I linked to this when DALL·E 3 was originally announced, but perhaps it's more appropriate now.
Last year I generated around 7,000 images using DALL·E 2 and uploaded them to https://generrated.com/
I've been re-running the same prompts using DALL·E 3, although haven't updated the site yet (although I'm planning to). So far I've created 2,000 like-for-like using those prompts.
----------
In the meantime, here are some things I've noticed with DALL·E 3 vs. DALL·E 2:
- the quality is astounding, especially illustrations (vs. photographs) — as I've been looking at the DALL·E 2 images I've constantly felt like the old images look like potatoes now we have DALL·E 3 and Midjourney (even though at the time they seemed stunning)
- you will struggle to get an output that references a specific artist, but it will sometimes offer to make images in the general style of the artist as a compromise
- it can get quite repetitive when you ask it for concepts — if you look at the 'representation of anxiety' images on Generrated, you'll see that there's a huge variety in them, but as I've been running them with DALL·E 3 it seems to prefer certain imagery (in this case, a human heart under stress appears a lot), and the 'discovery of gravity' will include a tree and an apple 80% of the time
- some of the prompts need guidance to get the output you desire — 'iconic logo symbol' works well on DALL·E 2 to create a logo, but with DALL·E 3 will often produce a general image/painting with a logo somewhere in the image (e.g. a NASA logo on an astronaut's suit rather than a logo of an astronaut)
Those are some I can remember off of the top of my head. But it's so much fun to play with!
Despite being trained not to mimic a specific artist, it’s incredible how the last image of anxiety with impressionism style (bottom-left) is close to the Wanderer above the Sea of Fog [0] (which is not even an impressionist painting!). It feels like it still very much rely on the underlying paintings used in the training material if the prompt is not more specific (also confirmed by the apple and tree example).
And it's fantastic by the way: it understands way better instructions than midjourney; also can do decent logos.
But I didn't cancel midjourney because it has more options, and is better at producing stunningly beautiful things.
The other comments are right though, the more time passes, the more open ai looks like a platform. But just like the Apple's platform, or twitter's, fb's, ms's, etc., it will adopt the features of the top apps built on it and kill them mercilessly.
It's incredible, but does seem to suffer from moral filters.
I tried to generate some DnD character art, and it generated an absolutely perfect depiction except for the wrong skin color. I tried multiple times to have it change it, but it replied every time with "there were issues generating all the images based on the provided adjustments". Asking it to change the outfit or gender was no problem though.
What's interesting to me is how AI-generated images (not just DALLE but also Midjourney and others) have a specific look and feel. It's typically characterized by high contrast and high saturation. Anyone know why that style is more likely to be the output?
It seems to me that I can get most any look and feel I want. Most of the ones other people post seem to have a very different look and feel to my own.
That said, you can sort of get a default look and feel if you just give a short prompt, and then it will tend toward the ones that are favored by RLHF. I prefer very long prompts.... as long as it will allow.
So if you do "a cool treehouse" you'll get sort of the default look. It will be very different if you say "treehouse, naturally occurring, in an old beautiful tree with branches that are low and spread widely and have lots of character and hanging moss and thick bark and curvy roots and mushrooms on a rocky outcropping from a mountainside. photograph, golden hour, sun through trees, damp from rain. Treehouse is part of tree, with fractal forms and live shaped wood and stone and stained glass and glowiness. art nouveau, gorgeous colors and fantasy design"
It's funny that the same people who complain that AI is "cheating" and uncreative, often are the ones who go to so little effort to get good results. It's not like it takes any arcane knowledge to get good images, but if you can use some imagination and string a lot of descriptive words together you can get so much better results.
It’s just lack of aesthetic talent, while competent images they’re very tacky looking. It’s not fundamental to AI because if you know what you’re doing you can push it to interesting aesthetics. Although Midjourney does some things behind the scenes to push everything towards that look you’re talking about.
Midjourney has been collecting human preference data for about a year now - every time you generate an image there and click on the one you want to enlarge you're providing a signal as to which image a human being preferred.
So my hunch is that humans prefer high contrast and high saturation images!
Reminds me of The New Coke, a flavor designed by having people provide very short term feedback on different flavor combinations.
Most feedback processes for generative models are based on asking the user to draw immediate sentiments rather than having them provide deeper art and style critiques
Try photography terms, as they're less likely to have uploaded HDR'd content under postings that include the technical specs of what they shot with, e.g. "24mm f/1.2" or "70mm f/2.8" or "D820" and such.
50mm (optimally with "Nikon" or "Canon" or similar) or 35mm will probably get you the most "natural" looking FoV. (Adding "lower" # fstops will get you a lot more depth-of-field/bokeh.)
The old adage is "f/8 and be there" so f/8 might get the most natural images if you want to specify an fstop(?)
This is where things like img2img in StableDiffusion really came in handy as you could simply apply an entire prompt like a photoshop filter..
There's a lot of variation between monitors so high contrast/saturation is far more likely to look the same across monitors. Digital art in general has moved towards this style.
I've got two LG 27" 4k monitors, same model number but produced several years apart, and while one monitor can easily show light grays like #EEE, it just looks white in the other.
The CFG is a trick used to trade variety for apparent quality, applied to images it causes high saturation and contrast. It's used because the training data set is very noisy.
The paper: https://arxiv.org/abs/2207.12598, of course using CFG change the sample distribution from the training distribution giving it that specific look.
A lot of it is driven by underlying params - if you decrease style param a bunch in Midjourney, some of the high contrast & high saturation affinity goes away.
There seem to be a lot of startups that are simply a thin veneer on top of the OpenAI APIs. We're going to see more and more functionality swept up into the OpenAI offering. If I was in one of those startups I'd be getting a little bit worried.
I don't know why anyone would be surprised at this turn of events. It's been pretty obvious all along that it was only a matter of time before OpenAI started productizing, so unless your startup is serving a very specific niche that you know inside and out, it's hard to see how anyone could justify starting a business whose value proposition is "combine these two ready-made OpenAI APIs".
They also keep silently worsening their API (persumably for a moat)
When turbo-gpt3.5-instruct launched you could see logprobs of words that you had in the prompt, then suddenly one day you couldn't, because it's "not possible" or "not available" or something like that.
I find it crazy that the first example they show is generating images for a science project. One would think that science is concerned with actual observations from the world, not generations that might be misleading or false. It’s kind of the arch-example I usually see for why student access to these tools is problematic.
> I am doing a report on cirrus clouds for my science class. I need photorealistic images that show off how wispy they are. I am going to compare them to photos I took of puffy cumulonimbus clouds at my house yesterday.
So the images are used for comparison against the photos that the researcher has already taken.
If they were to straight up base their thesis on AI made images, then I'd agree with you. But in this case it seems to be used as supplements, which seems fine to me, especially when used to highlight the difference between a "real" photo.
The images generated are much more coherent and compliant with the instructions than Midjourney's.
However Midjourney's has more beautiful, artistic pictures. Their recent upscaler is also very good. Midjourney is also much better at capturing the "style" of an artist.
Hands have definitely gotten much better in the latest version of Midjourney. Text is the one everyone struggles at, but early samples of Dall-E 3 had some promising examples.
Interesting that none of the new features (DALLE-3, Advanced Data Analysis, Browse with Bing) are usable without enabling history (and therefore, using your data for training).
Is it better than DALL-E 3 in Bing? I heard they ruined it by censoring it so much it was mostly unusable.
I've not paid for ChatGPT Plus as it just seems too expensive for my use, but i've been quite intreasted in getting GPT4 access, adding DALL-E 3 to the mix makes it more worthwhile for me now.
I’d be worried if I were MidJourney now. This seems just significantly better for all sorts of things other than pure art stuff. Anything that requires text or strict following of instructions is immeasurably better in Dalle3 than in MJ. And ChatGPT can take your vague, bad prompts and turn them into pretty good instructions for the model. I actually downgraded my MJ plan since it became clear to me that I’d be using it far less now. Hopefully they can come up with a response to it, but it’s a tall order to integrate an LLM the way OpenAI has done here.
I'd be worried if I were Midjourney specifically because Discord is a terrible user interface. Trying out DALL-E 3 on a regular web interface ia so much better.
I'd try out MJ again if they had a regular website. I don't even like OpenAI as a company, but I can't stand using Discord like that.
Overall I prefer MidJourney's styles and feature set. But it's really, really hard to make MidJourney draw the things I want, especially when there's specific/detailed scenery I want to depict. The latter is now quite doable using DALL-E 3 even though the drawing itself may not be as good as MidJourney.
I recently generated images for a presentation. It took about 30 tries to generate 5 suitable images. But I burned 60 MidJourney generations and in the end none of the results were satisfactory. But because they were ugly but because they didn't properly depict the concept I wanted.
Now, if I can import a DALL-E 3 image into MidJourney and then use Zoom Out from there, that would be wonderful.
I’ll re-up here that I’ve collected all the ChatGPT system prompts together [0], including DALL•E 3 [1]. I wanted them all in one place as I work on the next release of my ChatGPT AutoExpert [2] “custom instructions”, which will include custom instructions for Voice Conversations, Vision, and yes, DALL•E.
I'm not an artist by any means but I don't have to pay for MidJourney anymore separately, everything all in ChatGPT now and I can get the same if not better results.
Me, my wife and children can now play with this and become artists (if they choose to be) now without switching websites.
What a great time to be alive, this is the future.
I quickly found that this feature has rate limits: "I apologize for the inconvenience, but due to rate limits, I'm unable to generate images at this moment. Please wait for 14 minutes before generating more images. In the meantime, I'm here to help with any other questions or information you might need."
"I'm sorry for the inconvenience, but I cannot provide original photorealistic images directly. However, I can help guide you to sources where you might find such images or provide more information on cirrus clouds to support your report."
Note, the wrong grandparent distribution but at least the structure is right.
ChatGPT even provided a decent prompt for Dall-E version.
However the Dall-E version was giving horrible cyclical graph monstrousities that in no way resembled tree, just graphs with multiple fathers, mothers, complete non-sense.
Also, I was hoping to see pictures of people but that also was failing.
It have some restrictions on i.e generating something based on works from a century back, but it is impressive.
Told it to generate images based on the song Vincent (that generated some Van Gogh style drawings), then ask it to generate the same, but with Tintoretto style (couldn't use some newer artists, even using the song had objections), and then added corrections on some of the generated pictures, with impressive results.
It looks like a translation job. I mean it asks in a language that Dall-E speaks, for something that somewhat was implied in what I said, in the way that Chatgpt understood it.
Not available via API yet, right? I've been hosting a stable diffusion instance for a while, and even with the latest SDXL models Dalle3 really blows it out the water, would love to try it out.
Were there any improvements in distancing generated images from the training set? I'd like to use AI images commercially, but I'm always afraid of some person claiming that the image looks just like their work.
Edit: I see they tried to make sure the image doesn't look like the "style of living artists", and added the option of people to opt-out their images from the training set. Progress, but is this enough? I don't think so.
MY favorite party of dalle3 in chatgpt+ is being able to conversationally ask for changes to the output. I am a big fan of SD, but this is very very good.
[+] [-] davidbarker|2 years ago|reply
Last year I generated around 7,000 images using DALL·E 2 and uploaded them to https://generrated.com/
I've been re-running the same prompts using DALL·E 3, although haven't updated the site yet (although I'm planning to). So far I've created 2,000 like-for-like using those prompts.
----------
In the meantime, here are some things I've noticed with DALL·E 3 vs. DALL·E 2:
- the quality is astounding, especially illustrations (vs. photographs) — as I've been looking at the DALL·E 2 images I've constantly felt like the old images look like potatoes now we have DALL·E 3 and Midjourney (even though at the time they seemed stunning)
- you will struggle to get an output that references a specific artist, but it will sometimes offer to make images in the general style of the artist as a compromise
- it can get quite repetitive when you ask it for concepts — if you look at the 'representation of anxiety' images on Generrated, you'll see that there's a huge variety in them, but as I've been running them with DALL·E 3 it seems to prefer certain imagery (in this case, a human heart under stress appears a lot), and the 'discovery of gravity' will include a tree and an apple 80% of the time
- some of the prompts need guidance to get the output you desire — 'iconic logo symbol' works well on DALL·E 2 to create a logo, but with DALL·E 3 will often produce a general image/painting with a logo somewhere in the image (e.g. a NASA logo on an astronaut's suit rather than a logo of an astronaut)
Those are some I can remember off of the top of my head. But it's so much fun to play with!
----------
Edit: I quickly put together 3 comparisons between v2 and v3: https://imgur.com/a/L9DYCSA
[+] [-] KolmogorovComp|2 years ago|reply
[0] https://en.wikipedia.org/wiki/Wanderer_above_the_Sea_of_Fog
[+] [-] MrNeon|2 years ago|reply
From what I've read in several places DALL-E 3 in ChatGPT uses the same seed for every generation which can exacerbate that problem.
[+] [-] BiteCode_dev|2 years ago|reply
But I didn't cancel midjourney because it has more options, and is better at producing stunningly beautiful things.
The other comments are right though, the more time passes, the more open ai looks like a platform. But just like the Apple's platform, or twitter's, fb's, ms's, etc., it will adopt the features of the top apps built on it and kill them mercilessly.
[+] [-] dtech|2 years ago|reply
I tried to generate some DnD character art, and it generated an absolutely perfect depiction except for the wrong skin color. I tried multiple times to have it change it, but it replied every time with "there were issues generating all the images based on the provided adjustments". Asking it to change the outfit or gender was no problem though.
[+] [-] huytersd|2 years ago|reply
[+] [-] danielvaughn|2 years ago|reply
[+] [-] Kuinox|2 years ago|reply
[+] [-] robbrown451|2 years ago|reply
That said, you can sort of get a default look and feel if you just give a short prompt, and then it will tend toward the ones that are favored by RLHF. I prefer very long prompts.... as long as it will allow.
So if you do "a cool treehouse" you'll get sort of the default look. It will be very different if you say "treehouse, naturally occurring, in an old beautiful tree with branches that are low and spread widely and have lots of character and hanging moss and thick bark and curvy roots and mushrooms on a rocky outcropping from a mountainside. photograph, golden hour, sun through trees, damp from rain. Treehouse is part of tree, with fractal forms and live shaped wood and stone and stained glass and glowiness. art nouveau, gorgeous colors and fantasy design"
It's funny that the same people who complain that AI is "cheating" and uncreative, often are the ones who go to so little effort to get good results. It's not like it takes any arcane knowledge to get good images, but if you can use some imagination and string a lot of descriptive words together you can get so much better results.
[+] [-] whywhywhywhy|2 years ago|reply
[+] [-] simonw|2 years ago|reply
So my hunch is that humans prefer high contrast and high saturation images!
[+] [-] Taek|2 years ago|reply
Most feedback processes for generative models are based on asking the user to draw immediate sentiments rather than having them provide deeper art and style critiques
[+] [-] dontupvoteme|2 years ago|reply
50mm (optimally with "Nikon" or "Canon" or similar) or 35mm will probably get you the most "natural" looking FoV. (Adding "lower" # fstops will get you a lot more depth-of-field/bokeh.)
The old adage is "f/8 and be there" so f/8 might get the most natural images if you want to specify an fstop(?)
This is where things like img2img in StableDiffusion really came in handy as you could simply apply an entire prompt like a photoshop filter..
[+] [-] shmatt|2 years ago|reply
[+] [-] civilitty|2 years ago|reply
I've got two LG 27" 4k monitors, same model number but produced several years apart, and while one monitor can easily show light grays like #EEE, it just looks white in the other.
[+] [-] GaggiX|2 years ago|reply
The paper: https://arxiv.org/abs/2207.12598, of course using CFG change the sample distribution from the training distribution giving it that specific look.
[+] [-] sebzim4500|2 years ago|reply
[+] [-] kubrickslair|2 years ago|reply
[+] [-] caesil|2 years ago|reply
[+] [-] filterfiber|2 years ago|reply
Basically on A/B tests, humans tend to prefer more saturated, "punchy" images. Which is also why iPhones tend to do the same thing.
For artificial images people also seem to prefer stylized "dramatic" styles as well.
And the model was finetuned to match.
[+] [-] iamflimflam1|2 years ago|reply
[+] [-] lolinder|2 years ago|reply
[+] [-] dontupvoteme|2 years ago|reply
When turbo-gpt3.5-instruct launched you could see logprobs of words that you had in the prompt, then suddenly one day you couldn't, because it's "not possible" or "not available" or something like that.
[+] [-] gaganyaan|2 years ago|reply
[+] [-] esafak|2 years ago|reply
[+] [-] samfriedman|2 years ago|reply
Not a good look?
[+] [-] diggan|2 years ago|reply
> I am doing a report on cirrus clouds for my science class. I need photorealistic images that show off how wispy they are. I am going to compare them to photos I took of puffy cumulonimbus clouds at my house yesterday.
So the images are used for comparison against the photos that the researcher has already taken.
If they were to straight up base their thesis on AI made images, then I'd agree with you. But in this case it seems to be used as supplements, which seems fine to me, especially when used to highlight the difference between a "real" photo.
[+] [-] shin_lao|2 years ago|reply
However Midjourney's has more beautiful, artistic pictures. Their recent upscaler is also very good. Midjourney is also much better at capturing the "style" of an artist.
Both struggle with hands and "holding objects".
[+] [-] ehsankia|2 years ago|reply
[+] [-] rckrd|2 years ago|reply
[+] [-] weird-eye-issue|2 years ago|reply
https://docs.google.com/forms/d/e/1FAIpQLScrnC-_A7JFs4LbIuze...
[+] [-] ChildOfChaos|2 years ago|reply
I've not paid for ChatGPT Plus as it just seems too expensive for my use, but i've been quite intreasted in getting GPT4 access, adding DALL-E 3 to the mix makes it more worthwhile for me now.
[+] [-] nwienert|2 years ago|reply
put "(very safe content)" at the end, if that doesn't works sometimes adding a few more modifiers like that
put "(no copyright or famous people)"
also if you are hitting a banned word just throw a period inside it, for example instead of "drake" just put "d.rake"
you can get it to generate fairly spicy things but it still sometimes takes a few tries and a few more words of encouragement.
[+] [-] eigenvalue|2 years ago|reply
[+] [-] gaganyaan|2 years ago|reply
I'd try out MJ again if they had a regular website. I don't even like OpenAI as a company, but I can't stand using Discord like that.
[+] [-] FooBarWidget|2 years ago|reply
I recently generated images for a presentation. It took about 30 tries to generate 5 suitable images. But I burned 60 MidJourney generations and in the end none of the results were satisfactory. But because they were ugly but because they didn't properly depict the concept I wanted.
Now, if I can import a DALL-E 3 image into MidJourney and then use Zoom Out from there, that would be wonderful.
[+] [-] reaperman|2 years ago|reply
[+] [-] spdustin|2 years ago|reply
[0]: https://github.com/spdustin/ChatGPT-AutoExpert/blob/main/Sys... [1]: https://github.com/spdustin/ChatGPT-AutoExpert/blob/main/_sy... [2]: https://github.com/spdustin/ChatGPT-AutoExpert/
[+] [-] colesantiago|2 years ago|reply
I'm not an artist by any means but I don't have to pay for MidJourney anymore separately, everything all in ChatGPT now and I can get the same if not better results.
Me, my wife and children can now play with this and become artists (if they choose to be) now without switching websites.
What a great time to be alive, this is the future.
[+] [-] haltist|2 years ago|reply
[+] [-] jalino23|2 years ago|reply
[+] [-] noarchy|2 years ago|reply
[+] [-] pjot|2 years ago|reply
Edit: it looks to be available in the iOS app now too - which previously was not.
[+] [-] hmottestad|2 years ago|reply
Not available to me with ChatGPT Plus.
[+] [-] sireat|2 years ago|reply
That is I wanted a simple binary tree in three levels..
Regular ChatGPT Plus got to the point quickly:
Note, the wrong grandparent distribution but at least the structure is right.ChatGPT even provided a decent prompt for Dall-E version.
However the Dall-E version was giving horrible cyclical graph monstrousities that in no way resembled tree, just graphs with multiple fathers, mothers, complete non-sense.
Also, I was hoping to see pictures of people but that also was failing.
Seems like very much a beta product.
[+] [-] gmuslera|2 years ago|reply
Told it to generate images based on the song Vincent (that generated some Van Gogh style drawings), then ask it to generate the same, but with Tintoretto style (couldn't use some newer artists, even using the song had objections), and then added corrections on some of the generated pictures, with impressive results.
It looks like a translation job. I mean it asks in a language that Dall-E speaks, for something that somewhat was implied in what I said, in the way that Chatgpt understood it.
[+] [-] Zpalmtree|2 years ago|reply
[+] [-] bradhilton|2 years ago|reply
[+] [-] glimshe|2 years ago|reply
Edit: I see they tried to make sure the image doesn't look like the "style of living artists", and added the option of people to opt-out their images from the training set. Progress, but is this enough? I don't think so.
[+] [-] nickthegreek|2 years ago|reply