What is a "LoRa" in AI?

At all those web sites for using AI to create images or videos, one term appears often: "LoRa" = Low-Rank Adaptation. It turns out that these are extensions of the original model which add to the original model without that model having to be re-trained. Hereby a lot of computing time is saved.

Here is the explanation of LoRa from one of its inventors at Microsoft, Edward Hu:

Testing DeeVid AI

 DeeVid is a site which offers image-2-video, text-to-video, video-2-video, and lip-sync. Also AI images can be created.

Here is a SFW video created from a starting and an end image:

NSFW content is detected and is not being processed.

But SFW content is fine, like this:

Testing SEAART.AI - Update

One of the results that comes up in a Google search for "img2vid AI nsfw" is SEAART.AI.

While it looks initially promising, testing it initially was quite a disappointment.

First, one needs to pay something to get videos. There are credits and stamina. I have not yet been able to understand the difference between the two, but I figured that one can buy the credits, and the stamina are some kind of reward. For example, one gets 130 stamina credited every day, but then they disappear.

One can also get a subscriptions to become an SVIP, which includes credits and stamina and provides additional benefits.

Since I needed some credit, I purchased for about €20 a set: 

Gave me 17600 credits. Creating a video from an image costs 400 credits.

Which means that this would last for 44 videos.

Which also means one video costs about 0.45 € , which is a reasonable price.

If I wanted to get SVIP, I had these choices:

My try: I wanted to upload a NSFW image, then create a video from it. This was the picture of a partially dressed doll, with bear breasts but panties under a skirt.

Uploading a 2nd (end) video requires to be an SVIP, so I only could upload a start image.

After the upload was done, the image was incorrectly rotated 90 deg. This has not happened with any other AI app site, but I had this problem also with PizApp. Some issue in how my phone encodes the with and height. But Windows apps usually handle this well.

OK, then I tried to generate, adding a prompt about removing panties. 

Gave no warning, but then the result was blurred - it was apparently tagged as adult / NSFW, and I was not supposed to see this. In my setting for SeaArt I enabled mature imagery, but that did not seem to have any consequence...

I gave it another try - it still loaded the source image 90 deg rotated horizontally. This is a stupid bug which will prevent me using Searart.ai .

There are no immediate and direct links to terms of service or to the content policy... and there is no warning that when uploading a pic must not be more than PG-13.

The result this time was not NSFW, so I could see it. Looks good, except that it is 90 deg rotated because also the upload was in the wrong orientation.

The quality of the motion is not bad.

Overall this looks like a very promising site, with many options. But it currently has a few problems:

• loaded images are incorrectly rotated by 90 deg.  Update: this appears ok for other pictures.
• Despite having declared to allow NSFW, I cannot see any NSFW content, not even my own generated videos (at least the creation of those appears to be allowed)
• There is no help link. There is a feedback link, but it refers to feedback for specific images, and it is in Traditional Chinese.   Update: there is a customer service chat button, which I need to try out. 

Maybe in the future this site will develop better, but as of today I cannot recommend it.

Showing the power of VIDU

On the site CIVITAI I did experiment a bit with VIDU Q1. This tool allows to provide a start and and an end image, then it interpolates between these images and creates a video.

I had taken the following series of 9 photographs:

I now used the first 8 images of this series as start images, then set the image following each of these images as the end image. Each of the end images was again used as the start image for the next video.

With VIDU Q1 I then created the 8 individual video snippets. 

 

Then I concatenated them to this video:

Testing Civitai

 In principle this CIVITAI is a great and very comprehensive site: one can choose between a wide variety of models, approaches, techniques. Must actually be already an expert user in order to make full use of the many benefits and options.

But there is one shortcoming: one can only upload images up to PG-13. The system automatically (with the help AI of course) checks the content of the pic and classifies it. And when the resulting video is above PG-13, then it is hidden, and the paid credits are lost.

The site does have tons of NSFW AI content, and they seem to be more open to this than other AI sites. But they apparently had a battle with their credit card payment processor, so currently no direct credit card payment is possible. But they are very creative and have found a good solution: allowing the purchase of gift cards which then can be redeemed.

To please their linked services, such as VIDU (who are VERY strict and adamant about only being up to PG-13), CIVITAI limits uploading your own images (e.g. as starter images) to PG-13. One would have to create an AI image from scratch using CIVITAI or one of the other linked AI services (e.g. comfyUI) to generate an AI image, and then during the upload, the metadata for those are checked. If the uploaded image does not have those metadata, then it is not allowed.

So far this is the most comprehensive AI site which has a wide variety of imagery and videos. Very impressive is their partnering with VIDU who are providing the VIDU Q1 concept. This create very nice videos (4 sec long) from a starting and an end image.

Here is an example:

I also created this, ensembled of a series of short VIDU Q1 video snippets:

I have to see how I can use this site to create also more elaborate NSFW videos. For PG-13 content, this site is unrivalled by any other in terms of quality and content.

Their support seems to be slow... I sent an inquiry and have not yet heard back from them.

Cost per video of length 3-5 sec: $0.40, which is quite reasonable.

Testing Unlucid.ai

The web app Unlucid.ai is a very simple AI site: it has only a few preset models, which can be applied to a single picture and create a 4 second clip. 

There are 4 NSFW models which work relatively well when the source picture is more or less aligned to the starting situation of the model.

Here is the pricing for purchasing "gems":

Creating a video costs 30 gems, which is between $1.56 - $2.25 per video - relatively expensive.

The models to be used are fixed - maybe they will be changed once in a while over time, but currently there is not much change or customization possible.

But the resulting videos are OK.

 

A video with male dolls

With Vidnoz and other (audio) tools I created this video:

Testing Vidnoz

In summer last year I discovered Vidnoz, and I have experimented a bit with it. It is quite powerful and allows talking videos to be generated from still images.

The main point of Vidnoz is to create advertisements or presentations, for example for online training. 

https://aiapp.vidnoz.com/avatar/index.html

There appears no limitation on NSFW, I was able to upload all kinds of images. 

Overall the site is a bit confusing - there are a several Vidnoz products, which require separate subscriptions:

• Vidnoz AI
• Vidnoz Gen
• Vidnoz AI Voice
• Vidnoz AI Video Translator
• Vidnoz AI Interactive Avatars
• Vidnoz Flex

The individual tools have even more choices: 

There are more than 1200 different voices, in different languages. When speaking English, these "foreign" voices even can speak in their native accent.

Is a great tool, very capable, but a bit confusing with all these various tools and options.

Testing DreamFace

Already for a while I had the DreamFace app on my mobile phone. It allows there to swap faces and do a few things with faces.I did get a "Perpetual Pro" subscription which allows:

• unlimited standard avatar video
• unlimited standard AI photo
• high-def and no watermark
• 3 mit per video
• 5 weekly credits

It allows upload of a single image, then animates it, For example a body shake, dancing or pets speaking.

Nothing NSFW unfortunately.

The desktop version then has a few more benefits: https://tools.dreamfaceapp.com/ 

One can upload a video, then one can upload an audio or type a text, and then the video is animated with the text or the audio, lips moving properly. This is quite a useful feature, and I have used it already a few times.

Unfortunately it does not store the created setups, so one has to memorize the voices used and the text entered. Only the resulting videos are stored.

Here are two examples of the cats speaking (these are German cats, so they speak German...) :

And here is a video of a doll sitting near a lake:

When the audio / text is shorter then the source video, the video is cut short.

When the audio / text is longer than the source video, the video begins to roll backwards.

My dolls

Here is a living spreadsheet that is updated with information about my doll heads. As of today, there are 91 different heads and 38 different bodies.

This spreadsheet also shows a few specific details on how the dolls are used in the context of OneSixthuals: e.g. age, backstory, etc.

AI - Artificial Intelligence for my doll videos

In December 2021 I began taking pictures of dolls. These are dolls in scale 1:6, similar to the Barbie doll scale. These dolls look very realistic, and I realised that one could create scenes in which dolls interact.

So I began taking picture series, with still shot of the dolls, then putting captions and speech bubbles like in a cartoon. This way I was able to create "Graphic Novels". 

Then in summer 2024 I discovered that AI technology can actually animate those static doll pictures. So I experimented a bit and was fascinated that I could give my dolls a voice. Face animation is done automatically, fitting to the text they are speaking. 

And so I began experimenting, and I will share the results here on this blog. I will introduce a few of the apps that I have used, but I must state that this will just be a snapshot of the current situation: the technology is very much in flux, things are rapidly changing, and there are new innovations everyday. So my "reviews" may get outdated soon, but at least for now they may provide a starting point for those who are interested in follow these developments or give them a try.

Happy New Year 2025!

Dear all,

you may have noticed the change: this blog of oneSixthuals is starting anew. All the previous posts that were under the URL for this blog here are now under a new name: oneSexthuals . This was done to refocus on the two aspects of dolls: naughty and nice.

This blog here will be devoted to the "nice" aspects, where as the "naughty" aspects can be seen here: https://onesexthuals.blogspot.com/ 

I hope that I will be able to bring a few nice and interesting content here, which allows wide audiences enjoying the beauty of those dolls in scale 1:6.

So, happy new year to all of you from the dolls of oneSixthuals!