Is it possible to create own JAV using AI technology?

[therealpmuk]

LOL sorry, that made me laugh!

Maybe it's the model you're using, is it a base LTX? There is another newer(ish) one I saw that is supposed to be better with NSFW because it was trained only on people. I think... Also, have you tried using a LoRA that's NSFW if there are any for LTX? I'm still using Wan. After I finish my trailer I'll take a crack at your images and see if Wan has any better luck.

According to copilot, the neck-grabbing is down to LTX having been trained on socila media clips which are mostly people having arguments and slapping each other around. Go figure.

Yeah, I'm using a bas LTX model, but I'm going to move on to CogVideoX‑5B in comfyui just as soon as I can find enough disc space for the installation.

I am still weirded out by the fact that if someone's face is not in frame in the source photo it is hallucinated by the model. In my alley scene above, his head disappears for a moment or two behind hers. When I try to stitch on another clip, it emerges completely different!

 

[Casshern2]

I had to edit my post right before yours, I placed a link to the clips Wan was able to produce...in case you missed it, your last reply started on page 2 for me.

 

[zombierambo]

This will be possible within the next few years but currently the generation technology is not mature enough

 

[ramp_it_up]

I couldn't get him to grab her ankle after 20 attempts, either. But at least he didn't try to strangle her. What an awkward guy!

19.2 MB file on MEGA

mega.nz

You made that?

 

[Casshern2]

You made that?

Using his image as the source for image-to-video and a Wan 2.2 video model and ComfyUI, yes. Those are all straight one after another generations just merged together (outside of that workflow). I left all attempts, even ones that looked wonky. For full disclosure. Honestly, I thought Wan would have got it early on, but those are the first (and only ) 20 attempts. Maybe with the right prompting and luck with seeds it would make him grab her ankle one day!

 

[Casshern2]

This will be possible within the next few years but currently the generation technology is not mature enough

You're right regarding making a full-length JAV title. There is great stuff out now like the new LTX...hopefully Wan will bring it again soon, but, with lip sync and extended video times already a thing, it won't be too long. That being said, even now I think one could piece together something that resembles at least a quasi or alternative version of what one might call a JAV video. Not that great quality, but passable to some.

 

[Casshern2]

It's done!

The disclaimers/disclosures:

• all image sources are AI generated and are not meant to represent any actual people
• no effort was made to maintain character consistency, this was merely a proof of concept
• the intended free soundtrack ended up too short for the video length so it was replaced with another free soundtrack *
• the ending title was meant to be a fade to black, but Wan 2.2 didn't like a prompt as simple as that.
• any anatomical anomalies were noticed but ignored in the interest of time
• the time cues with the audio for some of the transitions were nothing more than luck
• no video editing software used, only Yamb, MKVMerge, MKVExtract, ffmpeg
• and oh yeah, it's uncensored

* my checklist was listing the intended number of seconds for each clip. But, I shouldn't have been surprised that each clip would not truly be exactly 2 seconds, or 4 seconds, and so on. Likely 2 seconds became 2.3 seconds or something like that for all the video generations, so the end result after combining all parts was longer than anticipated.

This was fun to do! I truly think, like @zombierambo says, as fast as this stuff is progressing it won't be long before something pretty passable as a form of JAV content will be possible. Just need the time and resources, really. With all the new LTX releases and supporting releases (lip sync, extended video, etc.) something pretty good can be produced right now. Probably even with Wan using proper prompting and good source images.

So, @Muz1234, yes, it is possible to create your own JAV using AI technology. If anything, currently, only for yourself and only what you consider to be JAV. Go learn this stuff and make what you want.

This will likely disappoint! Haha! But it was certainly quick and dirty. I didn't mess with the audio levels, so you might adjust your sound in case you have it loud, this might blare for you. Enjoy!

​

https://mega.nz/embed/KIlQjSbC#pvXmCShvCeMqC_C-iEw3Ag5NMxE_N4b19qJaTsB5L5g

 

[zombierambo]

It's done!

That is truly an impressive result, I think you have clearly established that there's no question at all whether this will happen but only how long it will take.

 

[therealpmuk]

I couldn't get him to grab her ankle after 20 attempts, either. But at least he didn't try to strangle her. What an awkward guy!

19.2 MB file on MEGA

mega.nz

You put me to shame @cashern2, and I didn't intend for you to spend so much time trying to prove me wrong, espevcially as is wasn't strictly JAV-related.

Great result though! All we need now is a model trained on ankle-grabbing. I dare say specialist fetsih, softcore and hardcore models will appear after a while.

 

[Casshern2]

Well done mate! This really shows potential, it's clear you've put a lot of effort into it, but considering the pace of development of AI models, it's quite possible that everyone will start producing their own videos in 2-3 years.

Thanks, it was a fun process! It's one of those things where you do something enough times, it eventually doesn't feel like a chore at all. I've been working in this stuff for a bit now, so it's a familiar hobby for me. And I hear you! 2-3 years if that for putting together a good length piece. I had an uncomfortable conversation with a guy I know about this type of stuff and he hit the nail on the head, in my opinion. For whatever content there is and you look at...if you're watching for, ahem. reasons...it's maybe a few 1-2 minute places OR less within a JAV title that does it, if you know what I mean. So, with this technology whatever your Mortal Kombat move is, you can just create clips of that to your heart's content.

(for those who don't know the reference of the Mortal Kombat move...it's the FINISH HIM!

 

[Casshern2]

That is truly an impressive result, I think you have clearly established that there's no question at all whether this will happen but only how long it will take.

Thanks! I have to think what some are doing right now who have better computers, better graphics cards, and a ton of time. With the right ingredients any niche, mainstream, or kink content can be produced. Hopefully this later won't have a negative impact on the real studios, now that I let that hit. Sales might drop if it becomes easier and easier to do. BUT I think the industry is safe for now.

 

[Casshern2]

You put me to shame @cashern2, and I didn't intend for you to spend so much time trying to prove me wrong, espevcially as is wasn't strictly JAV-related.

Great result though! All we need now is a model trained on ankle-grabbing. I dare say specialist fetsih, softcore and hardcore models will appear after a while.

Hey, friend! I hope I didn't give that impression, I wasn't trying to prove you wrong, just seeing if maybe a different model would have better luck producing what you were after. At least he eventually did touch near her ankle, but I think I know what you were looking for. Wouldn't mind seeing that myself!

And it wasn't much time at all, really. Just gave it a prompt queued 2 generations. Check the results, change the prompt to see what else might work, rinse and repeat. This stuff is largely unpredictable, as you know. Heck, maybe if I had stayed with the first prompt and let it generate 20 clips, one of them would have done it! Anyway, keep at it! LTX is the popular one again. If you had seen my early posts, that's what I started with then switched to Wan. I'm so comfortable with it, but I may start looking back to LTX with all the new bells and whistles. Hopefully Wan steps up again before long.

 

[Casshern2]

For any interested I posted the creation of this thing over in the thread where that belongs.

AI Video Generation - General Discussion, Tips, Tricks, Frustrations and Showcases

 

[r00g]

Wow, impressive work!

 

[spunkyD]

I know one of the current limitations is clip length, I think we're at 5-6 seconds? I've heard about a model called LongCat that can do videos up to 4min. Have you tried it?

Im really excited about this field, I have your mega.nz links open in another tab, gonna check them out.

 

[Casshern2]

I've heard about a model called LongCat that can do videos up to 4min. Have you tried it?

Hey, there, no I had not heard of that one. I'll do some research on it, thanks for the heads up! And, you're pretty much right, for the most part, depending on the size of the video frame, from 5-8 seconds per clip is the advertised default lengths for most models, unless you have a great graphics card!. But that got me curious about average length of cuts in scenes of a JAV title vs normal content like a TV show or movie.

Did a quick Google search: "in a modern TV show, what is the average length in seconds is each cut scene?"

AI Overview: In modern TV shows, the average duration of a single visual cut (or shot) is approximately 2.5 seconds. The "Cut" (Shot Length): Popularized by streaming platforms like Netflix, the rapid editing pace of modern shows features a new shot every 2.5 seconds to keep the audience's brain engaged and maintain story momentum.

2.5 seconds, wow! It's so normal that I never even realized that. Out of curiosity I used FFMPEG to count the number of cut scenes in my trailer to see if it was accurate. It was! Counted 32, that's how many shots I had including title cards. So now to run that against just one (sex) scene from a random full-length JAV. I'll post what I find, this is interesting to me.



EDIT: Finished already!

I used the second scene from the recent ALDN-593 My Sister-in-law (lol) Or Rather, She's All Mine During The Day Lol Honma Yuri.

The scene is approximately 34min in length. Using the script I used on my trailer this thing brought back 24 cut scenes. In a 34min full scene that's a whopping average 1.41 minutes per cut! That's like 84 second cuts! But that goes to show...sexy women in sexual scenes are not meant to be watched in quick cuts on Netflix haha! Guess it's best I stick to creating trailers of this stuff if anything!

EDIT 2: Had a thought this morning

This morning I had the idea of trying the same thing on an IV title, surely Image Video directors have a very different cutting style than a feature director. The results were as expected (pretty much).

I tried the same process on the recent MMR-AA377 Dream Of A Glimpse/Shimizu Renka

I analyzed the first scene which was approximately 14min long. This came back as having 42 cut scenes, which still is about an average of 20 seconds per cut scene? Is my math right?

 

[The Last Ninja]

Also worth mentioning are the continuation workflows where the model outputs the last frame of a generated clip, which can then be used as the starting point for the next generation. In theory, that means you're not really limited by the model's native clip length and can extend a scene as long as you can maintain consistency. Of course, keeping character and scene consistency over longer sequences becomes the real challenge.

Sorry if this has already been brought up earlier in the thread — I only skimmed the last few posts.

 

[Casshern2]

Sorry if this has already been brought up earlier in the thread — I only skimmed the last few posts.

It hadn't but even if it had been, that's always okay! We've all skimmed things and maybe got dinged for it now and again, haha! Have you done any video generations you'd like to share? Or tips, pointers, workflow ideas? I know I'd love to see some! My video generation post is lonely and I keep finding myself talking about the subject in other threads. Feel free to skim that one and join the conversation!

Speaking of, and I'll document in the other thread if I find any success, but I've been cooking up a workflow idea in my head for some time around long videos. Being a programmer (of sorts) I of course first thought of something recursive, but early on all I read was that you could not recursively send the last frame to the image input in a single workflow set of nodes to generate a new video. Probably different now but I've had this idea so long I'm going to try to figure it out, even if there is now a solution already.

 

[The Last Ninja]

It hadn't but even if it had been, that's always okay! We've all skimmed things and maybe got dinged for it now and again, haha! Have you done any video generations you'd like to share? Or tips, pointers, workflow ideas? I know I'd love to see some! My video generation post is lonely and I keep finding myself talking about the subject in other threads. Feel free to skim that one and join the conversation!

Speaking of, and I'll document in the other thread if I find any success, but I've been cooking up a workflow idea in my head for some time around long videos. Being a programmer (of sorts) I of course first thought of something recursive, but early on all I read was that you could not recursively send the last frame to the image input in a single workflow set of nodes to generate a new video. Probably different now but I've had this idea so long I'm going to try to figure it out, even if there is now a solution already.

Here's a link to a workflow based on this exact idea. It's not for the base WAN 2.2 checkpoint, though, but for the specialized SmoothMix WAN 2.2 model:

https://civitai.red/models/1847730/...irst2last-framemmaudio?modelVersionId=2091039

I'll attach a screenshot as well. You can see the GGUF setup there, along with the models I'm using in the loaders.

Unet Loader (GGUF) / High:
smoothMixWan22I2VV20_highQ5KM.gguf

Unet Loader (GGUF) / Low:
smoothMixWan22I2VV20_lowQ5KM.gguf

HIGH LORA LOADER:
wan_22\high_noise_model.safetensors

LOW LORA LOADER:
wan_22\low_noise_model.safetensors

I usually use the Q5KM versions of the checkpoints because my GPU can still handle them, and in my experience they produce slightly better results. I've also tried the Q4KS versions, and they're quite good as well while requiring less VRAM and generation time.

What really surprised me is how good the transitions between segments can be. By some kind of "magic," WAN seems to infer the direction of motion from the last frame, so the joins between segments often don't produce noticeable jumps. I might upload a sample video later.

I've also come across some truly monumental workflows on Civitai, where entire scene sequences were chained together, each with different LoRAs. I didn't study them in detail, but I assume the trick is that the last frame is automatically passed to the next stage, allowing an extended video to be generated with a single click.

By the way, you can drag the screenshot directly into ComfyUI, and it will load the workflow automatically. Of the three preview windows at the bottom, the first one shows the "raw" video while the interpolation and upscaling are still running, the middle one displays the final processed video, and the last window outputs the final frame that you can use to continue the generation.

 

[The Last Ninja]

I made a short demo video. It's made up of four chained segments:
https://mega.nz/file/ImVjyBTS#MI5qGKbYRMGkBPu43ro6xMapqhGCCObYmDBIjWUicLw

Adding a cross-dissolve effect between the cuts can make the transitions look even smoother:
https://mega.nz/file/IjsR1RzZ#-W4khFrUGEuyHOL4Pcbg6RcoP5wKM6v-XnszItlAAK0

To be honest, I've only made one video like this before, and I can already see that I got pretty lucky with the transitions in that one. In this example, they're a bit more noticeable. One thing I've learned is that it's worth trying to end each segment with less dynamic motion, because fast movement tends to make the transition points more obvious.

To be fair, I only generated two or three versions for each segment. If I'd generated more, I probably could have picked transitions that matched up even better.