Can’t wait for your V6 deep-dive post. From what I’ve seen photorealism is off the charts, but the model still struggles with text and complex prompt direction, at least relative to DALL-E 3.
Yes to photorealism and text not being quite at DALL-E 3 levels. (V6 can do short two-word signs well enough, but anything longer is likely to get scrambled.)
Wow, that is really impressive. I'm also bullish on MJ's ability to improve prompt adherence given that the OpenAI team has published some of their approach on modeling sentence/word fragments and extrapolating longer prompts based on training data.
Yeah for sure. From my deep-dive into the DALL-E 3 research paper, it's clear that there's nothing in OpenAI's approach that others can't replicate. They even list avenues for fixing text rendering and prompt comprehension. So I fully expect things to improve rapidly this year!
I love how goofy it is that an update came in while you were working on this. That's sort of how the singularity is supposed to happen: you just can't keep up with all the updates any more, and they just kind of happen before you can adjust to them.
To be fair, V6 has been in the works for many months and was originally planned for a September launch before being pushed back repeatedly. I knew it was coming out shortly because they were having a "rating party" to review the initial batch of images. It was just a coincidence that it happend on the exact day. Will try to cover it next Thursday, my last post of the year!
Can’t wait for your V6 deep-dive post. From what I’ve seen photorealism is off the charts, but the model still struggles with text and complex prompt direction, at least relative to DALL-E 3.
Yes to photorealism and text not being quite at DALL-E 3 levels. (V6 can do short two-word signs well enough, but anything longer is likely to get scrambled.)
But prompt adherence is pretty great so far. Here's a test I shared yesterday: https://substack.com/@whytryai/note/c-45803449
Wow, that is really impressive. I'm also bullish on MJ's ability to improve prompt adherence given that the OpenAI team has published some of their approach on modeling sentence/word fragments and extrapolating longer prompts based on training data.
Yeah for sure. From my deep-dive into the DALL-E 3 research paper, it's clear that there's nothing in OpenAI's approach that others can't replicate. They even list avenues for fixing text rendering and prompt comprehension. So I fully expect things to improve rapidly this year!
I love how goofy it is that an update came in while you were working on this. That's sort of how the singularity is supposed to happen: you just can't keep up with all the updates any more, and they just kind of happen before you can adjust to them.
To be fair, V6 has been in the works for many months and was originally planned for a September launch before being pushed back repeatedly. I knew it was coming out shortly because they were having a "rating party" to review the initial batch of images. It was just a coincidence that it happend on the exact day. Will try to cover it next Thursday, my last post of the year!
If I'm right about the singularity, they'll release versions 7 through 59 by tomorrow.
Midjourney is Skynet. We'll all be turned into hyper realistic photos before being erased from existence.
Dude.
We ARE the source material.
Duuuuuude.