Experiments with AI image to video generation

[broho1234]

Experiments with AI image to video generation


After seeing a few Youtube videos on the new open source and uncensored video generator Frame Pack, for example

You must be registered to see the links

I was curious to see what kind of results one can get when asking for "adult" video.

Here are a few results I got using one image and a text prompt.
All the input images below are screenshots from the game "My Pleasure":

1. Prompt: "The girl licks the tip of the penis."
Input image:



Video output:

View attachment 01_Result.mp4



















Total failure: the AI doesn't understand how to make the girl lick the tip of the penis, instead it seems to think it needs to add cum from the penis (which is not entirely wrong).

2. Prompt: "The upper body moves back and forth up and down while the penis stays inserted in the woman's asshole."
Input image:



Video output:

View attachment 02_Result.mp4



















The result is not too bad, the penis doesn't move much but the woman's hands do.

3. Prompt: "The girls say hi and wave to the camera."
Input image:



Video output:

View attachment 03_Result.mp4



















The result is passable, the girls do smile and wave to the camera, but their hands are too close to the camera.


4. Prompt: "The woman opens her legs slowly."
Input image:



Video output:

View attachment 04_Result.mp4


















The result is not bad, the AI makes the skirt falls between the woman's legs which is a possibility, the other one being the skirt is stretched by the thighs so the crotch remains visible.

 

[broho1234]

Let me know if you have done similar experiments and what results you got.

Which AI generator is best for adult content ?

 

[Kretua]

Hi!
Thank you for the invitation.



The video above was made with ComfyUI using a WAN 2.1 generator and a LORA.
Two of the most prevalent NSFW generators used by the majority.
One is WAN2.1 , the other is Hunyuan. The majority prefer to use WAN2.1 because of the greater choice of LORA.
I should note here that FramePack uses Hunyuan.
The FramePack version shown in the opening post of this thread is not really suitable for "adult" videos. FramePack Studio is better suited for this (respect modine2021), because of the LORA support.
Experience shows that valuable NSFW content can only be achieved with LORAs.

If you are using Pinokio framework, you will find FP Studio there, and WAN2.1 api. As well as a bunch of cool AI. api.
Note that the pinokio.computer site is currently down due to a domain dispute, but with a little searching and a couple of Git commands you can install it.
Sorry if I'm sharing obvious stuff.

Here is a small video generated by FP Studio:
View attachment FramePack_00001.mp4

 

[Kretua]

sorry for the two posts, but the forum engine won.

There are two serious flaws , one is the sticky tongue and the other is the missing dick end.
For the first one, it is obvious that blowjob LORA has been tried on the right hand figure, this can usually be remedied with the right prompt.
The second problem I have encountered several times but have not found a consistent answer.
First of all I suspected a difference between the LORAs, but e.g. under WAN2.1 sometimes the problem occurs sometimes not.
For example, it does it here:
View attachment blowjob two blonde women .mp4

 

[Thierry29]

hello !
Civitai has more and more lora for Hunyuan. But I don't know if these lora are compatible with framepack.
i have not installed framepack yet, it will be my choice because for now I only have 8gb of ram.
I also need to find a good tutorial to install it with Comfy UI. my PC doesn't want a pinochio.
I start with Flux to generate the basic images. Using an image to image or text to image workflow. The goal is to have persistent characters.

 

[Kretua]

hello !
Civitai has more and more lora for Hunyuan. But I don't know if these lora are compatible with framepack.
i have not installed framepack yet, it will be my choice because for now I only have 8gb of ram.
I also need to find a good tutorial to install it with Comfy UI. my PC doesn't want a pinochio.
I start with Flux to generate the basic images. Using an image to image or text to image workflow. The goal is to have persistent characters.

Hi.
FramePack Studio uses Hunyuan LORA, WAN probably doesn't work, but I haven't tried.
I would not recommend ComfyUI so much with 8GB VRAM.
I only mentioned pinokio because it installs all the dependencies itself, python, torch, visual, etc. And you can put Forge, ComfyUI under it. On my machine I have Forge, Comfy, FramePack everywhere.
If you still want to do it separately, there are many places where you can find help to install ComfyUI. Also on the civitai site there is a downloadable .BAT file that does it all.
"The goal is to have persistent characters."
Yes it is for everyone.

 

[broho1234]

Kretua and Thierry29,

welcome to the discussion.

Kretua, thanks for the useful advice, I only started dabbling in AI video generation.
I will take a look at FramePack Studio with specific LORAs.

Do you have any pointers where to find adult-oriented LORAs ?

 

[JAI_INVESTIGATOR]

These is using WAN 2.1 with LORAs and some good workflows, a bit of upscaling and some editing. The LORAs and workflow were sourced from Civitai.
(I tried the leg spread one as well but it didn't go so well as there isn't any lora for leg spreading)

View attachment 1.mp4
View attachment 2.mp4

 

[broho1234]

These is using WAN 2.1 with LORAs and some good workflows, a bit of upscaling and some editing. The LORAs and workflow were sourced from Civitai.
(I tried the leg spread one as well but it didn't go so well as there isn't any lora for leg spreading)

View attachment 4927320
View attachment 4927321

Very nice results, especially the second one.

Did you use the same prompts as me ?

Also, what happens if you let the generator run for 5 seconds (or more) as I did ? I'm curious to see what the results are.

 

[Kretua]

Kretua and Thierry29,

welcome to the discussion.

Kretua, thanks for the useful advice, I only started dabbling in AI video generation.
I will take a look at FramePack Studio with specific LORAs.

Do you have any pointers where to find adult-oriented LORAs ?

The easiest way to get LORAs is from the Civitai site.
It is important to log in to the site. No registration is necessary, just google/discord or other accunt.
Feel free to do it with any , it is safe.
Then in the search box above, type LORA.
On the page that appears, in the left hand side bar, type in what you are looking for lora. In the example, I put Hunyuan Video in the box and found 100.
Clicking on any one of them will bring up its data page, and there you click on the download button.
In FramePack Studio, you find the ...model/lora directory and copy what you downloaded into it, then launch the program.
It is important to restart the program when it is running, otherwise you will not see what you have copied.
Afterwards you can choose which lora you want to use in the appropriate drop down menu.If I remember correctly, the default value is 1.
It's worth using only one LORA at first and experimenting with the 0.5 - 1 strength.
by the way, be sure to read the advice on the LORAs datasheet, there is a lot of help there, for example using the right keywords (bL0wj0b) or similar.

(The example shown by JAI_INVESTIGATOR was done with ComfyUI, but I don't want to belch into it)

 

[Pif paf]

ok so it works with rtx20xx also in my laptop. As I was doing other things during the 25 sec ai video generation, I don't have any hard data how slow was that process. But whole framework installation from pinokio site worked without any problems and errors.
As the movie was genarated from some amateur site photo i won't post it for my safety reasons.

 

[JAI_INVESTIGATOR]

Very nice results, especially the second one.

Did you use the same prompts as me ?

Also, what happens if you let the generator run for 5 seconds (or more) as I did ? I'm curious to see what the results are.

This is 5 seconds. Only took my poor GPU about an hour

View attachment 3.mp4
















I used a different prompt:
A woman is licking a man's penis. The woman is kneeling and looking up with her tongue out. Then she moves her head toward the penis and starts to lick its tip. The view is POV.

The easiest way to get LORAs is from the Civitai site.
It is important to log in to the site. No registration is necessary, just google/discord or other accunt.
Feel free to do it with any , it is safe.
Then in the search box above, type LORA.
On the page that appears, in the left hand side bar, type in what you are looking for lora. In the example, I put Hunyuan Video in the box and found 100.
Clicking on any one of them will bring up its data page, and there you click on the download button.
In FramePack Studio, you find the ...model/lora directory and copy what you downloaded into it, then launch the program.
It is important to restart the program when it is running, otherwise you will not see what you have copied.
Afterwards you can choose which lora you want to use in the appropriate drop down menu.If I remember correctly, the default value is 1.
It's worth using only one LORA at first and experimenting with the 0.5 - 1 strength.
by the way, be sure to read the advice on the LORAs datasheet, there is a lot of help there, for example using the right keywords (bL0wj0b) or similar.
View attachment 4928589
(The example shown by JAI_INVESTIGATOR was done with ComfyUI, but I don't want to belch into it)

Yes, I used ComfyUI and basically some good workflows from Civitai. I also got my LORAs from there (but I use WAN not Hunyuan)

 

[broho1234]

The easiest way to get LORAs is from the Civitai site.
It is important to log in to the site. No registration is necessary, just google/discord or other accunt.
Feel free to do it with any , it is safe.
Then in the search box above, type LORA.
On the page that appears, in the left hand side bar, type in what you are looking for lora. In the example, I put Hunyuan Video in the box and found 100.
Clicking on any one of them will bring up its data page, and there you click on the download button.
In FramePack Studio, you find the ...model/lora directory and copy what you downloaded into it, then launch the program.
It is important to restart the program when it is running, otherwise you will not see what you have copied.
Afterwards you can choose which lora you want to use in the appropriate drop down menu.If I remember correctly, the default value is 1.
It's worth using only one LORA at first and experimenting with the 0.5 - 1 strength.
by the way, be sure to read the advice on the LORAs datasheet, there is a lot of help there, for example using the right keywords (bL0wj0b) or similar.
View attachment 4928589
(The example shown by JAI_INVESTIGATOR was done with ComfyUI, but I don't want to belch into it)

Thanks, very useful details.

 

[broho1234]

Found this video:

"Make Realistic Animations of your Daz Characters with AI" by Generation3dX:

You must be registered to see the links

which is a very gentle introduction to the subject of image to video AI generation.

It discusses various aspects:
-Daz/Blender animations vs. AI ones
-running an AI locally vs on a paying site where there is censorship
-etc.

It also shows a simple workflow in ComfyUI and the results he gets from it.

 

[broho1234]

This is 5 seconds. Only took my poor GPU about an hour

View attachment 4930723
















I used a different prompt:
A woman is licking a man's penis. The woman is kneeling and looking up with her tongue out. Then she moves her head toward the penis and starts to lick its tip. The view is POV.

Thank you (and your poor GPU) for the effort.
Also, congrats on an excellent result, one of the best I've seen so far of this type (only started looking into this).

Your workflow is really good at keeping the face consistent ("undeformed") under different angles.

I especially like that it manages to "keep her cuteness intact" which is a requirement if this is to replace the traditional animation techniques (the subject of the Youtube video above).

 

[Kretua]

Yes, people who use ComfyUI, including me, use workflows like this with different variants.
For example, I created this one.
DAZ3d image -> ComfyUI WAN2.1
View attachment Brandi001.mp4
View attachment Brandi002.mp4

 

[broho1234]

Yes, people who use ComfyUI, including me, use workflows like this with different variants.
For example, I created this one.
DAZ3d image -> ComfyUI WAN2.1
View attachment 4931901
View attachment 4931904

Very impressive !