6 ms·

I'd bet on this being true six months from now. DALL-E 2 came out in 2021-01 (19mo ago) and Midjourney in 2022-07 (13mo ago). While the quality of images has i

by jefftk 3y ago

I'd bet on this being true six months from now. DALL-E 2 came out in 2021-01 (19mo ago) and Midjourney in 2022-07 (13mo ago). While the quality of images has improved a lot, if you have something specific and moderately complex in mind it's still very hard to get there just from a textual prompt.

raincole 3y ago
But the tech did improve a lot in the past year. One year ago we had textual inversion. Now we have LoRA and control net. I know it might not be a scientific breakthrough, but in practice it's day and night.
- jefftk 3y ago
  I'd be happy to bet that the tech will continue to improve. But "producing a specific image with generative AI is sometimes almost impossible" seems very likely to still be true in February 2024.
  - raincole 3y ago
    Depending on your definition of "specific". Of course an AI can not produce an image that matches what's on your mind 100%. Because by definition it will require you to provide all the information, a.k.a. you need to make the image by yourself first. But I really don't think illustrators are as safe as the article implies. Yes, the jobs won't disappear overnight, but are there that much demands for illustration to support a future where every illustrator becomes 10x more productive than before? (I've done illustration commercially before, while it's not my main source of income and I'm junior level at best.)
    - jefftk 3y ago
      Want to try to turn this into something concrete enough that we can bet on?
      - raincole 3y ago
        I was kinda tempted, but I don't think there is such a good metric to measure it. (Since the job market is heavily influenced by interest rate etc... and if I were confident at predicting macro economics I would just bet on stock)
        jefftk 3y ago
        What if we picked 10 prompts that it seems like an AI should be able to depict well, but it can't yet? And then iff the best AI tool in February can do the majority of them you win?
        raincole 3y ago
        I'll bet against it tho. I don't actually believe pure text-prompt-to-image will improve much (not in a few months at least). I just believe there will be more non-text tools to guide AI, like LoRA and control net, and they will be more accessible. Control net kinda did what you said but in a different timeframe: it was quite difficult to tell AI to generate a person "sitting with their legs cross". Today, it's relatively easy to do this with control net, but still hard with text prompt only. Edit: and the sibling comment made me question myself why I would ever take a random bet on the internet.
        jefftk 3y ago
        We started this thread with: sambleckley> producing a specific image with generative AI is sometimes almost impossible justrealist> Who could possibly think this will be the case six months from now? me> I'd bet on this being true six months from now I thought you disagreed with me on this, but it sounds like maybe not?
        raincole 3y ago
        I was just stating the improvement on SD we've seen since DELL-E 2 and Midjourney came out wasn't just about "quality of image", but also about "have something specific and moderately complex in mind". Thus I mentioned textual inversion vs LoRA/ControlNet.
gedy 3y ago
> While the quality of images has improved a lot, if you have something specific and moderately complex in mind it's still very hard to get there just from a textual prompt. I think this gets to an important point. If whoever is paying the bills has something very specific in mind, they won't be happy with AI (or frankly many artists) at this point. But as a creative interpretation of something more general, it's actually really good, and I think with many low-importance works like the author describes as "furniture" - we really don't need to be that exact.
GaggiX 3y ago
Dalle 2 was out April 2022. I don't think it will be true six months from now. OpenAI has been red teaming a new version for months that goes way beyond anything we've seen today. You can see some leaks from this video: https://youtu.be/koR1_JBe2j0 https://youtu.be/koR1_JBe2j0
- jefftk 3y ago
  > Dalle 2 was out April 2022 Sorry, thanks! I accidentally gave the date for the first DALL-E. > OpenAI has been red teaming a new version for months that goes way beyond anything we've seen today. Any sign that it's better at understanding complex textual prompts, and not just at making high-quality images?
  - GaggiX 3y ago
    >Any sign that it's better at understanding complex textual prompts, and not just at making high-quality images? The video I have linked.
- gwern 3y ago
  I don't know if that's 'way beyond'. Those samples look similar to Imagen and Parti in terms of quality and following complicated text prompts. (Look at the PartiPrompts for the absurd prompts Parti can execute accurately.) I'm glad to see OA does have a successor for DALL-E 2, though: the service seems to have somehow gotten worse since release.
  - GaggiX 3y ago
    I shouldn't have said "goes way beyond anything we've seen today", but probably it goes way beyond anything we will be able to try in this 6 months period. Parti 20B is really good at language understanding, more meh Imagen. >the service seems to have somehow gotten worse since release. Yeah, I think it's pretty clear that they probably lowered the number of steps or used a similar strategy to reduce the computation.
iudqnolq 3y ago
But in general the better something is the harder it is to improve. From an outsiders' perspective it looks like we've had a series of achievements where AI worked impressively well considering it isn't a human. But afterwards we got an incremental grind that never got close to an AI that's just good, period. I'm very impressed people got self-driving car demos working. I couldn't do that. But year after year self-driving cars remained a demo, albeit an incrementally improving one.
- jefftk 3y ago
  > year after year self-driving cars remained a demo, albeit an incrementally improving one You can get a ride in Phoenix today. Order a Waymo from your phone, it shows up with no driver. https://waymo.com/waymo-one-phoenix/ https://waymo.com/waymo-one-phoenix/ The number of cities is still small, but it's not a demo anymore.
  - iudqnolq 3y ago
    It's limited to a small number of cities with optimal conditions and every ride has a human supervising. They're still spending R&D, rather than making a profit. That's a demo.
    - jefftk 3y ago
      > every ride has a human supervising What do you mean? Waymo has been running fully driverless operations with the general public since 2020: https://waymo.com/blog/2020/10/waymo-is-opening-its-fully-driverless.html https://waymo.com/blog/2020/10/waymo-is-opening-its-fully-dr...

by jefftk 3y ago

raincole 3y ago

But the tech did improve a lot in the past year. One year ago we had textual inversion. Now we have LoRA and control net. I know it might not be a scientific breakthrough, but in practice it's day and night.

jefftk 3y ago
I'd be happy to bet that the tech will continue to improve. But "producing a specific image with generative AI is sometimes almost impossible" seems very likely to still be true in February 2024.
- raincole 3y ago
  Depending on your definition of "specific". Of course an AI can not produce an image that matches what's on your mind 100%. Because by definition it will require you to provide all the information, a.k.a. you need to make the image by yourself first. But I really don't think illustrators are as safe as the article implies. Yes, the jobs won't disappear overnight, but are there that much demands for illustration to support a future where every illustrator becomes 10x more productive than before? (I've done illustration commercially before, while it's not my main source of income and I'm junior level at best.)
  - jefftk 3y ago
    Want to try to turn this into something concrete enough that we can bet on?
    - raincole 3y ago
      I was kinda tempted, but I don't think there is such a good metric to measure it. (Since the job market is heavily influenced by interest rate etc... and if I were confident at predicting macro economics I would just bet on stock)
      - jefftk 3y ago
        What if we picked 10 prompts that it seems like an AI should be able to depict well, but it can't yet? And then iff the best AI tool in February can do the majority of them you win?
        raincole 3y ago
        I'll bet against it tho. I don't actually believe pure text-prompt-to-image will improve much (not in a few months at least). I just believe there will be more non-text tools to guide AI, like LoRA and control net, and they will be more accessible. Control net kinda did what you said but in a different timeframe: it was quite difficult to tell AI to generate a person "sitting with their legs cross". Today, it's relatively easy to do this with control net, but still hard with text prompt only. Edit: and the sibling comment made me question myself why I would ever take a random bet on the internet.
        jefftk 3y ago
        We started this thread with: sambleckley> producing a specific image with generative AI is sometimes almost impossible justrealist> Who could possibly think this will be the case six months from now? me> I'd bet on this being true six months from now I thought you disagreed with me on this, but it sounds like maybe not?
        raincole 3y ago
        I was just stating the improvement on SD we've seen since DELL-E 2 and Midjourney came out wasn't just about "quality of image", but also about "have something specific and moderately complex in mind". Thus I mentioned textual inversion vs LoRA/ControlNet.

gedy 3y ago

> While the quality of images has improved a lot, if you have something specific and moderately complex in mind it's still very hard to get there just from a textual prompt. I think this gets to an important point. If whoever is paying the bills has something very specific in mind, they won't be happy with AI (or frankly many artists) at this point. But as a creative interpretation of something more general, it's actually really good, and I think with many low-importance works like the author describes as "furniture" - we really don't need to be that exact.

GaggiX 3y ago

Dalle 2 was out April 2022. I don't think it will be true six months from now. OpenAI has been red teaming a new version for months that goes way beyond anything we've seen today. You can see some leaks from this video: https://youtu.be/koR1_JBe2j0 https://youtu.be/koR1_JBe2j0

jefftk 3y ago
> Dalle 2 was out April 2022 Sorry, thanks! I accidentally gave the date for the first DALL-E. > OpenAI has been red teaming a new version for months that goes way beyond anything we've seen today. Any sign that it's better at understanding complex textual prompts, and not just at making high-quality images?
- GaggiX 3y ago
  >Any sign that it's better at understanding complex textual prompts, and not just at making high-quality images? The video I have linked.
gwern 3y ago
I don't know if that's 'way beyond'. Those samples look similar to Imagen and Parti in terms of quality and following complicated text prompts. (Look at the PartiPrompts for the absurd prompts Parti can execute accurately.) I'm glad to see OA does have a successor for DALL-E 2, though: the service seems to have somehow gotten worse since release.
- GaggiX 3y ago
  I shouldn't have said "goes way beyond anything we've seen today", but probably it goes way beyond anything we will be able to try in this 6 months period. Parti 20B is really good at language understanding, more meh Imagen. >the service seems to have somehow gotten worse since release. Yeah, I think it's pretty clear that they probably lowered the number of steps or used a similar strategy to reduce the computation.

iudqnolq 3y ago

But in general the better something is the harder it is to improve. From an outsiders' perspective it looks like we've had a series of achievements where AI worked impressively well considering it isn't a human. But afterwards we got an incremental grind that never got close to an AI that's just good, period. I'm very impressed people got self-driving car demos working. I couldn't do that. But year after year self-driving cars remained a demo, albeit an incrementally improving one.

jefftk 3y ago
> year after year self-driving cars remained a demo, albeit an incrementally improving one You can get a ride in Phoenix today. Order a Waymo from your phone, it shows up with no driver. https://waymo.com/waymo-one-phoenix/ https://waymo.com/waymo-one-phoenix/ The number of cities is still small, but it's not a demo anymore.
- iudqnolq 3y ago
  It's limited to a small number of cities with optimal conditions and every ride has a human supervising. They're still spending R&D, rather than making a profit. That's a demo.
  - jefftk 3y ago
    > every ride has a human supervising What do you mean? Waymo has been running fully driverless operations with the general public since 2020: https://waymo.com/blog/2020/10/waymo-is-opening-its-fully-driverless.html https://waymo.com/blog/2020/10/waymo-is-opening-its-fully-dr...