3 ms·

Very unlikely to be better than SDXL combined with its rich ecosystem of loras, controlnets, and other custom content in WebUIs like Automatic1111 or comfyUI an

Very unlikely to be better than SDXL combined with its rich ecosystem of loras, controlnets, and other custom content in WebUIs like Automatic1111 or comfyUI and their extensions. Would love to be proven wrong!

KerryJones 3y ago
Is your hypothesis that DALL-E 3 won't be using additional methods? Or what do you consider "better" -- I agree Automatic1111 and comfyUI make SDXL great for home use, but it feels a bit like an argument of PC vs Mac, where OpenAI is providing the Mac experience (polished, instantly easy to use) while PC offers more customization and nitty gritty.
- dragonwriter 3y ago
  > but it feels a bit like an argument of PC vs Mac, where OpenAI is providing the Mac experience (polished, instantly easy to use) while PC offers more customization and nitty gritty. 1984, the year the Mac was introduced (and when the more customizable, nitty-gritty Apple II series was their main seller) was also Apple’s all time peak in share of the personal computer market in terms of units sold. So…yeah, I think the comparison is apt, but maybe not in the way you think it is.
redox99 3y ago
MJ is already better than SDXL if you don't need any of the tools that SDXL has and MJ doesn't. So it isn't that hard to imagine Dalle 3 being better than SDXL.
- dragonwriter 3y ago
  > > Very unlikely to be better than SDXL combined with its rich ecosystem […] > MJ is already better than SDXL if you don’t need any of the [elements of the rich ecosystem beyond the base models] Yeah, I think you kind of missed the point there. (Also, not convinced you are right even there, MJ seems to be much worse at prompt-following than base SDXL model, and on other qualities in the range where subjective opinions are going to vary considerably on which is better, judging from the head-to-head comparisons with prompts I’ve seen, largely from people claiming that MJ is better so presumably not trying to subtly favor SDXL in the construction of the comparisons. Because of the ecosystem, its been a long time since I found MJ more useful than even SD 1.5-based toolsets.)
  - ehsankia 3y ago
    From my experience, while MJ is not as good at following the exact prompt, it's been much better at producing overall visually appealing images. I don't have as much experience with SDXL, but I've used plenty of Dall-E and other models; most of them generate results that follow the prompt well, but look more like a collage than a piece of art. Which may very well be what you're going for, but for doing more abstract and creative brain storming, I much prefer the Midjourney, which definitely has a specific "look".
- mrKei4 3y ago
  [dead]
mlsu 3y ago
To add a little bit to this: ControlNet. If you are actually a visual artist, I think the leader of the pack right now is controlnet, because you can exactly determine the visual structure of your image. While MJ or Dall-e may be better at "imagining" concepts, or have a more aesthetic sensibility (with Loras and custom-trained models, I'm not even sure about that) with controlnet you can very precisely specify how your image should be structured. This is closer to how traditional artists work. They don't go for the details (color, texture, shading) first. They do a sketch: what are the big forms in this image? How do they fit together? Then they begin filling in intermediate details. What is the color palette? Where are light sources? Which areas have contrast? Which do not? Only after they have done all of that preliminary work will they actually implement the frills on the dress, or the twirls of the mustache. If you are just a person who wants to make some pretty pictures, Midjourney (and, Dalle3, now) is probably your best bet. If you are an artist who wants to use an actual tool, you are using StableDiffusion. I think it's unlikely that the centralized "plug and play" Midjourney or OpenAI will ever be able to or interested in replicating the complex interface of stablediffusion. But there is a tremendous opportunity for a startup that can improve the UX of the complex workflows that are being developed by "ai artists." That's also why I am convinced that MidJourney / Dalle will not replace artists. You simply cannot, with a single prompt, replace the work of a true visual artist.
- krapp 3y ago
  >That's also why I am convinced that MidJourney / Dalle will not replace artists. You simply cannot, with a single prompt, replace the work of a true visual artist. They already are, because employers don't care about "true visual art," they care about cost and productivity, and getting an intern or someone outsourced to write prompts is both cheaper and faster than paying an actual artist, and capitalism dictates the path of least resistance is the path all competitive business must take. Companies are already replacing their creative staff with AI, or are planning to. AI generated art is already everywhere in advertising. And yes, they contain obvious errors that wouldn't exist with a real artist. And no, companies do not give a damn.
  - tommek4077 3y ago

KerryJones 3y ago

Is your hypothesis that DALL-E 3 won't be using additional methods? Or what do you consider "better" -- I agree Automatic1111 and comfyUI make SDXL great for home use, but it feels a bit like an argument of PC vs Mac, where OpenAI is providing the Mac experience (polished, instantly easy to use) while PC offers more customization and nitty gritty.

dragonwriter 3y ago
> but it feels a bit like an argument of PC vs Mac, where OpenAI is providing the Mac experience (polished, instantly easy to use) while PC offers more customization and nitty gritty. 1984, the year the Mac was introduced (and when the more customizable, nitty-gritty Apple II series was their main seller) was also Apple’s all time peak in share of the personal computer market in terms of units sold. So…yeah, I think the comparison is apt, but maybe not in the way you think it is.

redox99 3y ago

MJ is already better than SDXL if you don't need any of the tools that SDXL has and MJ doesn't. So it isn't that hard to imagine Dalle 3 being better than SDXL.

dragonwriter 3y ago
> > Very unlikely to be better than SDXL combined with its rich ecosystem […] > MJ is already better than SDXL if you don’t need any of the [elements of the rich ecosystem beyond the base models] Yeah, I think you kind of missed the point there. (Also, not convinced you are right even there, MJ seems to be much worse at prompt-following than base SDXL model, and on other qualities in the range where subjective opinions are going to vary considerably on which is better, judging from the head-to-head comparisons with prompts I’ve seen, largely from people claiming that MJ is better so presumably not trying to subtly favor SDXL in the construction of the comparisons. Because of the ecosystem, its been a long time since I found MJ more useful than even SD 1.5-based toolsets.)
- ehsankia 3y ago
  From my experience, while MJ is not as good at following the exact prompt, it's been much better at producing overall visually appealing images. I don't have as much experience with SDXL, but I've used plenty of Dall-E and other models; most of them generate results that follow the prompt well, but look more like a collage than a piece of art. Which may very well be what you're going for, but for doing more abstract and creative brain storming, I much prefer the Midjourney, which definitely has a specific "look".
mrKei4 3y ago
[dead]

mlsu 3y ago

To add a little bit to this: ControlNet. If you are actually a visual artist, I think the leader of the pack right now is controlnet, because you can exactly determine the visual structure of your image. While MJ or Dall-e may be better at "imagining" concepts, or have a more aesthetic sensibility (with Loras and custom-trained models, I'm not even sure about that) with controlnet you can very precisely specify how your image should be structured. This is closer to how traditional artists work. They don't go for the details (color, texture, shading) first. They do a sketch: what are the big forms in this image? How do they fit together? Then they begin filling in intermediate details. What is the color palette? Where are light sources? Which areas have contrast? Which do not? Only after they have done all of that preliminary work will they actually implement the frills on the dress, or the twirls of the mustache. If you are just a person who wants to make some pretty pictures, Midjourney (and, Dalle3, now) is probably your best bet. If you are an artist who wants to use an actual tool, you are using StableDiffusion. I think it's unlikely that the centralized "plug and play" Midjourney or OpenAI will ever be able to or interested in replicating the complex interface of stablediffusion. But there is a tremendous opportunity for a startup that can improve the UX of the complex workflows that are being developed by "ai artists." That's also why I am convinced that MidJourney / Dalle will not replace artists. You simply cannot, with a single prompt, replace the work of a true visual artist.

krapp 3y ago
>That's also why I am convinced that MidJourney / Dalle will not replace artists. You simply cannot, with a single prompt, replace the work of a true visual artist. They already are, because employers don't care about "true visual art," they care about cost and productivity, and getting an intern or someone outsourced to write prompts is both cheaper and faster than paying an actual artist, and capitalism dictates the path of least resistance is the path all competitive business must take. Companies are already replacing their creative staff with AI, or are planning to. AI generated art is already everywhere in advertising. And yes, they contain obvious errors that wouldn't exist with a real artist. And no, companies do not give a damn.
- tommek4077 3y ago