Experimenting with generative AI
Hello, this is @joviansummer(witness: @jswit). I may be a stranger to this art community, but hopefully some of you might find this post interesting.
I've been experimenting with generative AI for some time, and it was pretty interesting for me. I didn't use online services. Instead, I directly used open models such as Stable Diffusion and its derivatives combined with UI software such as Forge or ComfyUI.
The image attached to this post was done by me, with multiple back and forth with AI model, using image-to-image generation.
These are a few things I've found regarding this experiment, but considering that I'm neither artist nor AI expert, it would be better to take them with a grain of salt. Anyway, here they are:
Maybe it's just me, but it was really difficult to get what I want exactly with text-to-image generation.
AI could generate pretty good images, but there were anomalies sometimes. Especially eyes, fingers, or arms/legs.
If I manually draw/paint a draft image and then use AI to revise it, sometimes AI could present brilliant results. If you have specific visual in your mind, you need to draw/paint something so that AI can use it as a reference. Text prompt(text-to-image) was not enough in many cases.
Image generation using text prompt only may help artists getting some new ideas/inspirations.
You need a powerful GPU for image generation. It's possible to do run image generation AI models with CPU, but it's very slow.
I'm testing an experimental service jsup.GEN to connect Steemit and generative AI. Currently it's limited to text-to-image generation. Maybe I should look into image-to-image generation also.
Personally I don't think AI will replace real artists, at least for now. Instead, AI may help artists speed up their works greatly without compromising visual quality.
Thank you for reading, and have a wonderful day!
@joviansummer's STEEM projects
@jswit witness:
I'm running a STEEM witness node. I'd really appreciate it if you vote for my witness account @jswit. (https://steemitwallet.com/~witnesses)
[ENG] Introducing @jswit witness project
@jsup curation:
[ENG] Introduction to @jsup curation project
[ENG] Using @jsup curation project as an auto-voting agent
[ENG] jsup 2.0 - make your upvote great again
Steemit-Search:
Steemit-Search: a simple website for post searching
jSTEEM:
jSTEEM project - STEEM blockchain explorer on Telegram messenger
Upvoted! Thank you for supporting witness @jswit.
До jsup.GEN я почти не пользовалась генерирующим картинки AI. Как вы и сказали, AI нужна большая мощность, чтобы по текстовому описанию воплотить замысел художника. Художнику легче самому нарисовать, что он хотел изобразить.
По моим наблюдениям генерирующий AI подходит людям, умеющим работать с большим количеством картинок: в интернете, в обработке. Точно также они могут долго смотреть и отбирать AI-картинки.
jsup.GEN в Steemit отличается от этого. В первую очередь он знакомит с технологией. Удачно, что при этом он не напрягает пользователя большим количеством картинок. И я наконец-то могу, не напрягаясь, переводить свои тексты в картинки для любителей картинок. =)
Кроме того, мне понравилось изучать что умеет рисовать jsup.GEN.
AI can bring you pretty good images sometimes. The problem is that it's an expensive technology now. Maybe it will become more affordable in the future. Thanks!