Nice pictures. Really wish I'd started my GenAI benchmark back when DALL-E 2 were still around, just to have them as a fun little baseline for how much things have changed.
The only DALL-E 2 images [1] I could dig up in my archives that matched any of the bench prompts was the Alexander the Great prompt:
A painting of Alexander the Great riding a hippity hop into battle. The hippity hop toy was an old children’s toy that looked like a giant rubber ball with rubber handles that one straddle and ride.
And while it's not exactly the same, DALL-E Mini [2] (from back in 2021) is still around and you can run it locally using pretty modest hardware.
Images without prompts and statistics are... meaningless. We don't know if/how these adhere, whether they took a thousand hours or a thousandth of a second.
> “Would you tell me, please, which way I ought to go from here?”
> “That depends a good deal on where you want to get to,” said the Cat.
> “I don’t much care where—” said Alice.
> “Then it doesn’t matter which way you walk,” said the Cat.
If you do regular image gen, you can see that these images demonstrate memorized styles, broken anatomy, bad text layout (or hallucinated text), etc. Even the old Dall-E watermark is a throwback.
I don’t do image gen regularly… so I’d love more context. I can see some less good features in the images but I’m not fluent with what today’s generators would do. I agree with GP that it would be quite useful to know about the prompts & iteration effort. If generation times changed, they could be relevant & helpful.
I've been running a benchmark for comparing generative image models with a strong emphasis on complex prompt adherence since the days of SDXL back in 2023.
It’s not really intended to be viewed chronologically but there is a progress metric that shows how models have improved their scores over time, which might help you compare earlier models with later models such as Ideogram4 or Moss-Image released in 2026.
Oh wow, throwback!
Pretty remarkable how looking thru these janky old images actually did give me a sort of fuzzy nostalgic feeling. The tells/flaws are such a recognizable style.
I remember absolutely abusing the DALLE endpoints when we first got access after the waitlist ended. This was in my old job, where work looked a lot different from today. Also just a vastly different time in my life, funny how much time flies in four years.
(I do wish this had a before/after sort of comparison to demonstrate newer sota models handling the same prompts)
Oh my gosh. Makes me want to wake up my comfyUI setup on a RTX 2080 egpu which still works like a charm today even. I think the last time I touched it was about two years ago!
I always found the commercial stuff to be complete rubbish. I was astounded by how I could just spend a couple of hours using open models and open source software to produce better results.
Thanks for the insight, I honestly think 11gb VRAM is more than enough for image generation. The rest lies in just using a good model and having a basic understanding of the parameters and how to prompt.
What surprised me was that I basically "solved" image generation locally when it became a thing and then lost interest immediately after.
Commercial models/systems are stupid if you ask me. There is just no need for them anymore unless one needs ludicrous scale and has a fat wallet. And can tolerate all the shit that comes with nsfw bias.
Interesting post but I would love to see the process behind these images. I agree that the base capability has definitely increased but man, editing via prompts still remains a pain in the neck. I've found that it's difficult to be able to maintain stability in certain areas of an an image, while making changes in others, and avoiding artifacting across the images.
For a side project that's very art-asset heavy (and that I'm picky on), it takes me about an hour to get from an initial sketch up to a final image that I'm fairly happy with.
One technique which seems to work well for stuff with text/numbers is to ask that a JSON file be generated which contains the text, then edit that, then direct that the image be regeneratd with the updated JSON content.
I've been moderately curious if one could get such a description of the visual aspects of an image and make get consistency/repetition thus.
Nice pictures. Really wish I'd started my GenAI benchmark back when DALL-E 2 were still around, just to have them as a fun little baseline for how much things have changed.
The only DALL-E 2 images [1] I could dig up in my archives that matched any of the bench prompts was the Alexander the Great prompt:
And while it's not exactly the same, DALL-E Mini [2] (from back in 2021) is still around and you can run it locally using pretty modest hardware.[1] - https://imgpb.com/ZyzuX
[2] - https://github.com/borisdayma/dalle-mini
Images without prompts and statistics are... meaningless. We don't know if/how these adhere, whether they took a thousand hours or a thousandth of a second.
> “Would you tell me, please, which way I ought to go from here?”
> “That depends a good deal on where you want to get to,” said the Cat.
> “I don’t much care where—” said Alice.
> “Then it doesn’t matter which way you walk,” said the Cat.
(Lewis Carroll)
A bit of context would be nice. I have literally no clue what I am looking at.
These are not meaningless.
If you do regular image gen, you can see that these images demonstrate memorized styles, broken anatomy, bad text layout (or hallucinated text), etc. Even the old Dall-E watermark is a throwback.
Generation times are orthogonal.
I don’t do image gen regularly… so I’d love more context. I can see some less good features in the images but I’m not fluent with what today’s generators would do. I agree with GP that it would be quite useful to know about the prompts & iteration effort. If generation times changed, they could be relevant & helpful.
I don't do extremely regular image gen, but I've been running ComfyUI on my 3090 for, jeeze, two years?
Here's my homelab blog: https://www.gavmor.com/regular-labs/
Maybe I'm just not that observant, but all these pictures are just... "LGTM."
When I'm generating, I have really specific adherence/performance goals.
Spring 2022 was not an adherence/performance era.
I think we can still appreciate these snapshots in time.
In the future, this will be an aesthetic. Some already consider early AI jankiness a style.
No modern or over-time comparisons?
In some ways I prefer looking at these over the hyper-polished 2026 image generators tbh.
I've been running a benchmark for comparing generative image models with a strong emphasis on complex prompt adherence since the days of SDXL back in 2023.
It’s not really intended to be viewed chronologically but there is a progress metric that shows how models have improved their scores over time, which might help you compare earlier models with later models such as Ideogram4 or Moss-Image released in 2026.
https://genai-showdown.specr.net
Just ask any hyper-polished 2026 image generator for images in the style of August 2022 DALL-E 2.
Oh wow, throwback! Pretty remarkable how looking thru these janky old images actually did give me a sort of fuzzy nostalgic feeling. The tells/flaws are such a recognizable style. I remember absolutely abusing the DALLE endpoints when we first got access after the waitlist ended. This was in my old job, where work looked a lot different from today. Also just a vastly different time in my life, funny how much time flies in four years.
(I do wish this had a before/after sort of comparison to demonstrate newer sota models handling the same prompts)
Wrote up a blog entry super recently about the nostalgia of the flaws - https://blog.alexbeals.com/posts/i-miss-crappy-ai-art
Oh my gosh. Makes me want to wake up my comfyUI setup on a RTX 2080 egpu which still works like a charm today even. I think the last time I touched it was about two years ago!
I always found the commercial stuff to be complete rubbish. I was astounded by how I could just spend a couple of hours using open models and open source software to produce better results.
For low ram solutions we have fun with Qwen 2.1 or zimage turbo
Thanks for the insight, I honestly think 11gb VRAM is more than enough for image generation. The rest lies in just using a good model and having a basic understanding of the parameters and how to prompt.
What surprised me was that I basically "solved" image generation locally when it became a thing and then lost interest immediately after.
Commercial models/systems are stupid if you ask me. There is just no need for them anymore unless one needs ludicrous scale and has a fat wallet. And can tolerate all the shit that comes with nsfw bias.
technically impressive but also soulless in a way. maybe i'm already biased knowing they're ai, but artistically they seem so uninteresting.
Interesting post but I would love to see the process behind these images. I agree that the base capability has definitely increased but man, editing via prompts still remains a pain in the neck. I've found that it's difficult to be able to maintain stability in certain areas of an an image, while making changes in others, and avoiding artifacting across the images.
For a side project that's very art-asset heavy (and that I'm picky on), it takes me about an hour to get from an initial sketch up to a final image that I'm fairly happy with.
One technique which seems to work well for stuff with text/numbers is to ask that a JSON file be generated which contains the text, then edit that, then direct that the image be regeneratd with the updated JSON content.
I've been moderately curious if one could get such a description of the visual aspects of an image and make get consistency/repetition thus.