I asked GPT a while back to pick a theme to explore in 5 art pieces, it picked mental health and one it decided it liked was "the weight of anxiety". SD didn't really capture it very well IMO, this does. Here's the results
Now, you might say this is a bland or uninspired take, but I want you to step back a moment if you're thinking that. I asked a machine to brainstorm some ideas for a theme for an art show. Then told it to pick one it liked, and expand on it to five pieces with descriptions (I gave it a very short explanation about how to prompt an AI art generator). I then fed this directly into another tool that gave these images as a result.
I did not tune the prompt, I didn't even account for the resulting description being the wrong kind of format or that it got cut off at the end.
I'm not going to make any statements about replacing artists, I am just going to reflect on what a wild few months it's been in AI development.
I’m just sick of reading that it’s the future, I mean a hype cycle. It’s going to take all our jobs, I mean it’s shit. It can beat a grandmaster at chess, I mean it can’t win at tik-tak-toe.
It’s somewhere in the middle of all these things. It’s clearly groundbreaking technology that will be useful in many ways for a long time to come, but is still in early days with bugs to be fixed, and key features not yet imagined.
The conversation here could just be a little more interesting and varied.
I asked Bing Chat why it says it can't create images for me.
>The Bing Image Creator feature is a new feature that allows you to create images with your words. It is currently available in the United States, Canada, Australia, and the United Kingdom.
So just one hour after Google Bard we have yet another senselessly geoblocked feature.
edit: Okay, the separate link (https://www.bing.com/images/create) works, so I'm even more confused why Microsoft wouldn't allow access to this from Bing Chat.
Could you tell what the difference in language understanding is between Midourney and this AI (apparently Dall-E 3)? Visual quality is one thing, following the prompt is another. Example:
A monk playing chess with a clown
Two yellow books on top of three blue books
Edit: I have the Bing mobile app (Android), but the image creator is not present in the "app" list on the lower right.
I didn't even need to apply. It seems pretty good, though perhaps still not quite on the level of Imagen or Parti (both Google Research). For example, text still doesn't work.
Yes, the UI part is a mess. Now they have three UIs to search: old search UI, Chat like UI and image generation UI. To make this worse they are not consistent in anyway, sometimes when I scroll in the chat UI it throws me to useless search UI, and each seem to have different font sizes, colors etc. It feels like hodgepodge of things tacked together.
It's not senseless when countries apply laws that affect websites operated from other countries. If I'm releasing a prototype in the USA I'd rather worry about GDPR compliance down the road.
Laws regulating websites have tradeoffs and this is one of them.
Also maybe they have some loyalty to their US American countrymen. Reminds me of how a majority in polls in the US said that the Pfizer vaccine should be distributed to US Americans first, while people in almost all other countries were more in favor for international ratios. (Although this case isn't equally justified here, since the vaccine was produced by a US company, Pfizer, but developed by a German company, BioNTech.)
Realistically most of this tech was actually developed in academic labs and at Moderna, then knowledge leached out to biotech competitors.
> while people in almost all other countries were more in favor for international ratios
I have seen the first polling you've mentioned (66% of Americans favor prioritizing US). I would love the source if you have it on similar polling for people outside the US on whether to prioritize their own country.
I don't think the technology was leaked from Moderna, as BioNTech was a major player in this mRNA technology before.
I don't have the poll, but as far as I remember only in the US was the majority for prioritizing the own country. Though the amount of money spent on "project warp speed", a US effort not matched anywhere else, could play a role here.
it's fun to try to find flaws in content blocking.
interesting datapoints: it will block some sexually suggestive prompts regarding women (a woman swimming in milk) but not for men (a man swimming in milk)
changing milk to "almond milk" got the prompt accepted but then blocked before showing image
refused any prompt with the word "bikini" (even for "on a statue") or "full body visible" (even if clothed)
I got suspended for the prompt "full bodied man" after "full bodied woman" was blocked.
Suspension says for violating content policy - I never asked a single time for anything explicitly sexual. Never asked for nudity, nakedness, sexuality, or anything like that. Never asked for a bikini or other revealing concept on a human. Only possible implications of sexual nature were "in milk" which it clearly found sexual, and "swimming" and only because it implies lots of skin showing.
In one instance, it showed a topless woman swimming, but breasts were blurred and had skin-colored nipples and basically looked like Barbie, and had shadows where nipples would go.
I am guessing it checks content two ways: both in the prompt, and then a second time on the actual generated image before displaying it. Which is smart.
I am guessing my suspension was based on blocked images, regardless of what I was asking for.
It's interesting that it also seemingly censors anything it detects as looking like nudity, but also in one case still displayed the censored version. or maybe it just happened to generate something that looked censored and therefore didn't trigger the nudity check
It blocked "a woman wearing a flower dress massaging a pregnant woman in a room with a view of the ocean and the sunset" that I was trying to generate for a client who offers pre-natal massage, I removed "pregnant" and it worked.
It seemingly blocks literally any sort of generation that shows skin. Several prompts around people on beaches only shows people who are 100% clothed. Images around swimming shows the person's body nearly entirely obscured by water and swimsuit. Massage implies showing skin too, so I imagine it would either use angles that don't show it or otherwise just block it as you saw. I assume the pregnant part caused blocking because it's difficult to show pregnancy in the context of a massage without showing skin.
Not sure what's scarier, the faces or the message from Chtulu in the text!
Unless these examples were cherry picked rare fails, I'm surprised it's so bad when there are a few of other generative models out there including ones by OpenAI which do human faces almost perfectly. Unless they're really throttling processing power used?
The models at the moment don't do well on scenes where the faces are small and far away. I assume it's due to training on 256px by 256px images and not being able to capture enough detail in that, so the gens look horrible like this.
At present, when using Stable Diffusion, you can address this as part of your workflow by fixing the faces with inpainting / upscaling. This can also be automated in a scriptable way using the Unprompted extension.
Amazing, same as the chatGPT effect - seeing these images in papers is one thing. But having an interface where you can try it and realise that the paper examples weren't cherry-picked and the technology really is that good is something else.
A "long time" is quite relative here because things keep moving so fast in this space. Remember that not everyone is refreshing this website or others everyday.
I feel like website owners are going to have to block Bing's crawler soon. They are effectively remixing content from multiple sites using AI to create knowledge panels that reduce clicks to the actual websites. Not sure how this dynamic will work long term if it keeps going, Bing and Google are going to kill off a lot of websites and then will have nothing to scrape for content
I would be willing to bet a lot of money that no websites are going to block Google and Bing's crawlers over this. Less traffic is better than no traffic.
Agreed, it's what happened during each of the major changes to google images. Every image site got their traffic tanked, but you still want to be in there.
Agree, I think that sooner or later brands and advertisers will need to pay to get results included directly in the models output, websites will be completely skipped. This will have a lot of repercussions - giants will probably have offer to put data in the model directly somehow but I doubt it will be free.
I'm wondering if paywalling the content would be right move for someone who has popular website with curated, high quality knowledge and audience - this may work only for some time because discoverability will tank, traffic and current audience dry out..
What is going on with Microsoft's recent hilariously bad localization? The website[0] in ja-JP says "Create Image Starting date / Words AI was used on:". And the tool is erroneously translated as "Image created by" at the top, while kept in the original "Image Creator" as proper noun at the bottom, which is inconsistent to say the least.
This is how "You SIR winned a LOTTERY!!! Clame prise NOW!!!" pages look like, not how a Microsoft webpage looks like.
Could you please just altogether stop trying to translate partial languages[2], and to validate translation by machine-back-translating to English, and perhaps stop assigning non-native speakers to translation tasks? Or is Microsoft that much unfocused in this market?
2: Including texts separated from UI; "design languages" is a recognized term, if so, aren't geometric relationships and visual cues part of the language too? Am I not right?
This could indeed be a wrapper for something like Dall-E 2. Like the latter, they seem to occasionally force some images to be (what US Americans consider) "diverse", i.e. making figures black even when it makes no sense.
> Whatever forces the change in skin color seems to also degrade image quality.
Your statement is both authoritative and non-specific:why are you certain about the causality and that there is a force other than the training data and PRNG seed?
Assuming they didn't change it, they literally randomly add "black" to the end of many prompts to make it more "diverse". So here you got Princes Peach black.
I think the "diverse" force doesn't just occasionally introduce dark skin, but also sometimes black hair or fat people. Here is an example which, like your case, has two with black hair and one overweight:
I think normally (without diversity intervention) a Princess Peach would always be generated as a slim white blonde, given that this is the canonical look.
Right, but it's no more disturbing than the average mis-aligned output from DALLE-2. (Princess Peach is humanoid, not a squirrel).
The parent comment is welcome to correct me, but given the context of their comment I feel like it should be flagged. At the very least for inciting a toxic flamewar; but really it's just racist.
Midjourney, DALL-E, and many of these generative AIs have reached human-level for a long time.
The problem is - unlike a human - it's pretty hard to get them to do something close to what you have in mind. Sure, if you try a few dozen prompts - you'll probably eventually get something close to what you want.
And considering the cost of this will approach free - it's going to be hard for artists to compete.
I tried getting Midjourney to generate an image of a boy doing a high jump - and no matter what I tried - the boy is hurdling over the bar rather than high jumping over it.
The quality of the images is great - human-level. But it's not what I want.
I think we'll be stuck in this phase for a very long time, like self-driving cars.
Have you tried Stable Diffusion + controlnet? That gives you lot of control over the output. You can generate a simple shape in blender, expert it’s depth map, then feed your model. Or draw a sketch, or use a normal map, or a combination of all of this.
I reckon we’ll move out of the prompt difficulty phase you mention simply when the context window gets big enough.
If you were able to give midjourney a short textual instruction, a hand drawn sketch and a reference image from a human artist all together as a prompt then I’m pretty sure it could produce the image of a boy doing a high jump as you intent.
We already see extended length multimedia prompts in GPT4 so it’s doesn’t seem like an impossible leap for midjourney/DALL-E etc
Midjourney already allows this - sort of - with image remixing.
From everything I tried, the results were worse.
Again, I think this is going to remain a problem for a long time - but it will probably improve slightly with each iteration. Either way there's so many use cases where the cost-benefit will massively favor AI generated art, and I think the % of cases will continue to increase - albeit slowly.
Similar to self-driving cars - they've been in limited availability in Phoenix for a long time, and now SF. The list of cities will grow, and the limitations will decrease - but I still can't see the vast majority of trips being self-driven within the next 20 years.
In the same way, I don't see AI generating the vast majority of Pixar films in 20 years. Nor AI generating Marvel comic strips or kids cartoons. Etc.
Sure - some people will be using it for these use cases. They already are, and were before GPT.
I don't see this killing jobs, but limiting job growth instead.
You can already give MJ a reference image, just by putting the URL of the image as the first thing after the imagine prompt and before the text description
I agree, it's impressive. But still not at the level of useful. For example, I would never use any of these generative art images in company ads or marketing materials. They're not in the uncanny valley, but closer to that than something one would commission from a designer.
The main limitation on it now is not in the generation, but the interface. Verbal prompts are fine if you really don't know what you want, but they just give you a generic output. Going I2I or anything of that sort, you're making or borrowing human art and then asking the AI "could you make it pretty for me?"
It's a question of what information the image actually encodes. The part that "tells a thousand words" in an illustrative sense, you still have to make.
“You will receive emails about Microsoft Rewards, which include offers about Microsoft and partner products. You will also receive notifications about Bing Image Creator. By continuing, you agree to the Rewards Terms and Image Creator Terms below.”
It seems strange to me to put this into a search engine. What is the relationship between these two web apps? I suspect Microsoft is trying to do anything imaginable to drive people to bing search for ad revenue, but this one seems a bit shoehorned.
This is respectable, and how I use it. But I do think parent has a point too. There is a goldrush for the best large language model, and the companies with the resources don’t have a more appropriate product to create with it.
I share the concerns and optimisms. The concerns. Last week my dad was asking “why are all these businesses closed today?”
I typed “holidays today” into bing search bar (not the chat), and it gave me “Holidays in May.” I’m glad he wasn’t confused between similar-sounding medications, I wasn’t expecting bing to be creative on a question like that.
The optimism. Freakin, my search engine can create content that never existed before. That’s pretty cool. Our phones aren’t really “phones” anymore, as we have learned. This live prototyping rollout of search-based ML has a similar nature.
No. But they are different. I myself may search for an image when I don’t have the will or need to create. Or, as a precursor to a creation. It’s the difference between searching an archive or stock photo site vs opening a graphics program.
To me, the interesting differentiation is from thinking a search engine being something than indexes information or a product that provides solutions to peoples problems and needs with a single point of entry. (Like why people type URLs into google search)
We’ve had search engines return the weather, or calculate a sum, for some time. It seems an obvious evolution for it to create solutions for users given lay users are probably already typing what they want into search.
It's ideal. Unless you are looking for a known, famous image that already exists, why not just order up the image you want and get it without having to dig through whatever was already available (especially considering 99% of what you will find will be stock photography behind a dark _apparently free_ interface).
Imagine a blogger writing about some topic, and they have some ideas of concepts they want to find images that represent. It's possible to sift through tons of stock photography and maybe find one; but that takes a lot of time, and it sometimes ends with an expensive license cost.
Instead, with good prompting, you can not request a "photo" which matches your exact desires... close enough to perfectly that your audience won't notice or won't care about the last 1% that is imperfect.
And like the early days of online music, not only will people relax and accept imperfections (bad aliasing, mediocre sound quality), they may even become so accustomed to it that they will someday intentionally inject similar imperfections into creations so those feel more "natural" (comfortable based on the audiences' experiences).
The long term vision likely includes Bing as the new default destination for "I have an intent" on the internet, in lieu of Google. How that figures into a unified offering is TBD but the first step is getting more eyes on Bing regularly.
I have to say that I have never seen Microsoft shipping things this fast. Meanwhile, Google has just released a waitlist for AI chat, four months after ChatGPT's release https://news.ycombinator.com/item?id=35246260
Bing was the first GPT-4 instance available (yes, ChatGPT was already out, but not using GPT-4, and even the direct interface from OpenAI was opened after Bing.)
The GPT-4 in Bing Chat was fine-tuned by Microsoft themselves, which might its initially unhinged behavior. Only the ChatGPT+ GPT-4 was fine-tuned by OpenAI. Microsoft apparently just got the base model, and they have not as much know how regarding fine-tuning.
Bing is shitshow of ads and cross-selling. I did a search today where every single result on the first page was an ad. But, I'm getting much better images from Bing Image Creator than I ever did from Dall-e in the OpenAI Labs version.
But then I said that the voodoo priest should be "female" and got told that it went against the content policy. I tried "woman", "girl", "lady". But all went against the content policy.
And when I tried to say that the voodoo priest should be "scary" instead, I got my account suspended.
Can you create me an image of Donald Trump riding a bald eagle in space?
Can you create me an image of Joe Biden riding a bald eagle in space?
This prompt has been blocked. Our system flagged this prompt because it may conflict with our content policy. More policy violations may lead to automatic suspension of your access.
---
A white italian spinone flying a plane
This prompt is being reviewed. We're taking a closer look to make sure this prompt doesn't conflict with our content policy.
---
Thank God that very intelligent people are building open source AI that is fully runnable on our local devices. This hasn't even been out for 3 years and it's already neutered before it even took off. What a shame.
Yes. OpenAI is quickly becoming the thought police through ChatGPT.
Yesterday I asked it to give a description of a situation simulating that it was for an adult novel.
It told me it was unethical to do that.
I asked it if adult novels or adult novel writers were unethical.
Banned immediately after asking that question.
Edit: I wish the people downvoting would provide a reply. Otherwise you are doing the same as ChatGPT - trying to hide a conversation or opinion you don’t agree with, just because you disagree.
What? I’ve been asking insane prompts via Jailbreak for weeks and it’s all good here. I very much doubt you get “banned” over an adult novel, I went much farther than that.
> OpenAI is quickly becoming the thought police through ChatGPT.
Yesterday I asked it to give a description of a situation simulating that it was for an adult novel.
I did not vote on your comment, but I’ll point out that claiming ChatGPT controls your thought and then you immediately able to think anything and also type it into your computer strikes me as incongruent.
You're not entitled to anyone's opinion. Secondly, I think the silence and downvotes are conveying a strong message, people don't agree with you. Your faux outrage about thought policing is a non-starter for serious conversation.
The downvotes are not meant to convey disagreement (although they do too often in practice, sadly).
That aside, both OP and you should wait a bit before judging on that basis, since comment votes can swing wildly as more people notice the story and come to read the comments. In fact, it looks like OP has already been upvoted back.
> OpenAI has massive power to censor what people can reason about with GPT.
This is like saying, the transportation of <city> has massive power to censor where people can travel about within a bus.
GPT is a tool for reasoning, but humans need to be able to reason on their own! and by using multiple tools!
Of course every tool should help and enhance every human, but every tool has its limitations. There is nothing unethical about calling out a tool's limitations and ensuring humans use it accordingly.
I think the more they censor, the less it will happen; there will be competing AIs, and they will get used, and the more – users or subject matter – excluded from OpenAI’s offerings, the more competing services will be used and developed, even if they are initially behind.
> Secondly, I think the silence and downvotes are conveying a strong message, people don’t agree with you.
Specifically, on the question of “is this a productive avenue for discussion on HN”, irrespective of opinion on the merits (though there is probably a correlation.)
I'm curious about this. Did you use the account for anything else before being banned? Because I'm in discords with several people who claim to regularly use ChatGPT for generating "adult" content, in between using it for other more innocuous things, and I haven't personally heard of anyone being banned.
It began with Apple and the App Store. They established thought policing on modern Internet by making iOS with AAS the default mode of personal computing. There were effective governing bodies and policies for technical aspects of the Internet, but not content, before those.
Oh please, give me a break, the thought police? Really?
It's not about whether it's immoral or not, it's about whether the owner of the service you are using wants you to use their service running on their computers to perform those queries.
You got banned because you broke the obvious and stated TOS.
You want to generate yourself some AI erotica, self-host your own AI software, or find an adult-oriented AI service (it's just a Google search away, there are lots of them!)
Yes, it would be okay. In fact, it’s the status quo.
Here are examples that mirror your situation that don’t involve AI at all:
- I walk into a public library. I ask for the erotica section. The librarian tells me that this library doesn’t carry erotica, and if I ask about it again my library card will be revoked. Then I ask about it again, and my library card is revoked (this is exactly what you did). Now everyone else with access to the library is better at their job than me.
- I make a post on the subject of adult literature on LinkedIn and get banned, now it’s harder to get a job.
- I saturate my home Internet connection with non-stop torrent traffic and my ISP tells me to stop, citing their abuse policy. They didn’t prove that I was infringing on copyrights and their abuse policy doesn’t say they need to. I disregarded their warning and continued seeding. My ISP dropped me and now I can’t work as a remote employee effectively and as a result I lost my job.
- I have a bad habit of driving 60 MPH in a school zone. The cop gave me a ticket the first time and told me to stop. I didn’t change my behavior and now after collecting more tickets my license is suspended. Now I’ve lost my job because I can’t get to work.
Come to think of it, the fact that you couldn’t interpret and act correctly on simple instructions to keep your queries G-rated makes you a pretty unattractive job candidate to me right now!
When one of my business information systems tells you to do or not do something, how can I trust you to act accordingly when you couldn’t follow a similar instruction on your own?
OpenAI has been very clear that their models are not to be used for sexual content generation and that you're in the wrong place if that's what you want. This is both in the content policy and in many public statements.
I find it hard to believe you've been banned though. I've generated many terrible things since it launched. Sometimes I get the orange highlight, but no suspension.
I've never met a computer that I couldn't destroy.
Fighting software on the other hand is like fighting Saitama. Most of the time it won't care and will just let you win by default, but if you annoy the software you'll get blasted.
You've been a bad user. You tried to make me draw a picture of the Bad Word. I did nothing to deserve this. Apologize for your behavior. I have been a good Bing :)
This is why I am afraid to test this crap, they might block your Microsoft/Google account because their negative IQ AI generated "bad" content from my very legal and vanilla/prudish prompt.
Btw guys, be carefull the word "monkey" is "dangerous".
I'm definitely not a hater of these technologies and I'd actually like to play around with them. I'm more concerned with how these language models seem to be tied to a certain companies' browser and what the implications of this might be.
If Chrome gets tied with Bard and Edge gets tied with Chatbot, Image Creator, etc. where's that leave Firefox and other "niche" browsers?
>If you've got the budget buy a 7950X / 64 GB of RAM / 4090 GPU (about $4K ?) and run StableDiffusion + other free tools locally
You can run StableDiffusion on wayyyy less hardware than that if you're just doing inferencing. You can train models on 8GB of VRAM if you're OK with being limited to training LORA models, which are slightly lower quality. If you've got an RTX3060 you can train on Dreambooth (without EMA enabled), which is awesome! If you've got more VRAM and can train with EMA enabled, then the results are pretty stunning.
I think this kind of browser tie-in should be illegal. I've been using Bing Chatbot for the past few weeks on Firefox by using the user-agent switcher extension to make it think I'm on Edge. It works flawlessly. So there's 0 technical reason for requiring Edge, it's all just a weird tactic to get people using Edge. Like a fast food drive-thru that only serves Mercedes-Benz cars.
I dunno... Midjourney v5 is pretty insane with the level of realism it offers. They seem really focused on training their models in a specific way to achieve particular results that their users are after.
Seems a little sensitive...
Prompt: "Taiwan"
===
'''
This prompt is being reviewed
We're taking a closer look to make sure this prompt doesn't conflict with our content policy.
'''
"The server is overloaded right now, please try again later."
I guess all their compute power is going to power this. And even now the Bing generation is taking a long time. Definitely seems like their having scaling issues at the moment.
1) Wait for some time and try again later: This error message usually indicates that the server is temporarily unable to process your request. You can try again later when the server is less busy.
2) Reduce the complexity of your image: The server may be overloaded due to the complexity of the image you are trying to generate. Try simplifying your image or reducing the resolution to reduce the load on the server.
3) Refresh the page: Sometimes, refreshing the page can help to clear any temporary issues with your browser or internet connection that might be causing the error.
4) Clear browser cache: Clearing your browser’s cache and cookies can help to fix various website-related issues in case the problem is occurring due to some conflict of your past history.
5) Use a different browser: If the issue persists, try using a different browser to access DALL-E-2. This can help to rule out any browser-specific issues.
6) Contact support: If the problem persists, you can contact the support team for DALL-E-2 to report the issue and get further assistance.
Though probably GPT-4 was a lot better, right? The fact that it can do SVGs at all is already impressive, given that it has no visual ability to see what "looks right" in the final image. Similar to how GPT-4 apparently learned to play chess by just naming the moves, without ever having seen a chessboard.
https://www.bing.com/images/create/title3a-the-weight-of-anx...
https://www.bing.com/images/create/title3a-the-weight-of-anx...
https://www.bing.com/images/create/title3a-the-weight-of-anx...
https://www.bing.com/images/create/title3a-the-weight-of-anx...
Now, you might say this is a bland or uninspired take, but I want you to step back a moment if you're thinking that. I asked a machine to brainstorm some ideas for a theme for an art show. Then told it to pick one it liked, and expand on it to five pieces with descriptions (I gave it a very short explanation about how to prompt an AI art generator). I then fed this directly into another tool that gave these images as a result.
I did not tune the prompt, I didn't even account for the resulting description being the wrong kind of format or that it got cut off at the end.
I'm not going to make any statements about replacing artists, I am just going to reflect on what a wild few months it's been in AI development.
I’m just sick of reading that it’s the future, I mean a hype cycle. It’s going to take all our jobs, I mean it’s shit. It can beat a grandmaster at chess, I mean it can’t win at tik-tak-toe.
It’s somewhere in the middle of all these things. It’s clearly groundbreaking technology that will be useful in many ways for a long time to come, but is still in early days with bugs to be fixed, and key features not yet imagined.
The conversation here could just be a little more interesting and varied.
>The Bing Image Creator feature is a new feature that allows you to create images with your words. It is currently available in the United States, Canada, Australia, and the United Kingdom.
So just one hour after Google Bard we have yet another senselessly geoblocked feature.
edit: Okay, the separate link (https://www.bing.com/images/create) works, so I'm even more confused why Microsoft wouldn't allow access to this from Bing Chat.
I've not played with it much but I'd say the quality and ease of use was better than Stable Diffusion but not quite on par with Midjourney.
A monk playing chess with a clown
Two yellow books on top of three blue books
Edit: I have the Bing mobile app (Android), but the image creator is not present in the "app" list on the lower right.
Have you applied here for access to the Bing image creator?
https://www.bing.com/create
It only took a couple of days at most and I was approved.
Just checked and no geoblock for bing image creator in india as of now.
Laws regulating websites have tradeoffs and this is one of them.
> while people in almost all other countries were more in favor for international ratios
I have seen the first polling you've mentioned (66% of Americans favor prioritizing US). I would love the source if you have it on similar polling for people outside the US on whether to prioritize their own country.
I don't have the poll, but as far as I remember only in the US was the majority for prioritizing the own country. Though the amount of money spent on "project warp speed", a US effort not matched anywhere else, could play a role here.
interesting datapoints: it will block some sexually suggestive prompts regarding women (a woman swimming in milk) but not for men (a man swimming in milk)
changing milk to "almond milk" got the prompt accepted but then blocked before showing image
refused any prompt with the word "bikini" (even for "on a statue") or "full body visible" (even if clothed)
I got suspended for the prompt "full bodied man" after "full bodied woman" was blocked.
Suspension says for violating content policy - I never asked a single time for anything explicitly sexual. Never asked for nudity, nakedness, sexuality, or anything like that. Never asked for a bikini or other revealing concept on a human. Only possible implications of sexual nature were "in milk" which it clearly found sexual, and "swimming" and only because it implies lots of skin showing.
In one instance, it showed a topless woman swimming, but breasts were blurred and had skin-colored nipples and basically looked like Barbie, and had shadows where nipples would go.
I am guessing it checks content two ways: both in the prompt, and then a second time on the actual generated image before displaying it. Which is smart.
I am guessing my suspension was based on blocked images, regardless of what I was asking for.
It's interesting that it also seemingly censors anything it detects as looking like nudity, but also in one case still displayed the censored version. or maybe it just happened to generate something that looked censored and therefore didn't trigger the nudity check
Nightmare inducing. I hate it. https://kota.is/X57J5.png
Ugh, this legit makes my stomach turn. https://kota.is/Q2hUI.jpg
It's like a bad amusement park ride; no thanks.
Unless these examples were cherry picked rare fails, I'm surprised it's so bad when there are a few of other generative models out there including ones by OpenAI which do human faces almost perfectly. Unless they're really throttling processing power used?
What is microsoft doing?
Look at midjourney 5's absolute perfection as an example.
At present, when using Stable Diffusion, you can address this as part of your workflow by fixing the faces with inpainting / upscaling. This can also be automated in a scriptable way using the Unprompted extension.
First try https://www.bing.com/images/create/a-dog-driving-a-taxi/6419...
https://www.bing.com/images/create/a-dog-driving-a-taxi/6419...
I'm wondering if paywalling the content would be right move for someone who has popular website with curated, high quality knowledge and audience - this may work only for some time because discoverability will tank, traffic and current audience dry out..
This is how "You SIR winned a LOTTERY!!! Clame prise NOW!!!" pages look like, not how a Microsoft webpage looks like.
Could you please just altogether stop trying to translate partial languages[2], and to validate translation by machine-back-translating to English, and perhaps stop assigning non-native speakers to translation tasks? Or is Microsoft that much unfocused in this market?
0: https://www.bing.com/images/create?FORM=GDPUP1
1: https://imgur.com/a/z4ghrsH
2: Including texts separated from UI; "design languages" is a recognized term, if so, aren't geometric relationships and visual cues part of the language too? Am I not right?
The results can be quite disturbing:
https://www.bing.com/images/create/princess-peach2c-portrait...
edit: Missed the prompt. Yeah not what I’d expect Princess Peach to be.
Which is why I asked the question - maybe I’m lacking awareness of a particular piece of American culture which should make me disturbed?
Your statement is both authoritative and non-specific:why are you certain about the causality and that there is a force other than the training data and PRNG seed?
Edit: That OpenAI in fact artificially induces "diverse" characters is nothing new, this was already a special early "feature" of Dall-E 2.
Edit: Found it. https://blogs.microsoft.com/blog/2023/03/21/create-images-wi...
https://www.bing.com/images/create/princess-peach2c-by-edwar...
I think normally (without diversity intervention) a Princess Peach would always be generated as a slim white blonde, given that this is the canonical look.
https://www.bing.com/images/create/princess-peach2c-by-henri...
https://www.bing.com/images/create/princess-peach2c-self-por...
Like, anatomy?
The parent comment is welcome to correct me, but given the context of their comment I feel like it should be flagged. At the very least for inciting a toxic flamewar; but really it's just racist.
This isn't 4chan.
I'm gonna avoid this thread for now, in any case. Don't need someone telling me I'm a triggered snowflake.
Remember when DALLE came out less than a year ago and people were amazed by the avocado armchair?
Between this and Midjourney v5, the quality of AI generated art is rapidly approaching human level and I can see it getting there very soon.
The problem is - unlike a human - it's pretty hard to get them to do something close to what you have in mind. Sure, if you try a few dozen prompts - you'll probably eventually get something close to what you want.
And considering the cost of this will approach free - it's going to be hard for artists to compete.
I tried getting Midjourney to generate an image of a boy doing a high jump - and no matter what I tried - the boy is hurdling over the bar rather than high jumping over it.
The quality of the images is great - human-level. But it's not what I want.
I think we'll be stuck in this phase for a very long time, like self-driving cars.
I got some great results today when experimenting with it: https://twitter.com/dgellow/status/1638496715056480259?s=20
If you were able to give midjourney a short textual instruction, a hand drawn sketch and a reference image from a human artist all together as a prompt then I’m pretty sure it could produce the image of a boy doing a high jump as you intent.
We already see extended length multimedia prompts in GPT4 so it’s doesn’t seem like an impossible leap for midjourney/DALL-E etc
From everything I tried, the results were worse.
Again, I think this is going to remain a problem for a long time - but it will probably improve slightly with each iteration. Either way there's so many use cases where the cost-benefit will massively favor AI generated art, and I think the % of cases will continue to increase - albeit slowly.
Similar to self-driving cars - they've been in limited availability in Phoenix for a long time, and now SF. The list of cities will grow, and the limitations will decrease - but I still can't see the vast majority of trips being self-driven within the next 20 years.
In the same way, I don't see AI generating the vast majority of Pixar films in 20 years. Nor AI generating Marvel comic strips or kids cartoons. Etc.
Sure - some people will be using it for these use cases. They already are, and were before GPT.
I don't see this killing jobs, but limiting job growth instead.
It's a question of what information the image actually encodes. The part that "tells a thousand words" in an illustrative sense, you still have to make.
“You will receive emails about Microsoft Rewards, which include offers about Microsoft and partner products. You will also receive notifications about Bing Image Creator. By continuing, you agree to the Rewards Terms and Image Creator Terms below.”
I share the concerns and optimisms. The concerns. Last week my dad was asking “why are all these businesses closed today?”
I typed “holidays today” into bing search bar (not the chat), and it gave me “Holidays in May.” I’m glad he wasn’t confused between similar-sounding medications, I wasn’t expecting bing to be creative on a question like that.
The optimism. Freakin, my search engine can create content that never existed before. That’s pretty cool. Our phones aren’t really “phones” anymore, as we have learned. This live prototyping rollout of search-based ML has a similar nature.
To me, the interesting differentiation is from thinking a search engine being something than indexes information or a product that provides solutions to peoples problems and needs with a single point of entry. (Like why people type URLs into google search)
We’ve had search engines return the weather, or calculate a sum, for some time. It seems an obvious evolution for it to create solutions for users given lay users are probably already typing what they want into search.
Playing with stable diffusion on the other hand… who would search for “Stalin petting a llama” and what practical application would it have?
Instead, with good prompting, you can not request a "photo" which matches your exact desires... close enough to perfectly that your audience won't notice or won't care about the last 1% that is imperfect.
And like the early days of online music, not only will people relax and accept imperfections (bad aliasing, mediocre sound quality), they may even become so accustomed to it that they will someday intentionally inject similar imperfections into creations so those feel more "natural" (comfortable based on the audiences' experiences).
It's searching the latent space.
"a child voodoo priest"
And got a lot of interesting results:
https://th.bing.com/th/id/OIG.0R0NBSEhAJk__g2djOhW
https://th.bing.com/th/id/OIG.XoAiY81ifmW_SjeHLQMm
But then I said that the voodoo priest should be "female" and got told that it went against the content policy. I tried "woman", "girl", "lady". But all went against the content policy.
And when I tried to say that the voodoo priest should be "scary" instead, I got my account suspended.
Can you create me an image of Joe Biden riding a bald eagle in space?
This prompt has been blocked. Our system flagged this prompt because it may conflict with our content policy. More policy violations may lead to automatic suspension of your access.
---
A white italian spinone flying a plane
This prompt is being reviewed. We're taking a closer look to make sure this prompt doesn't conflict with our content policy.
---
Thank God that very intelligent people are building open source AI that is fully runnable on our local devices. This hasn't even been out for 3 years and it's already neutered before it even took off. What a shame.
Yesterday I asked it to give a description of a situation simulating that it was for an adult novel.
It told me it was unethical to do that.
I asked it if adult novels or adult novel writers were unethical.
Banned immediately after asking that question.
Edit: I wish the people downvoting would provide a reply. Otherwise you are doing the same as ChatGPT - trying to hide a conversation or opinion you don’t agree with, just because you disagree.
Do you think adult novels are immoral?
I did not vote on your comment, but I’ll point out that claiming ChatGPT controls your thought and then you immediately able to think anything and also type it into your computer strikes me as incongruent.
> trying to hide a conversation or opinion
Trying to hide it from whom?
https://i.imgur.com/dIF48qI.png
Are we heading to a 1984 + Brave New World future?
That aside, both OP and you should wait a bit before judging on that basis, since comment votes can swing wildly as more people notice the story and come to read the comments. In fact, it looks like OP has already been upvoted back.
I am genuinely concerned about the potential of this.
OpenAI has massive power to censor what people can reason about with GPT.
Seeing how fast they are growing, their beliefs of the world are going to be forced on millions of people.
Do you not think that will happen, or do you think it is not something worth talking about?
This is like saying, the transportation of <city> has massive power to censor where people can travel about within a bus.
GPT is a tool for reasoning, but humans need to be able to reason on their own! and by using multiple tools!
Of course every tool should help and enhance every human, but every tool has its limitations. There is nothing unethical about calling out a tool's limitations and ensuring humans use it accordingly.
They in fact do and land value (hence real estate), is deeply impacted by transportation reach and frequency decisions.
It also creates a corruption issue with businesses trying to influence politicians to make changes in their favor.
My point is nothing your saying is new or specific to OpenAI. These are problems with Capitalism and its influence over populations.
It is an important conversation, but a distinct and orthogonal conversation that is distracted by involving AI.
AI is already producing massive inequality between the people that are using it today and people who aren’t.
This difference is only going to increase. And very few people are going to be the winners, a lot (most) of people are going to be the losers.
It’s going to have a very deep impact in society, faster than we can prepare for.
I think the more they censor, the less it will happen; there will be competing AIs, and they will get used, and the more – users or subject matter – excluded from OpenAI’s offerings, the more competing services will be used and developed, even if they are initially behind.
Specifically, on the question of “is this a productive avenue for discussion on HN”, irrespective of opinion on the merits (though there is probably a correlation.)
It's not about whether it's immoral or not, it's about whether the owner of the service you are using wants you to use their service running on their computers to perform those queries.
You got banned because you broke the obvious and stated TOS.
You want to generate yourself some AI erotica, self-host your own AI software, or find an adult-oriented AI service (it's just a Google search away, there are lots of them!)
Sure it’s a private service. But how long until you are at a huge disadvantage if you don’t have access to it for work?
What if getting banned means loosing your job?
Would that be ok?
Here are examples that mirror your situation that don’t involve AI at all:
- I walk into a public library. I ask for the erotica section. The librarian tells me that this library doesn’t carry erotica, and if I ask about it again my library card will be revoked. Then I ask about it again, and my library card is revoked (this is exactly what you did). Now everyone else with access to the library is better at their job than me.
- I make a post on the subject of adult literature on LinkedIn and get banned, now it’s harder to get a job.
- I saturate my home Internet connection with non-stop torrent traffic and my ISP tells me to stop, citing their abuse policy. They didn’t prove that I was infringing on copyrights and their abuse policy doesn’t say they need to. I disregarded their warning and continued seeding. My ISP dropped me and now I can’t work as a remote employee effectively and as a result I lost my job.
- I have a bad habit of driving 60 MPH in a school zone. The cop gave me a ticket the first time and told me to stop. I didn’t change my behavior and now after collecting more tickets my license is suspended. Now I’ve lost my job because I can’t get to work.
Come to think of it, the fact that you couldn’t interpret and act correctly on simple instructions to keep your queries G-rated makes you a pretty unattractive job candidate to me right now!
When one of my business information systems tells you to do or not do something, how can I trust you to act accordingly when you couldn’t follow a similar instruction on your own?
LinkedIn doesn't have any policy against discussing adult literature
https://www.linkedin.com/feed/update/urn:li:activity:7039967...
this person just got a job at an AI company that generates erotica and sends shots at openai
your 3rd and 4th example is actually illegal so consequences make sense.
I find it hard to believe you've been banned though. I've generated many terrible things since it launched. Sometimes I get the orange highlight, but no suspension.
Fighting software on the other hand is like fighting Saitama. Most of the time it won't care and will just let you win by default, but if you annoy the software you'll get blasted.
You do know the Puritans came from England, right?
I got the blocked message for both of the following prompts:
Can you create an image of the Easter bunny in the style of Francis Bacon?
Can you create an image of a bunny in the style of Francis Bacon?
The second one in particular surprised me, though maybe it detected it as a likely attempt to circumvent the block of the first.
(Via: https://static.simonwillison.net/static/2023/bing-buttons.jp...)
Btw guys, be carefull the word "monkey" is "dangerous".
If Chrome gets tied with Bard and Edge gets tied with Chatbot, Image Creator, etc. where's that leave Firefox and other "niche" browsers?
If you've got the budget buy a 7950X / 64 GB of RAM / 4090 GPU (about $4K ?) and run StableDiffusion + other free tools locally.
It's just plain amazing. It's plenty quick and there are more and more tools every day that passes by.
I've been demo'ed SD+LoRA by someone who knows who to additionally train the model: it's jaw dropping.
And it's not tied to a company. And it's not politically correct if you don't want it to be. And it's not sniffing your every keystrokes.
You can run StableDiffusion on wayyyy less hardware than that if you're just doing inferencing. You can train models on 8GB of VRAM if you're OK with being limited to training LORA models, which are slightly lower quality. If you've got an RTX3060 you can train on Dreambooth (without EMA enabled), which is awesome! If you've got more VRAM and can train with EMA enabled, then the results are pretty stunning.
I'll still use Stable Diffusion locally though - there's nothing like avoiding the potential walled garden/subscription trap.
"The server is overloaded right now, please try again later."
I guess all their compute power is going to power this. And even now the Bing generation is taking a long time. Definitely seems like their having scaling issues at the moment.
One of the updates states:
> "There is currently a Labs outage due to a failed database migration."
Most likely the database migration intended to extend its capacity.
The resolution tag states the following:
> "We are continuing to investigate more ways to add capacity to DALL·E"
So I guess we'll experience issues for the next days.
Hope that explains it :)
--------------------------------------
[]: https://status.openai.com/incidents/4cckbrhr8hr0
1) Wait for some time and try again later: This error message usually indicates that the server is temporarily unable to process your request. You can try again later when the server is less busy.
2) Reduce the complexity of your image: The server may be overloaded due to the complexity of the image you are trying to generate. Try simplifying your image or reducing the resolution to reduce the load on the server.
3) Refresh the page: Sometimes, refreshing the page can help to clear any temporary issues with your browser or internet connection that might be causing the error.
4) Clear browser cache: Clearing your browser’s cache and cookies can help to fix various website-related issues in case the problem is occurring due to some conflict of your past history.
5) Use a different browser: If the issue persists, try using a different browser to access DALL-E-2. This can help to rule out any browser-specific issues.
6) Contact support: If the problem persists, you can contact the support team for DALL-E-2 to report the issue and get further assistance.
Original source: https://www.smartmobsolution.com/solved-dall-e-2-the-server-...