4 ms·

Dall-E also doesn't allow you to say "in the style of X" and similar. Which I guess appease some concerns from artists, but also makes it harder to guide it whe

Dall-E also doesn't allow you to say "in the style of X" and similar. Which I guess appease some concerns from artists, but also makes it harder to guide it where you want it. So it's not just about wanting to make explicit images, it's about not having a different entity control what you can and cannot do.

yieldcrv 3y ago
I'm able to observe that as well, I'm also able to observe that caring about that isn't necessary in most use cases and the convenience still wins. I do branding and presentations all the time with graphics from Midjourney and rapidly iterate in that discord chat in various branches compiled in parallel all day. Make something as convenient, cheaper, private/offline, and faster and I'm there. Otherwise the ideology is irrelevant to me, and the unsaid secret is that that's true for most everyone else too.
- matsemann 3y ago
  It was more a rebuttal to you claiming that it only has explicit images going for it. Almost like you were labeling those using SD to use it only for your claimed reasons.
- CuriouslyC 3y ago
  Midjourney is good for inspiration, but as more people use it things that are just whole cloth generations are going to start to look same-y and cheap. Some iteration/post processing will be necessary not to look low rent.
genewitch 3y ago
with SD you can use LoRA, and "artist style" is, to my understanding, the easiest thing to make a model of. With LoRA (Low Rank Adaptation) you describe everything in an image that is not what you want the model to model. With a style, You can just use BLIP/CLIP or deepbooru to describe all of your images. At worst, you might have to remove other styles/artists in the tags. the model will only learn about the style. supposedly. I don't follow art enough to have a favorite style; i don't have enough source material on hand to do a style model, nor do i know enough to go out and grab a dataset. As an aside, textual inversion also made some great inroads into this sort of thing. So smallest size to largest: textual inversion(megs), Locon/lycoris/lora(tens of megs), full model (gigs!). The accuracy range over all compositions follows the same respective order, as well.
- efreak 3y ago
  Loras are 1 to hundreds of megs for sd15 (I've seen 1mb Lora and also 700mb Lora merges) and I think around 1gb+ for sdxl. Hypernetworks exist as well for sd15 at least, they're up to I think around 80mb?

yieldcrv 3y ago

I'm able to observe that as well, I'm also able to observe that caring about that isn't necessary in most use cases and the convenience still wins. I do branding and presentations all the time with graphics from Midjourney and rapidly iterate in that discord chat in various branches compiled in parallel all day. Make something as convenient, cheaper, private/offline, and faster and I'm there. Otherwise the ideology is irrelevant to me, and the unsaid secret is that that's true for most everyone else too.

matsemann 3y ago
It was more a rebuttal to you claiming that it only has explicit images going for it. Almost like you were labeling those using SD to use it only for your claimed reasons.
CuriouslyC 3y ago
Midjourney is good for inspiration, but as more people use it things that are just whole cloth generations are going to start to look same-y and cheap. Some iteration/post processing will be necessary not to look low rent.

genewitch 3y ago

with SD you can use LoRA, and "artist style" is, to my understanding, the easiest thing to make a model of. With LoRA (Low Rank Adaptation) you describe everything in an image that is not what you want the model to model. With a style, You can just use BLIP/CLIP or deepbooru to describe all of your images. At worst, you might have to remove other styles/artists in the tags. the model will only learn about the style. supposedly. I don't follow art enough to have a favorite style; i don't have enough source material on hand to do a style model, nor do i know enough to go out and grab a dataset. As an aside, textual inversion also made some great inroads into this sort of thing. So smallest size to largest: textual inversion(megs), Locon/lycoris/lora(tens of megs), full model (gigs!). The accuracy range over all compositions follows the same respective order, as well.

efreak 3y ago
Loras are 1 to hundreds of megs for sd15 (I've seen 1mb Lora and also 700mb Lora merges) and I think around 1gb+ for sdxl. Hypernetworks exist as well for sd15 at least, they're up to I think around 80mb?