14 ms·

It's really amazing how DALL-E missed the boat. When it was launched, it was a truly amazing service that had no equal. In the months since then, both Midjourne

• aabbcc1241 4y ago
  and huggging face doesn't even require you to signup/login to have a try
• irrational 4y ago
  Equal or better quality? I suppose it depends on what you are trying to create, but that hasn't been my experience at all.
  • johnfn 4y ago
    Which one are you comparing against? I've tried hundreds of prompts between SD and DALL-E and get comparable results. Midjourney was lagging for a while, but the new --testp parameter is really remarkable, which, in my view, makes it superior not only to Stable Diffusion but also to DALL-E as well.
    • yreg 4y ago
      My experience is that with prompts that fit into OpenAI's limiting content policy DALL-E text2img results are usually much better. And I use SD like 95% of the time, so it's not the case that I would be more used to DALL-E.
      • KaoruAoiShiho 4y ago
        I need some examples because I don't really see it for the vast majority of usecases.
        • yreg 4y ago
          Here I wanted to illustrate the game Waffle[0], first attempt with Dalle was pretty good, not true for SD: https://labs.openai.com/s/rCzJwauuiaIj1Pd3IyJGaHS3 https://labs.openai.com/s/rCzJwauuiaIj1Pd3IyJGaHS3 Here I wanted an illustration of a nuclear plant in a japanese landscape, first attempt with Dalle produced multiple good results. I tried SD and MJ (back when MJ didn't use SD) as well, had trouble even with multiple attempts: https://labs.openai.com/s/FxhxtMFe3kFS8msV8vekRAJ3 https://labs.openai.com/s/FxhxtMFe3kFS8msV8vekRAJ3 There are others, but anyway I think my examples are not important since it will be always easy to cherry pick prompts that yield the best results in model X. In my experience SD is good at producing (especially non-photo-realistic) art that looks pretty and DALL-E is better at following a specific prompt when I know what exactly I want. Of course I recognise your experience might (and probably does) differ. [0] - https://wafflegame.net/ https://wafflegame.net/
    • gpt5 4y ago
      An easy example of DALL-E superiority is its ability to combine two different concepts together. For example, DALL-E performs extremely impressively on prompts in the format of “a still of homer Simpson in The Godfather” (replace character and movie as you wish). with the other two it’s a lot of misses
      • TillE 4y ago
        Yeah you're right that Stable Diffusion produces garbage for that prompt. I'd love to see a site with lots of examples of the same prompt fed into various models, I assume someone has already made that.
  • hooloovoo_zoo 4y ago
    Agreed; SD barely follows prompts at all.
    • zimpenfish 4y ago
      > Agreed; SD barely follows prompts at all. I would heartily disagree - I've generated ~6.5k images using SD locally and most of them could be linked to the prompt they came from.
      • hooloovoo_zoo 4y ago
        Doesn’t ‘most of them could be linked to the prompt they came from’ strike you as damning with faint praise?
        • twojacobtwo 4y ago
          > SD barely follows prompts at all. > ...and most of them could be linked to the prompt they came from. You made it sound as if there is almost no connection between the prompt and the images and zimpenfish said that the majority could be linked, implying a strong connection. He/she doesn't have to be praising it at all to counter your claim.
        • zimpenfish 4y ago
          Not hugely - e.g. taking the 38 prompts including "a painting by William Adolphe Bouguereau" (which is easily the worst of the modifiers for me), 10 of them I'd say were "no clue to the prompt". For the 56 Munch images, 54 were good and 2 were quibbles ("an isopod as an angel" had no isopod but did have an angelic human - is that a pass or no?) (Which is probably better than you'd get from a human given the exact same prompts.)
      • itintheory 4y ago
        Have you seen a decent tutorial for setting up SD locally? I've been using it through huggingface, but that seems pretty limited.
• minimaxir 4y ago
  I am surprised OpenAI didn't adjust the price of DALL-E 2 given the rise of free/low-cost competitors. Granted, DALL-E appears to be buckling under demand regardless so the supply/demand curve doesn't warrant a price drop yet.
  • ortusdux 4y ago
    Name brand recognition goes a long way.
• mrtksn 4y ago
  It's almost as if OpenAI got the right idea at the beginning but sometime somewhere maybe in a meeting room, contrary to their initial openness goals they decided to be closed walled garden for a product that doesn't exist. IMHO giving a prompt to generate image is amazing but isn't a product because you can't actually produce useful stuff(it's great as an exploration tool). It seems to me, OpenAI rushed into monetisation and control before having a killer app. On the other hand Stable Diffusion emerged as a free tool where large community can experiment and search for the killer app together. People started adapting it into other tools and workflows and so far it seems like the magic is in finding prompts that make the device generate good quality outputs. Earlier today I saw announcement about lexica.art(Stable Diffusion prompt tool) getting funded.
  • fennecfoxy 4y ago
    Eh tbh I think it's directly related to the way that they decided to handle GPT and its iterations; they were worried that GPT would be used by bots to better spam the entire Internet, and to be honest they're right. I'm guessing that worry translated to image generation as well. It's a Pandora's box thing I guess.
  • minimaxir 4y ago
    > IMHO giving a prompt to generate image is amazing but isn't a product because you can't actually produce useful stuff Making art/weird pictures doesn't have to be useful, as that use case is the entire reason MJ/SD went viral.
    • npunt 4y ago
      Indeed, "it's just a toy" is often the place these things start
    • hugozap 4y ago
      True, but if you can't integrate it into your workflow it will stay as a toy (and that's ok)
      • 4y ago
• Gregioei 4y ago
  You do understand how fast this space is moving and how new everything is right? Your criticism is in my opinion not valid. Do they need to react to the market? Perhaps depends on what there goal even is. Is dall-e 2 fun to use and cost wise totally fine? For me yes. But I also have people running SD with a hacky webui on some good GPUs for free. How many people actually have access to it. Is there also a good benchmark on which tool is inherent better? Because it is also totally fine to have multiple offerings. I really don't sure if you ever seen product development for yourself. Dall-e clearly took the potential misuse risk much further than others.
  • naillo 4y ago
    Just as user friendly as dalle but for stable diffusion and more free credits: https://beta.dreamstudio.ai/dream https://beta.dreamstudio.ai/dream
    • Gregioei 4y ago
      I did a short test and created already 20 pictures withhe same dall e prompt without a result as good as dall-e. And in another test the faces are super shitty. Dall e also gives you 4 pictures per credit and dream 1. So good to have more options I think. Two different products feeling different.
      • naillo 4y ago
        I generally find stable diffusion outputs better than dalle so it's surprising you say that. A good prompt makes a big difference though.
        • Gregioei 4y ago
          This doesn't make my experience less true. But I also played around with sd. I still think my original comment is valid.
          • 4y ago
• fullshark 4y ago
  Or maybe they got undercut by an open source implementation and that was inevitable no matter what. How can you compete with free?
  • notahacker 4y ago
    It's easier to compete with free (most paid products do) if most of the people interested in AI generated art have been paying for their service for months rather than browsing for alternatives. Especially since their supposed advantage is better prompt understanding rather than image quality; easy to dismiss StableDiffusion if your first impressions of it are "doesn't understand me like DALL-E" rather than "wow, this is magic" The "waitlist" model might work when the product isn't ready for prime time or the exclusivity is a part of the pitch, but it's greatly overrated in other respects. I got a "The Wait Is Over" email to tell me I'm off the waitlist and able to use a not-exactly-new stock trading app this week as the UK economy crashed. Yeah, thanks, but no thanks...
• kylevedder 4y ago
  I've spent a significant amount of time playing with the variety of Diffusion models available and DALLE 2 tends to produce much better quality images. The other killer feature is DALLE 2 has support for in-fill.
• ericd 4y ago
  This has been the dominant story going around, I guess because people want it to be true since they're pissed at OpenAI for not being so open, but StableDiffusion's text2image is nowhere near as good as DALL-E 2 in my experience. DALL-E 2 is incredible at that, StableDiffusion is not. But maybe it doesn't matter, because many times more people are playing around with StableDiffusion, such that the absolute number of good images being shared around is much higher with StableDiffusion, even if the average result isn't great.
  • johnfn 4y ago
    > I guess because people want it to be true since they're pissed at OpenAI for not being so open This is honestly not my experience at all. When I first tried SD and MJ, I did so with a very clear and distinct feeling that they were "knock-off DALL-Es" and I strongly doubted that they would be able to produce anything on the level of DALL-E. Indeed, I believed this for my first couple hundred prompts, mostly because I didn't know how to properly prompt them. After using them for around a month, I slowly realized that this was not the case, and in fact they were outperforming DALL-E for most of my normal usage. I have a bunch of prompts where SD and MJ produce absolutely beautiful and coherent artwork with extremely high consistency, that when sent to DALL-E, give significantly worse results.
    • wunderbaba 4y ago
      It depends on what you're generating, complex prompts in DALLE ("a witch tossing rapunzels hair into a paper shredder at the bottom of the tower") blow midjourney and stable diffusion out of water. But if all you're doing is the equivalent of visual mad Libs: "Abraham Lincoln wearing a zoot suit on the moon.", then SD and MJ suffice.
      • Terretta 4y ago
        With several thousand images on each, I agree with this -- to a degree. Dall-E does seem more aware of relationships among things, but using parens and careful word order in some of the SD builds can beat it. By contrast, even most failed images from MidJourney could still be in an outsider art gallery. MJ aesthetic works, while Dall-E seems like a 9 year old was taken hostage and clipped out Rapunzel and the paper shredder from magazines and pasted them onto a ransom note. That said, I have not been able to get any of Dall-E, MJ, or SD to give me a coherent black Ford Excursion towing a silver camping trailer on the surface of the moon beneath an earthrise. At cost per image, I could pay to get complex concepts such as this rendered via any number of art-for-hire sites at less expense and guaranteed results.
• logicchains 4y ago
  Maybe OpenAI is learning the same lesson compiler and runtime vendors learned a couple decades ago: it's very hard to compete with open source.
• diebeforei485 4y ago
  I suspect it was driven by moral panic and not necessarily business considerations.
• nextstep 4y ago
  this is something that people only on HN would write/believe. Missed the boat on what? Giving away free images from a prompt? This is all early days and these demos are neat but the real value is yet to be seen. Maybe when this technology is licensed and integrated into Photoshop or Instagram or something like that.
• miles 4y ago
  > Midjourney and Stable Diffusion emerged and got to the point where they produce images of equal or better quality than DALL-E I cannot speak to DALL-E's results, as the signup process is currently broken (after providing email, name, and phone number, was met with "We’re experiencing a temporary issue with signups due to a vendor outage. We apologize for the inconvenience!"), but the Stable Diffusion results I've been getting are not just unusable, but downright bizarre... here are the four images it produced for "morihei ueshiba doing a double backflip": https://imgur.com/a/EvkQpBT https://imgur.com/a/EvkQpBT
  • miles 4y ago
    Finally was able to get the signup process sorted (discovered that I had to use a different email address than the one I had originally requested beta access with); DALL-E's results for the same prompt were more human at least: https://imgur.com/a/OahhDS4 https://imgur.com/a/OahhDS4 .
• josephcsible 4y ago
  Saying "missed the boat" makes it sound like it was just bad luck and not OpenAI's fault, but I'd argue that it was their fault. They could have made DALL-E open source; they just chose not to.
• matsemann 4y ago
  What really exited me about SD was how many creative things it was used for because people could modify and use it. Just the first week I saw tens of different cool projects here on HN. With Dall-E, I have only ever seen prompts+images.
• shadowgovt 4y ago
  It's a little heartbreaking because arguably, OpenAI tried to do the responsible thing here: come up with a sustainable business model to make AI-generated images profitable while respecting trademarks and controlling for some objectionable content. Very corporate; very above-the-board. Emad Mostaque, a millionaire hedge-fund manager with money to burn, spent approximately $600,000 to train a model and dumped it out for public consumption: no account for how it will be used, no concern about any sociopolitical consequences, damn the torpedoes and straight ahead. He basically burned down a potential industry space and hugely complicated an ongoing conversation on how these tools will interact with / disrupt the lives and livelihoods of artists... But he also basically changed the world overnight. Hashtag-squad-goals, am I right? There's a lesson to be learned here. I haven't decided what it is yet. Though I note that it's a lesson that probably applies to few people who don't have $600,000 to set aflame.
  • dalmo3 4y ago
    > dumped it out for public consumption: no account for how it will be used, no concern about any sociopolitical consequences I hadn't heard the story of how stable diffusion was created. Sounds like the guy is a true hero from your description. And only for $600k? Imagine if he decided to "burn" the rest of his millions on similar initiatives.
    • shadowgovt 4y ago
      I unfortunately lack the imagination to think of similar initiatives that could be addressed in such a fashion.
• datacruncher01 4y ago
  I think the same thing is going to happen to the new models as well. Something better and more efficient is going to eat their lunch. Maybe we'll see more application specific models and a general model sitting on top of that to compost results together down the road.
• skybrian 4y ago
  This gets upvoted due to inexplicable DALL-E hate, but on the other hand I'm keeping my DALL-E account and cancelled my MidJourney account because the DALL-E account doesn't cost me anything when I don't use it. Having an account I barely use is great because I can go generate an image whenever I want for comparison purposes. (Furthermore, if I don't use it very often, I'm in the free tier due to the 15 free credits a month.) Also, do you realize that Stable Diffusion is also running a pay-for-usage model at dreamstudio.ai? I like that too.
• hartator 4y ago
  Yes, it does feel they shot themselves in the foot. Their marketing was excellent, but somehow pushed expectations too much and underdelivered. It also felt very elitist. Not very tinkers in a garage that this generation of stable diffusion feels like.
• obert 4y ago
  My thought on timing this is about Responsibility. I don’t think other options put any or so much consideration about AI impact. Perhaps we’re just finding out that people don’t care. Yet?
• AlexanderTheGr8 4y ago
  Can you link to the hundred-page document? I believe you are talking about prompt engineering, and I would love to get more information about it. I am struggling with figuring out good prompts.
  • srfwx 4y ago
    this one maybe? http://dallery.gallery/wp-content/uploads/2022/07/The-DALL%C2%B7E-2-prompt-book-v1.02.pdf http://dallery.gallery/wp-content/uploads/2022/07/The-DALL%C...
• dqpb 4y ago
  I think it's hard to move fast and be the morality police at the same time.
• danielfoster 4y ago
  For something that is supposed to be intelligent, sometimes the restrictions around DALL-E make no sense. My recent request to generate an image of two cats sleeping together was not allowed because apparently this is “adult content.”
• mberning 4y ago
  It reeks of “product management”. It’s getting managed out of relevancy.
• Keyframe 4y ago
  I'm still heavily exploring these new tools from an artist perspective. I never managed to get a run on Midjourney, but between DALL-E and SD there are quite a few differences. Broadly speaking, DALL-E seems to better get a hang on photographic results and interpreting "what I meant". With stable diffusion it's a lot of fiddling and putting manual emphasis on certain keywords until just getting it right. Overall, pricing will need to be adjusted over time as well. I set out on an experiment the other day that you can see here: https://twitter.com/Keyframe/status/1574338738808934400 https://twitter.com/Keyframe/status/1574338738808934400 I went about trying to utilize stable diffusion for an imaginary concept project (concept for characters of a remake of TMNT, heh). Process was similar to how I'd do it with another artist more than if I drew it alone. It was back and forth, from rough outlines and then honing into details. Inpainting and img2img helped A TON and I hope I'd get dreambooth running soon as well since that will be a game-changer in the combination of things. Between exploration phase, detailing, alternatives, and manual painting and over painting, I'd say PER OUTPUT final image I created in the region of a thousand or so interim images. Process overall did take a lot of time but not as much as completely manual and I didn't feel like I had as much control as manual of course, but I did feel ultimately bold enough that I thought I had creative control. With dreambooth I expect it to close the gap. Overall, I was extremely pleased with the experiment and I'll continue exploring it, even though I'm not doing artwork professionally anymore. And so far no, it's not going to replace artists. It's another tool removing labour, but adds time on direction needed. Ultimately it'll be another brush in the toolbox.
• bergenty 4y ago
  I’m not totally up to date on this but doesn’t DALLE still have the best quality.
  • Terretta 4y ago
    No
  • MetaWhirledPeas 4y ago
    Not from what I've seen. I took some prompts I saw used in Midjourney and was getting relatively lame results from DALL-E. Not a great comparison, but still.
• smrtinsert 4y ago
  Agreed. That miss was brutal, although freely distributed will always best gated experience in my book. Stable Diffusion is already in everyones hands.
• westoncb 4y ago
  I was perplexed by this happening since I'd have figured the level of expertise in OpenAI's executive team would make them unlikely candidates to lose out strategically (vs against something more out of their control). But maybe it was overconfidence? Or maybe what Stable Diffusion did was really just unpredictable.
  • moralestapia 4y ago
    I think they believed no one else could come out with something like what they had, plus some vanity/arrogance maybe. Like back then when they held their GPT model from the public because "it could destroy the world" or some similarly deranged argument. >the level of expertise in OpenAI's executive team would make them unlikely candidates to lose out strategically Sure, to be honest they're making mistakes that even complete marketing noobs know to avoid. Anyway, it's nice to see others are eating their lunch and sharing it with us.
• api 4y ago
  Even better with Stable Diffusion at least you could run it locally. Running software locally and using our desktop or portable supercomputers for something other than web browsing. What a novel concept. But how is this possible without cloud?
• nerderloo 4y ago
  OpenAI's DALL-E paper started all this generative image stuff. You've gotta give them credit for publishing the paper to public.
  • suyash 4y ago
    GANs have existed long before and such art was being made years ago.
• allisdust 4y ago
  All their actions seem to arise out of arrogance and misplaced belief that no one can replicate their work. Even before DALL-E, with GPT etc the sheer arrogance that led them to think that they could decide how their tools can be used, who can use them and for what propose was revolting. Frankly it spoiled the whole ML/DL revolution for me which till that point of time was a little more open than other fields. True that companies like Google don't let you use their models like Alpha* freely or even open source them but they aren't dangling access like these pricks who called themselves open were doing. I hope they will become irrelevant in the coming years.
• bowsamic 4y ago
  Imo they simply thought that no one was going to be able to compete with them
• DocTomoe 4y ago
  Also, and this is a big one: DALL-E has an overzealous filter, which blocks seemingly harmless prompts as "violating content guidelines". Look, I get it: They don't want to be in the news for producing porn or gore. But if you block a prompt and threaten account closure on repeated such blockings, at least tell us what we did wrong.
• forgingahead 4y ago
  Truly proves the saying, "Get Woke, Go Broke". All this pearl-clutching over safety really did a disservice to them. In all fairness, their release of Whisper[0] last week is actually really amazing. Like CLIP, it has the ability to spawn a lot of further research and work thanks to the open source aspect of it. I hope OpenAI learns from this, downgrades the "safety" shills, and focuses on producing more high-quality open source work, both code and models, which will move the field forward. [0]: https://github.com/openai/whisper https://github.com/openai/whisper
• bambax 4y ago
  Yeah the reason there's no "waitlist" is nobody's waiting to get in anymore. You don't need a bouncer if there's no queue at the door.