1
« Last post by redhome on Today at 01:05:38 pm »
I've started seeing some VIPER AI images on Pixiv over the last year or so. (And a few posted here as well.)
It seems to be getting closer, especially for the faces though Cala's unusual hair is rarely done right and the Rise girl's clothing is often wrong or miscolored as I think it tends to confuse the Anne Mitter's uniform with other, similar styles.
I had not heard of anyone using AI to plug in uncensored parts to otherwise original art though. What I'd seen of this before was presumably handmade or used assets from other series copied onto the original. It would be great to see some of this, too to know if it looks any better than the handmade or copy-paste uncensoring.
That's AI for you. AI model training seems easy, but is very hard to master when it comes to getting the right output. Usually, the wrong input (images & tagging) is all that matters. The things you describe could be "hallucination" by the AI model - or the image creator chose to let the AI have more "creativity", if you will. Although, that term is not true since it's all about statistics and randomness. In the post above I tried to explain the "burn in" effect. It's good to have this, to get the character (e.g. Cala) right, but it's bad, if the burn in is too strong. If it's too strong, you will always get the same outputs no matter what you "prompt for". (e.g. "Cala in a bikini, on a mountain, sunrise in the background" - but what you get is Cala in a cave with a monster...). So you need to get something in the middle, which allows for things not found in the source material (to create completely new stuff), but this also means the output may differ (hair, clothing, colors). The training parameters itself may also have a very big impact on the outcome. But this is way too complex to go into detail here. And let's not talk about hands. The Achilles' heel of AI image models.  My guess is, that image models will get better over the next years, but as usual it takes time. Maybe I am not so critical about the output, if the style is the right one - and the characters are (very) recognizable. With or without clothes.  I try to get into the uncensoring part with a few screenshots, if I have time!
2
« Last post by redhome on Today at 12:53:52 pm »
I knew it was only a matter of time until AI generated Viper images would be done. The days of crude looking and poorly pieced together results from combining existing images and gifs into new gifs and images are practically over lol.
They look great btw! The V16 Rise girls especially; that Karin though  , though I'd say the Cala images from the AI could use a little work. I'm sure I wouldn't be the only person one here who can identify many of the images that have been used/referenced or which characters have been combined. Makes me feel like an expert 
There are at least two possibilities to train an AI image model to let it "know" what to learn and what to ignore (in both cases the source material must be high resolution, which is a real problem when trying to do something like that with VIPER): - Characters/Clothing/Concepts: I need to tag everything in the image - except - the things that should "burn in". Lets say I want to train a character named Karin, we know she usually has twin braids, blonde and long hair, glasses etc. - those things are not described by me, so the AI associates those specific traits always with the character/concept. Things not described by me, burn in.
- Style: The contrary to 1. is true here. I tag everything seen in an image. I don't want anything (except the style) to "burn in", but I like the AI to learn the general art-style, not the characters or anything else in the picture. This way the AI model is later able to generate images that look like VIPER artwork, but can be a lot of different things. I could create Lara Croft using the VIPER V-16 style, for example.
Coming back to your input about Cala: I didn't tag the images properly (see 2.), plus I didn't uncensor the source material. This way I can prompt for "a young woman on mars, moons in the sky" and I will get Cala being "taken" by some monster (because it burned in). It's an early model. But I need some time to refine it. Another thing is making things complicated: I cannot use similar images twice. That means, all images presented to an AI should/must be different, so that the AI model learns the concept (character et al) by seeing it/her/him/... in different positions etc. The Cala base images are different, but the things happening to her are usually the same.  That's why you recognize those poses/scenes easily.
3
« Last post by JG on Today at 09:17:08 am »
It has been brought to my attention that registrants are not receiving the registration validation email. Further testing has concluded that private messages are also not being sent.
I don't really know what the problem is. The emails are just sitting there without moving forward. I've rebooted the server and I've tested sending emails directly from the Outlook account and don't yet know what the problem is, as there's no record in the log of SMTP failures - they just aren't getting sent or moving out of queue. This may take a while to figure out the root cause and fix.
In the meantime, if you have recently registered and did not get your validation email, please send an email to sognadigitalmuseum@gmail.com from the email account you are trying to register and your registration will be completed manually. Apologies for the disruption of service and extra manual step.
4
« Last post by JG on Today at 08:19:34 am »
I've started seeing some VIPER AI images on Pixiv over the last year or so. (And a few posted here as well.)
It seems to be getting closer, especially for the faces though Cala's unusual hair is rarely done right and the Rise girl's clothing is often wrong or miscolored as I think it tends to confuse the Anne Mitter's uniform with other, similar styles.
I had not heard of anyone using AI to plug in uncensored parts to otherwise original art though. What I'd seen of this before was presumably handmade or used assets from other series copied onto the original. It would be great to see some of this, too to know if it looks any better than the handmade or copy-paste uncensoring.
5
I knew it was only a matter of time until AI generated Viper images would be done. The days of crude looking and poorly pieced together results from combining existing images and gifs into new gifs and images are practically over lol. They look great btw! The V16 Rise girls especially; that Karin though , though I'd say the Cala images from the AI could use a little work. I'm sure I wouldn't be the only person one here who can identify many of the images that have been used/referenced or which characters have been combined. Makes me feel like an expert
6
« Last post by redhome on Yesterday at 04:14:44 pm »
And here's a short preview on the RSR based LoRA. The inputs were mostly censored, so I need to uncensor them before training the LoRA again. Plus, the input were not "tagged" correctly (you need to tag the contents of an image, to let the "AI" know what it needs to learn and what to ignore).
7
« Last post by redhome on Yesterday at 04:09:34 pm »
Hi there! Still new here and I thought maybe some of you (even if the topic itself might be controversial) could be interested in something I am working on right now - or already worked on, to be honest. It also includes shameless self promotion. One of my hobbies (and partly my profession, too) is AI research, which also includes creating the basis of what you would call "AI models" to generate either text or images. I am not overly sciency about it and some of you might already have worked in that field in one way or another. Since I am also a really big fan of the VIPER series (for obvious reasons), I thought I could give it a spin and created a so called "LoRA" ( https://github.com/microsoft/LoRA) of the VIPER Rise series. It's a tedious process, but in the end you are able to create really neat images, which are not identical, but very close to the original (artworks). In my case it's a "style" (others are "character", "clothing" etc. or really any concept you can imagine). If you like, you may visit my VIPER Rise style LoRA here at Civitai and give it a spin yourself ( https://civitai.com/models/420185). If you own a beefy graphics card, you can even do that 100% locally. Currently, I am working on a new "style LoRA" based on the RSR artworks/games. Funny thing I noticed: I can easily "uncensor" original artworks/game images using something called "inpainting" and utilizing my LoRA (mentioned above). This is either done in the various web UIs (a1111, forge, ...) or by using "Krita" ( https://krita.org/en/) with the so called diffusion plugin ( https://github.com/Acly/krita-ai-diffusion). If you're interested, you can just reach out to me. Here are some sample images I created.
8
« Last post by JG on Yesterday at 10:48:00 am »
Well, I mean it doesn't help within your own market. But US dollars and British Pounds (and to a bit lesser extent, Euro as well) should have greater buying power in Japan. Five years ago if I wanted to buy something priced 5000 yen, I'd expect to pay about $45 USD. Now I can buy it for about $32. So all this art that used to start about 20,000 yen used to be about $180 and should now be about $130. Instead its approaching twice that for even the cheapest art (I'd say 35,000 is the new floor)
9
« Last post by uphillbattle on April 29, 2024, 11:22:53 pm »
Games, art books, and most figurines have pretty much stayed in the same price range for the last 10 years or so. Its really only the original art and telephone cards that seem to be accelerating upwards.
And with the Japanese yen dropping like a ten ton weight over the last couple of years compared to dollar (see attachment), we westerners really should be seeing better prices now than we did, but we aren't. People just gone crazy I guess.
I wish that were so, but my cost of living has gone up dramatically the past few years, practically offsetting any better pricing we aren’t seeing, so its actually been worse for me personally  
10
« Last post by JG on April 29, 2024, 07:41:05 pm »
Games, art books, and most figurines have pretty much stayed in the same price range for the last 10 years or so. Its really only the original art and telephone cards that seem to be accelerating upwards.
And with the Japanese yen dropping like a ten ton weight over the last couple of years compared to dollar (see attachment), we westerners really should be seeing better prices now than we did, but we aren't. People just gone crazy I guess.
|
Recent Posts