitting on top".

It is all a big pile of biases that happen to be close enough to our world view to impress it. I mean, you might get a chuckle out of an LLM that has learned that "Tyrone is a thug" but you don't want a robocop programmed that way because it won't produce justice.

muzani
DALL-E 2 feels like it has higher quality output, but it's narrowed towards a certain kind of output.

For example, a "bone axe" will just be a regular axe, with no attempt at making it bone. Midjourney understands that context, but it's still a little dumb with words and sometimes makes an axe with a bone.

Seeing GPT-3, it's probably smart enough to figure out really complex phrases. It just doesn't.

I think OpenAI has been biasing heavily for safety, targeting G rating, no deepfakes, no horror. Midjourney doesn't mind a little horror or nudity, but will avoid porn and furry stuff.