Image generation has come a long way since August 2022 (hbrooks.com)
43 points by hunterbrooks 2 days ago
gavmor 3 hours ago
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)
thenthenthen 8 minutes ago
A bit of context would be nice. I have literally no clue what I am looking at.
echelon an hour ago
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.
dahart 14 minutes ago
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.
gavmor an hour ago
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.
echelon 19 minutes ago
pants2 2 days ago
No modern or over-time comparisons?
In some ways I prefer looking at these over the hyper-polished 2026 image generators tbh.
vunderba 2 days ago
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.
DonHopkins 3 hours ago
Just ask any hyper-polished 2026 image generator for images in the style of August 2022 DALL-E 2.
fleshers 2 hours ago
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)
karim79 4 hours ago
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.
k12sosse 3 hours ago
For low ram solutions we have fun with Qwen 2.1 or zimage turbo
karim79 25 minutes ago
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.
slopinthebag 29 minutes ago
technically impressive but also soulless in a way. maybe i'm already biased knowing they're ai, but artistically they seem so uninteresting.
cellsstudio 3 hours ago
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.
WillAdams 2 hours ago
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.