71–80 of 163 posts
Very impressive. I hope we get the ability for "meta" things to work, like asking for a rectangular image, a 16x16 spritesheet, etc. Also, not using Bing search, sorry MS.
Is it now only Dalle 3? I remember it was a mix Dalle 2 and 3 like two days ago.
I had incredible results asking for architectural drawings earlier. Then a few minutes ago, I broke down and started prompting for supermodels. It did a terrific job the first few times.
But after like three of them getting blocked (I didn't actually ask for anything inappropriate) it starting giving me something that looked like unmitigated Stable Diffusion 1.5.
Lol.
Bing is kinda desperate it seems. I went to install GPT on my device yesterday and the first app result was a sponsored one - bing - telling can you can earn prizes by using the app. Don't know if they are more interested in growing the number of users or collecting that sweet data. Probably both.
There's an LLM morphing your queries somewhat before submitting to Dall-e and you can jailbreak that. https://twitter.com/madebyollin/status/1708204657708077294 https://media.discordapp.net/attachments/1023643945319792731...
How do you jailbreak it?
genuinely curious - is it hard for an advanced AI model to differentiate the intention of the prompt and then if it's mature content may be not generate the image?
Earlier quoted context omitted.
>Except this happens with every product OpenAI has released. Their own homegrown versions always have more features or are less hamstrung Like ?
ChatGPT currently has voice conversations for all paying subs. DALL-E 2 has inpainting and outpainting. Bing Chat vs paid ChatGPT is no contest: the guardrails Microsoft has put around Bing Chat make it a much worse experience. I use these products every single day of my life. It's night and day.
Right and Bing has had that for a bit. Worse voices and not hands off but the TTS is also a new model so it's not like switching is out of the question.
>DALL-E 2 has inpainting and outpainting
No one is saying Dalle-3 won't have inpainting/outpainting but there's zero guarantee you will be able to do that from cGPT. You'd think they'd demo that if they were but we'll see.
Paid users have been stuck on 4k context even though the model has more. Bing has been running the longer context model from the beginning with web page/pdf features. Bing also recently introduced "personalized answers" which is just retrieval augmented memory over all your chats, a feature cGPT really should have had by now.
There's an LLM morphing your queries somewhat before submitting to Dall-e and you can jailbreak that. https://twitter.com/madebyollin/status/1708204657708077294 https://media.discordapp.net/attachments/1023643945319792731...
`baby girl playing with a rabbit realistic image` prompt gave `Unsafe image content detected`. anyone else facing the same? looks like the content safety policies are too aggressive. genuinely curious - is it hard for an advanced AI model to differentiate the intention of the prompt and then if it's mature content may be not generate the image?