ChatGPT Images 2.0 Still Can't Draw the Seven-Legged Spider I Want(will-keleher.com)
will-keleher.com
ChatGPT Images 2.0 Still Can't Draw the Seven-Legged Spider I Want
https://will-keleher.com/posts/chatgpt-image-2-still-cant-draw-a-seven-legged-spider/
4 comments
Your example doesn’t quite work btw! It removes two legs rather than just one.
It’s possible for sure, but it really doesn’t match most people’s intuitions for what these models should be able to do
It’s possible for sure, but it really doesn’t match most people’s intuitions for what these models should be able to do
I’m a doofus, I uploaded wrong final image because it took a few tries.
It might take a little while for the cache to clear, but once it does, it should show the updated image.
https://imgpb.com/OWKdy
And totally agreed, the systems can be highly unintuitive. To generate better imagery, they need more training data but ironically, the more training data they have, the more they tend to overfit when asked for something that goes against that data.
Meanwhile, you can walk up to a child, any child, show them a picture of a four-leaf clover for the very first time, then say, “Can you draw a five-leaf clover?” and they can instantly do it. They don’t need enormous datasets covering every possible variation.
It might take a little while for the cache to clear, but once it does, it should show the updated image.
https://imgpb.com/OWKdy
And totally agreed, the systems can be highly unintuitive. To generate better imagery, they need more training data but ironically, the more training data they have, the more they tend to overfit when asked for something that goes against that data.
Meanwhile, you can walk up to a child, any child, show them a picture of a four-leaf clover for the very first time, then say, “Can you draw a five-leaf clover?” and they can instantly do it. They don’t need enormous datasets covering every possible variation.
Nice! It's great to confirm that it can in fact do it, and it's interesting that even with the clear guidance from an expert it took a few tries.
https://genai-showdown.specr.net/ is fascinating btw!
https://genai-showdown.specr.net/ is fascinating btw!
> If I truly wanted to use an LLM image generator to create this spider, I’m sure it’d be possible.
Naturally. The easiest way to do it would be to generate a silhouette of a spider in the requested art deco style until you get one you like. Then, either quickly erase the leg manually or circle it with a red highlighter, and ask a editing model such as Nano Banana or GPT Image to remove that leg.
These kinds of problems where you’re overriding a very common, strongly learned object from the training data have historically been very difficult for most generative systems.
It’s the same reason the nine-pointed star example on my generative AI showdown site kills 90% percent of models. Even relatively simple prompts, like a five-leaf clover, were hard for a long time.
Practical demonstration
https://imgpb.com/OWKdy