- Marianan ID
- 92
- Joined
- Aug 27, 2023
- Threads
- 21
- Messages
- 426
- Reaction score
- 565
- Awards
- 43
- Location
- Araara Island
- Website
- birb-site.neocities.org
LV
2
Offline
It's been a bit over a year since public heard about AI and possible use cases for it. At first there was a hype for image generation: dalle and other similar web based image generation services. Later attention shifted towards chatgpt. One of the most disgusting displays of normies poking around with something beyond their understanding. After it came turn for voice generation. "President gaming" edits powered by elevenlabs took tik tok and other short media formats by storm.
"Ey bro look Trump and Obama are saying gamer words while rushing B site in CS:GO"
ehh...
Fortunately OpenAI (company behind GPT models) is doing its best to make it as unappealing to users. Untroducing payed tier, applying shit ton of model censoring and filters. Even their greatest model, gpt4 is way more stupid now in comparison to its release half a year ago.
But back to the topic. I want to focus on the image generation, as I think I have the most experience with it. There are couple of words that come up whenever talking about generated images:
1. Shit ton of details to the image
2. Overly saturated
3. Cross breading / riding the hype
Those are the main issues that I have with this field, let me explain
1.As we consume more and more media, we start to pay less attention to it. I think this is one quite good example
Cozy classroom and all... Until you actually pay any attention to it.
Notice chairs merged into the desks. The light that doesn't match with the windows on the right. Whatever is happening with the window on the left. I will go easy on the board / clock / items on the shelves. As the "powerful AI" always had issues with it.
>Bruh stop roasting your shitty renders
Sucks. This is a 2nd preview pic from a model with over 333k downloads https://civitai.com/models/4823/deliberate
Is it really that hard to google "Cozy Japanese classroom" ?
2. Saturation
This is actually something that can be avoided... If creator wants to do anything about it. But most decide to not and just autopilot it
This issue is most visible with "realcartoon" and other models that try to merge realistic images of people with absurd colors and "fantasy elements". Gonna use Front promo image for "realcartoon3d V8" model. A god dam model with 46k downloads as of writing.
I will keep my list of complaints brief, as this point is not supposed to be about image features, just general ramblings about unnecessary over-saturation.
1. black something hanging from the right ear. Also something similar not attached to ear hidden in her hair to the left
2. Double "top head"
Does it really has to be so saturated ? Thank god it's not an image of something / someone outside. Because then colors would be way off the charts.
3. Cross breading.
This is purely technical complain, so I will keep it brief.
Models are trained on images. Many thousands before desirable effect is achieved. And there's only said amount of good quality images on the internet. At some point AI generated images sneak into the training data. Making the "AI art" look even more generic.
1. Overly saturated colors
2. Absurd amount of details
I hope that you start to see my point. It tries to catch attention with colors and wow users with details. While failing to propely do either of them
Will there be another wave of AI generated content that normies will jump on ?
Yes. There's still video left. There's absurd amount of work being done to make it work. Fortunately it's far from "type what you want to see and wait 3 minutes". I had a go at it myself in May... Results were bad to say the least. I will be getting a new GPU next week. I might come back to this subject and post results if something acceptable comes from it. I also got quite out of the loop, so it might be a bit closer than I think
thanks for coming to my ted talk
"Ey bro look Trump and Obama are saying gamer words while rushing B site in CS:GO"
ehh...
Fortunately OpenAI (company behind GPT models) is doing its best to make it as unappealing to users. Untroducing payed tier, applying shit ton of model censoring and filters. Even their greatest model, gpt4 is way more stupid now in comparison to its release half a year ago.
But back to the topic. I want to focus on the image generation, as I think I have the most experience with it. There are couple of words that come up whenever talking about generated images:
1. Shit ton of details to the image
2. Overly saturated
3. Cross breading / riding the hype
Those are the main issues that I have with this field, let me explain
1.As we consume more and more media, we start to pay less attention to it. I think this is one quite good example
Cozy classroom and all... Until you actually pay any attention to it.
Notice chairs merged into the desks. The light that doesn't match with the windows on the right. Whatever is happening with the window on the left. I will go easy on the board / clock / items on the shelves. As the "powerful AI" always had issues with it.
>Bruh stop roasting your shitty renders
Sucks. This is a 2nd preview pic from a model with over 333k downloads https://civitai.com/models/4823/deliberate
Is it really that hard to google "Cozy Japanese classroom" ?
2. Saturation
This is actually something that can be avoided... If creator wants to do anything about it. But most decide to not and just autopilot it
This issue is most visible with "realcartoon" and other models that try to merge realistic images of people with absurd colors and "fantasy elements". Gonna use Front promo image for "realcartoon3d V8" model. A god dam model with 46k downloads as of writing.
I will keep my list of complaints brief, as this point is not supposed to be about image features, just general ramblings about unnecessary over-saturation.
1. black something hanging from the right ear. Also something similar not attached to ear hidden in her hair to the left
2. Double "top head"
Does it really has to be so saturated ? Thank god it's not an image of something / someone outside. Because then colors would be way off the charts.
3. Cross breading.
This is purely technical complain, so I will keep it brief.
Models are trained on images. Many thousands before desirable effect is achieved. And there's only said amount of good quality images on the internet. At some point AI generated images sneak into the training data. Making the "AI art" look even more generic.
1. Overly saturated colors
2. Absurd amount of details
I hope that you start to see my point. It tries to catch attention with colors and wow users with details. While failing to propely do either of them
Will there be another wave of AI generated content that normies will jump on ?
Yes. There's still video left. There's absurd amount of work being done to make it work. Fortunately it's far from "type what you want to see and wait 3 minutes". I had a go at it myself in May... Results were bad to say the least. I will be getting a new GPU next week. I might come back to this subject and post results if something acceptable comes from it. I also got quite out of the loop, so it might be a bit closer than I think
thanks for coming to my ted talk
