Ok it’s been just over a year since Dall-E launched. Dall-E is OpenAI’s image generation tool that is now available within the paid version of ChatGPT.
Back in the beginning on 2023 it was a bit of a joke about how bad rendering of fingers and hands was in image generators.
![](https://djangoandy.com/wp-content/uploads/2023/12/image-9-1024x537.png)
But by the middle of 2023 rendering of fingers and hands was ok.
It also somewhat sucked at horses.
![](https://djangoandy.com/wp-content/uploads/2023/12/image-10.png)
![](https://djangoandy.com/wp-content/uploads/2023/12/image-11.png)
There were a couple of issues here:
- Horses have “long and spindly” legs which image generators struggled with. Hopefully this will get fixed in 2024. While it’s not shown in these example to get 5 legged horses was not uncommon.
- Most horses generated are what a non horse person would say is a horse. Most horses generated were cartoonish and slab faced and bodied. Horses have a very subtle refinement to their conformation that horse people immediately recognise if it’s even slightly wrong. The sample images look cartoonish and out of proportion.
This image shows a cartoonish horse. Fine for most applications – but a horsey person immediately picks the conformation faults of the animal. Also I asked for a buckskin and got a palomino.
![](https://djangoandy.com/wp-content/uploads/2023/12/image-16.png)
ANYWAY
Dall-E is also pretty average at generating text.
I received this Meme:
![](https://djangoandy.com/wp-content/uploads/2023/12/image-12.png)
I provided this prompt within ChatGPT 4:
Prompt: Can you recreate this image to say “horses” and not “books”? use “horse” for “book” of course
It could not. Here is the result:
![](https://djangoandy.com/wp-content/uploads/2023/12/image-13.png)
There are few issues here:
- The text is nonsense.
- The flowchart has not been kept – Dall-E created a new one.
- The cropping is non optimal.
The graphics are a nice touch.
Lets try again:
Prompt: ok you need to try again please
![](https://djangoandy.com/wp-content/uploads/2023/12/image-14.png)
Same issues as before – but I admit I like the graphics more.
One more go:
Prompt: Take the original image and only change the word “book” to “horse” and the word “books” to “horses”. Keep the font size and font face. This is the only change I require.
Result is still average:
![](https://djangoandy.com/wp-content/uploads/2023/12/image-15.png)
Welp – looks like we are at the limits of Dall-E. Text is a real problem for it.
So what I’ll do is try this experiment again in 2024 – and suss another blog article for youse – so you can see Dall-E and ChatGPT improve over time.