Microsoft Brings OpenAI's DALL-E Image Creator To the New Bing (techcrunch.com) 28
Microsoft today announced that its new AI-enabled Bing will now allow users to generate images with Bing Chat. From a report: This new feature is powered by DALL-E, OpenAI's generative image generator. The company didn't say which version of DALL-E it is using here, except for saying that it is using the "very latest DALL-E models." Dubbed the "Bing Image Creator," this new capability is now (slowly) rolling out to users in the Bing preview and will only be available through Bing's Creative Mode. It'll come to Bing's Balanced and Precise modes in the future.
The new image generator will also be available in the Edge sidebar. The right prompts will generate the now-familiar square of four high-res DALL-E images. There's one major difference, though: there will be a small Bing logo in the bottom left corner. The early Bing AI release was missing a few guardrails, but Microsoft quickly fixed those. The company is clearly hoping to avoid these issues with this release.
The new image generator will also be available in the Edge sidebar. The right prompts will generate the now-familiar square of four high-res DALL-E images. There's one major difference, though: there will be a small Bing logo in the bottom left corner. The early Bing AI release was missing a few guardrails, but Microsoft quickly fixed those. The company is clearly hoping to avoid these issues with this release.
The driving factor behind adoption of any new tech (Score:2)
Re: (Score:2, Funny)
Re: (Score:3)
I was originally replying to call you a sick freak, but I now see you said anime, not animal. All good.
Re: (Score:2)
When they're done pimping Bing (Score:3)
maybe Microsoft could try fixing the engine. You know, for people who use a search engine for serious purposes. Now that would be novel...
Re: Why doesn't MS make it do serious things (Score:1)
But mirroring all the crap and creeps on the Intertubes is the lower hanging fruit because there's much more of it to bot-train with. Smart & Sane is more expensive.
By the way, where's CatGPT where you video-chat with cute cats? That's gotta be popular, and people wouldn't expect cats to be a smart as humans.
"You bought an outdated bot, silly head!" (Score:1)
Dall-E is yesterday's tech, everyone wants ChatGPT now. Thus, they probably got a better licensing deal. (Where is the merged DallGPT? by the way)
(Title ref: https://www.youtube.com/watch?... [youtube.com] )
Re: (Score:2)
Re: (Score:1)
It amazes me how much AI is improving without having common sense. Bots can now stick to a theme in chats, creepy ones at that [theguardian.com]. Maybe we can get practical AI withOUT adding in common sense?
I was thinking that if neural-net-based bots could be coordinated with something like Cyc, it could finally have common sense. To the best of my knowledge, the merge experiments are not going so well...yet. Maybe they can use a genetic algorithm to breed better coordination rather than try to figure out how to merge them.
Re: (Score:2)
The very peak of "dumb"? (Score:1)
> They can't create or understand ideas,
They can indeed create. Just not "understand". However, my point is that maybe bots will be able to do more than we thought WITHOUT "understanding" (as we know it). We don't know the upper limit on usefulness without the understand part. In other words, the max usefulness level without "understanding" why things are the way they are. Maybe there's a better name for this concept? The level might be higher than we think.
I'm sure we are not at the peak of Understandi
Re: (Score:2)
Re: (Score:2)
Maybe we can get practical AI withOUT adding in common sense?
I guess it depends on what you mean by "practical AI", since even simple automation is definitely valuable.
Copyright free stock images (Score:3)
May be some legal liability here (Score:1)
Since DALL-E is prone to include some direct aspects of copyrighted work, I wonder if that would open up Microsoft for copyright infringement lawsuits.
Re: (Score:2)
Indeed, training images are represented in weights and can often be retrieved from them with negligible changes. The artificial neural net contains a lossy compressed version of the corpus. DMCA gives some limited exemptions for caching, but this goes well beyond that.
OpenAI could get away with it by calling it research, but Microsoft have almost certainly added a separate trained AI to detect images too close to a restrictive license corpus image, which just removes those from the output candidates. It doe
Re: (Score:2)
>can often be retrieved from them with negligible changes
"Often"? I do not think that word means what you think it does.
To retrieve a decent likeness of an image used in training, the image has to be severely over-represented in the training corpus. There have been a very small number of such images identified at this point (less than 200 so far, last I heard), and we can expect a few handfulls more. That's it, out of training sets ranging from 600k to billions.
Images that are only in the training set on
Glaze will keep em honest (Score:1)
That should keep art thieves & plagiarism at bay (for a while)
Re: (Score:2)
Glaze is useless: it has a horribly high error rate going both ways. People have been posting tests on Reddit and it seems like it is good at detecting some images from some models, and useless at everything else, AND it way too often marks human-generated art as AI.
It's like the ChatGPT detection systems: too faulty to be used for anything good, and most likely to lead to a lot of mistaken accusations and heartache...and possibly lawsuits.
Didn't run straight from the chat for me (Score:3)
I couldn't get it to generate images right from Bing "Chat". I had to do it from here:
https://www.bing.com/images/cr... [bing.com]
Re: (Score:2, Flamebait)
Re: (Score:2)
Re: (Score:1)
It's not like Bing has a lot of customers to lose. When you are a distant second, gambling looks shinier. Sticking in bot toys may just strike fad oil.
Re: (Score:2)
It censors innocent things almost all the time (Score:2)
It's very sensitive at the moment.
I asked it to create a teddybear balancing on a thightrope balancing a vase on his nose, that one got flagged. I could world it differently and after a few tries using rope instead of thightrope then it got it, but wrong - but it accepted it. Apparantly it didn't like the word thight because the word tightrope was mispelled.
It will also flag and censor words that contain body, for example I asked it to create a man with the face of a VHS player with the body of a butler cle
Can it do porn? No? (Score:2)
Then I am not interested. Porn is probably the only real use-case for this.