summaryrefslogtreecommitdiffstats
path: root/content/notebook/fediverse
diff options
context:
space:
mode:
Diffstat (limited to 'content/notebook/fediverse')
-rw-r--r--content/notebook/fediverse/ai-art.md29
1 files changed, 24 insertions, 5 deletions
diff --git a/content/notebook/fediverse/ai-art.md b/content/notebook/fediverse/ai-art.md
index 49f491b..6ea5a8b 100644
--- a/content/notebook/fediverse/ai-art.md
+++ b/content/notebook/fediverse/ai-art.md
@@ -1,18 +1,22 @@
1--- 1---
2schema_type: DigitalDocument 2schema_type: DigitalDocument
3title: On AI Art -- Artists Weren't Happy When Photography Was Invented, Too 3title: On AI Art
4date: 2022-09-30 4date: 2022-09-30
5last_update: 2022-09-30 5last_update: 2022-11-13
6 6
7references: 7references:
8 - label: Reply to my Post 8 - label: "2022-09-30 -- Reply to my original post"
9 url: https://merveilles.town/@jbauer/109088036845654325 9 url: https://merveilles.town/@jbauer/109088036845654325
10 - label: My Response 10 - label: "2022-09-30 -- My Response"
11 url: https://mk.vulpes.one/notes/95s8h9ajtp 11 url: https://mk.vulpes.one/notes/95s8h9ajtp
12 - label: "2022-11-13"
13 url: https://mk.vulpes.one/notes/97i33e337z
12--- 14---
13 15
16## 2022-09-30 -- Artists Weren't Happy When Photography Was Invented, Too
17
14::: alert 18::: alert
15I posted the title of this page on fedi. Someone replied and I elaborated on my views with the following post. 19I posted the title on fedi. Someone replied and I elaborated on my views with the following post.
16::: 20:::
17 21
18The primary reason AI art is widely getting banned is because a lot of people are wowed by the novelty of this technology and post their shitty results. But there are also people who spend a lot of time on tuning their prompts, running the results through img2img several times and touching things up manually in Photoshop. They're spending real effort on getting good results. 22The primary reason AI art is widely getting banned is because a lot of people are wowed by the novelty of this technology and post their shitty results. But there are also people who spend a lot of time on tuning their prompts, running the results through img2img several times and touching things up manually in Photoshop. They're spending real effort on getting good results.
@@ -27,3 +31,18 @@ Artists often use references and look up to other artists and adapt qualities fr
27Stable Diffusion was trained by looking at art as well. It doesn't have a database of every single picture. Instead it recognizes various concepts and puts them in some N-dimensional space -- it's memories. So one point in this space captures one concept it has seen in many different pictures. The prompt simply determines which of these concepts the AI will try to use (up to 75 with SD). 31Stable Diffusion was trained by looking at art as well. It doesn't have a database of every single picture. Instead it recognizes various concepts and puts them in some N-dimensional space -- it's memories. So one point in this space captures one concept it has seen in many different pictures. The prompt simply determines which of these concepts the AI will try to use (up to 75 with SD).
28 32
29Do you see the parallels? At least from my limited understanding, this doesn't seem much different from humans looking at pictures and recognizing shading, lighting, painting,... techniques and selectively using them in their own art. 33Do you see the parallels? At least from my limited understanding, this doesn't seem much different from humans looking at pictures and recognizing shading, lighting, painting,... techniques and selectively using them in their own art.
34
35## 2022-11-13 -- The Concept of "Lost Commissions"
36
37An argument I hear often is that using an AI means I can get a piece in an artist's style when it should've been a commission. This doesn't really work for my situation because I don't copy an individual artist's style in my generated art -- it's a remix of things from both the original data and the data I added. I don't think commissioning ~10 artists at once is even possible, and if it was you'd better have lots of money.
38
39> Then commission an artist whose style is close enough
40
41Sure, I guess I could do that.\
42But this is where the second problem comes in: The way this usually goes with good artists (who are usually also popular) is that commissions are closed, the queues are full for the next 3 years, they never respond, and/or you have to keep track of them on Twitter for those 5 nanoseconds they're open before 10 billion followers snatch all slots.\
43I went through it all and I hated it. I won't even try commissioning an artist anymore if I notice any of these things because this is too stressful.
44
45And now there is AI that allows me to create good art of my random ideas. It's art that otherwise never would've existed because without the instantaneous feedback of AI I wouldn't even know if the idea worked or if I liked the implementation.\
46On the flip side, with ideas that proved to be good such as the fox-corvid-gryphon, I am now considering creating more art of it with my own abilities and maybe commission someone if there ever is a chance. So if anything, AI has *increased* my willingness to buy commissions.
47
48This whole situation is strikingly similar to the music/video/gaming industry, piracy and the concept of "lost sales". Food for thought.