Fan Art Sexbot [LlamaMann Games] - Fanart

[BlondieAngelEyes]

At least knowing that Luma can decently get a walk animation is nice to know.

LUMA Cheerleader Walk


LUMA KREA Enhanced Cheerleader Walk [AIS 0.6][Render]

Sadly, Bailey's hands obtained the AI curse of too many fingers. It's also nearly impossible to keep Ashley colored cateared headset attached. Which isn't surprising since I doubt there is enough video training data with those type of headsets.

Ashley looks like she is coming to kick some ass! ha Cool AI

 

[DoubleXX]

Trying to make a Steph LoRA(first time trying something like this). In the mean time here some Ashley.

edit: If I make more I'll post them in this comment. I have added more twice.

 

[Hieronymous]

Wow!
These look very realistic, one could easily think she's a real woman from a quick glance.

This is probably a best case test since that Sam face is so incredibly realistic. I chucked broho1234 Sambot image into the HEDRA AI along with some audio. Audio is from the Milfylicious game.

 

[Hieronymous]

Since MAGNIFIC is beta testing their ReLight tool. I tried it out. It's quirky, but has serious potential in the future of game development from what I'm seeing.

Daphne Test:



So, I went from KREA upscale -> MAGNIFIC Relight = Muddy ReLit picture -> KREA upscale to fix all messed up relighting problems and the fine details that got removed. Daphne's shoe got messed up, but otherwise the relighting seems to look pretty passible.


Ashley Test:



Same steps as above. KREA enhance > MAGNIFIC RELIGHT > KREA enhance


GROUP Test:



Same steps as above. KREA enhance > MAGNIFIC RELIGHT > KREA enhance
Only thing of note seem that ReLight hates pink/red for some reason I don't understand. As it gets removed in almost every relighting I've tried. As usual, Daphne's light up headset gets vanished 50% of the time.


CHEERLEADER Test:

Locker room


Football Field


Not perfect, but if you needed to rapidly iterate some entirely new backgrounds to concept out some new locations for our characters in phase 2, this could give some decent ideas very quickly.

 

[broho1234]

This is probably a best case test since that Sam face is so incredibly realistic. I chucked broho1234 Sambot image into the HEDRA AI along with some audio. Audio is from the Milfylicious game.

Very impressive !

 

[broho1234]

Since MAGNIFIC is beta testing their ReLight tool. I tried it out. It's quirky, but has serious potential in the future of game development from what I'm seeing.

Daphne Test:

View attachment 3786201 View attachment 3786205 View attachment 3786244

So, I went from KREA upscale -> MAGNIFIC Relight = Muddy ReLit picture -> KREA upscale to fix all messed up relighting problems and the fine details that got removed. Daphne's shoe got messed up, but otherwise the relighting seems to look pretty passible.


Ashley Test:

View attachment 3786262 View attachment 3786266 View attachment 3786269

Same steps as above. KREA enhance > MAGNIFIC RELIGHT > KREA enhance


GROUP Test:

View attachment 3786295 View attachment 3786296 View attachment 3786298

Same steps as above. KREA enhance > MAGNIFIC RELIGHT > KREA enhance
Only thing of note seem that ReLight hates pink/red for some reason I don't understand. As it gets removed in almost every relighting I've tried. As usual, Daphne's light up headset gets vanished 50% of the time.


CHEERLEADER Test:

Locker room
View attachment 3786317 View attachment 3786320 View attachment 3786323

Football Field
View attachment 3786317 View attachment 3786331 View attachment 3786332

Not perfect, but if you needed to rapidly iterate some entirely new backgrounds to concept out some new locations for our characters in phase 2, this could give some decent ideas very quickly.

Very nice !



The image above, is it taken from the indian version of Sexbot ? </jk>

 

[Hieronymous]

Very nice !

View attachment 3810694

The image above, is it taken from the indian version of Sexbot ? </jk>

Nah. Relight just evens out all the skins tones seen on photo. I'm guessing the inclusion of Steph might have fudged that, or that the prom initial photo was very dimly lit and fooled the AI to think they were all the same skin tone.

The AI doesn't care for ethnicity, it's only looking at an image depth map and remapping the lighting from that. While AI is fast, half the results you can throw away as bad, 25% have AI issues of forgetting anything they can't see in 0.5s and the remaining 20% have issues with hands. The "Good" results that get posted on social media is the 5% that shows that there might be a practical use case for AI photogen.

In order to make AI useful, I can see why people have to use workflows on things like ComfyUI to make practical usable results on a consistent basis.

Plugging in things and rolling the slot machine leads to a ton of garbage results wasting a ton of time. Fixing borked results is usually more trouble than it is worth most of the time.

 

[broho1234]

Nah. Relight just evens out all the skins tones seen on photo. I'm guessing the inclusion of Steph might have fudged that, or that the prom initial photo was very dimly lit and fooled the AI to think they were all the same skin tone.

The AI doesn't care for ethnicity, it's only looking at an image depth map and remapping the lighting from that. While AI is fast, half the results you can throw away as bad, 25% have AI issues of forgetting anything they can't see in 0.5s and the remaining 20% have issues with hands. The "Good" results that get posted on social media is the 5% that shows that there might be a practical use case for AI photogen.

In order to make AI useful, I can see why people have to use workflows on things like ComfyUI to make practical usable results on a consistent basis.

Plugging in things and rolling the slot machine leads to a ton of garbage results wasting a ton of time. Fixing borked results is usually more trouble than it is worth most of the time.

Thanks for the detailed reply, it's good to learn from someone skilled in AI image generation.

 

[Hieronymous]

Thanks for the detailed reply, it's good to learn from someone skilled in AI image generation.

It's mainly why I threw in the tests for the group shot of the girls. AI seems to work well on a singular person most of the time, but throw anything it's not well suited to handling yet and you can get some weird results.

I'd like for AI to be at the level where making a game with renders isn't a pain in the ass, but it's not quite there just yet. Like it feels like its about 75% cooked and I'm not sure if the last 25% will quite get there.

I'm just messing around with the tools because I can see a future where some aspects of AI might be useful in some specific use cases. I think the problem is that people think AI can do everything like a multitool. If you don't know where AI's weak spots are, then you can't know how to the best places to utilize are and where it would fail at.

 

[majdaddin]

... and I'm not sure if the last 25% will quite get there.

I think the problem is that people think AI can do everything like a multitool.

Well known 80/20 problem meets humans' infinite techno optimism. Cars can stay in their lane autonomously, so next year they will be perfectly fully autonomous. Or we did the first 80% quickly, and the next 20% will take another few decades. I'm very curious where we are in a few years, when AI will have to have proven itself. I don't think the good stories will be from AI, nor most cars, nor the best customer service, nor the best 'girlfriends' ;-) But maybe that's just my techno pessimism.

Some of the AI art in this thread IS amazing though. It needs a human supervisor, but some of it is really really good.

 

[Hieronymous]

Daphne Dance test > Motion powered by S2Haru Kokain dance

What I mean by half baked. This was stupidly fast to do [1min], but as usual. Hands, messed up. Headset. Gone. Turning around/side shots have problems.

But there are tiny glimmers of hey, that looks pretty decent interspersed into that.

I'll probably see if I can run some testing on if a higher resolution starter image helps the output. That or find a dance that doesn't have any turning around and the face facing forward the entire time.

Pulled a quick KREA enhance to see what would happen to the block hands.

Looks even more passable now. AI just keeps selling me on hope, so it's not hard to see why people keep dumping money into it.

 

[Hieronymous]

Seeing as there was some hope of translating dance moves to the models fairly well.
I gave it a shot on a much faster dance along with camera motion with Ashley.

Some bits of her hands gets lost given how fast the moves are coming out, but overall, much better than I would have originally hoped.

*notes* I originally wanted to try cheerleading dances, but AI's have no hope in figuring out what pompoms are sadly. Maybe in a few more years.

 

[Hieronymous]

Ashley Walk Test Background Retained

Mapped the Ashley model on top of a video while keeping the background of the video. It could be a bit cleaner around the edges of the model replacement, but passes the casual glance.

Her shirt is messed up in the back, but considering the AI couldn't see any of it, that shouldn't surprise anyone. The AI also guessed at the length of Ashley's hair, which is why you see the artifacting from the original model's hair swinging since I picked her for the same hair length of the Ashley model. The walk is a bit awkward because the original video, the model has heels on, while the replacement is in sneakers.

Not sure yet if just leaving the background white, inserting the background after the fact would look better yet.

 

[yourmomma]

Nah. Relight just evens out all the skins tones seen on photo. I'm guessing the inclusion of Steph might have fudged that, or that the prom initial photo was very dimly lit and fooled the AI to think they were all the same skin tone.

The AI doesn't care for ethnicity, it's only looking at an image depth map and remapping the lighting from that. While AI is fast, half the results you can throw away as bad, 25% have AI issues of forgetting anything they can't see in 0.5s and the remaining 20% have issues with hands. The "Good" results that get posted on social media is the 5% that shows that there might be a practical use case for AI photogen.

In order to make AI useful, I can see why people have to use workflows on things like ComfyUI to make practical usable results on a consistent basis.

Plugging in things and rolling the slot machine leads to a ton of garbage results wasting a ton of time. Fixing borked results is usually more trouble than it is worth most of the time.

It's not just the lighting, the AI pretty much made them a variation of a clone. The facial features are nearly the same.

Otherwise kind of cool. The animations are interesting.

 

[Hieronymous]

Testing out some Michael Jackson moves to limit test the AI character models.

1982 Thriller Dance moves on the Main Character.

I couldn't get a good copy of the Main Character in the Ghostbusters outfit that showed his entire body. So I used him in the prom suit.


1995 Dangerous Dance moves on the Main Character.

*Notes* AI can't handle some of Jackson's most famous dance moves sadly. Moonwalking bugs out [If AI could barely handle walking properly, how could it possibly understand a moonwalk?], and so does the super spin where Michael spins in place.

=====================================
Melvin Michael Jackson Dance testing:

1987 Smooth Criminal moves on Melvin.

I know Melvin's Tux is messed up. The only photo of him with decent lighting had one of his arms across his chest.

I figure as an idea to mull over is that phase 2, our characters gain some dance moves.
Might try seeing if KREA will fix up the hands tomorrow, otherwise this looks fairly promising so far.

 

[Hieronymous]

Bailey Tomb Raider AI motion KREA Test

Still not entirely sold on letting the AI mess around with the background. It removed a ton of detail on the stonework / bark of the tree and decided it was vitally important to animate a tiny bit of wind movement to the leaves of the vines.

For those wondering, the motion is being driven by a recent tweet by Ireland Reid when showing off her Lara Croft cosplay.

 

[Hieronymous]

Someone asked me what the bad AI generations looked like:

So here's todays cursed images:
Since other AI tools started rolling out their takes on relighting. I started testing them out.

Used Daphne to see if the hot pink color could be retained. Nope. None of the relighting keeps neon colors.

Starting Image

As usual, the cat ears of the headsets gets wiped. Oddly, the background isn't usable either. Though I guess that's expected for a another AI tools first try at relighting.

Starting Image

This one ended up with black pants, mostly looking ok.
This one oddly removed her pants entirely. Pervy AI.
This one turned Daphne into a ghost. Yes. The AI hallucinated that.

I think I'll need to grab some HD room photos to feed into the AI to get less buggy generated images. I could probably use the hallucinated image generation for dream sequences I guess.

 

[broho1234]

Someone asked me what the bad AI generations looked like:

So here's todays cursed images:
Since other AI tools started rolling out their takes on relighting. I started testing them out.

Used Daphne to see if the hot pink color could be retained. Nope. None of the relighting keeps neon colors.

Starting Image
View attachment 3822789 View attachment 3822791 View attachment 3822792
As usual, the cat ears of the headsets gets wiped. Oddly, the background isn't usable either. Though I guess that's expected for a another AI tools first try at relighting.

Starting Image
View attachment 3822800
View attachment 3822801 This one ended up with black pants, mostly looking ok.
View attachment 3822807 This one oddly removed her pants entirely. Pervy AI.
View attachment 3822809 This one turned Daphne into a ghost. Yes. The AI hallucinated that.

I think I'll need to grab some HD room photos to feed into the AI to get less buggy generated images. I could probably use the hallucinated image generation for dream sequences I guess.

You're saying these are bad, I think they're amazing - I'm looking at the faces and bodies, not the background.





The faces in the two images above are very realistic (esp. the 1st one), way more than the original.
The bodies are quite impressive too, especially the 2nd one where the pants were removed - it seems that the AIs are getting better at giving us what we want .

In any case please post more, I'm looking forward to seeing what you come up with.

 

[Hieronymous]

Managed to clobber together a 10s LUMA loop of Emily and Ashley confronting each other.


Tried to see what KREA would do to the clip. This is at 25% AI strength. Lots of loss in talking detail. But the models/hair/fabric looks nicer.


KREA at 45% gives really nice hair simulation, but almost all the lip movement is gone at this level of AI strength. And their eyes seems unfocused or at times cross eyed.

 

[Hieronymous]

Since KLING AI finally went public today. I can finally compare how well it does against LUMA.

KLING AI walk test

Major loss in details compared to LUMA. But, the movement is pretty natural looking and hair/fabric movement is pretty accurate. As usual, Daphne's headset lost an ear.

**Notes** double checked the KLING parameters. Only High speed low quality video generations available right now. Will retest when I can try High Quality video generations.

Still gives me a general idea of what KLING can do.