Vintage and Retro Celeb Fakes (Stable Diffusion)

[Papyrus1000]

c

 

[Papyrus1000]

Some Beverly D'Angelo images:

 

[Papyrus1000]

For anyone interested (I know some were), here is the Nicola Bryant (1983 era) model that I didn't think was good enough to upload. To be honest, I have never pushed this model, but I don't think it's as bad as I thought it was. I'm sure anyone more interested than me can really get something out of it.

 

[Toon-51-#]

Thank You

 

[Papyrus1000]

Thinking about seeing what can be done with Luma - not much, really, if your seed image falls foul of their censor, which is likely. But here's one Emma Samms image converted into a brief clip.

 

[Papyrus1000]

Though I was able to share the Emma Samms Imgur vid above, I am currently having difficulty sharing the Carrie Fisher/Slave Leia experiments (you're not missing much, most are pretty bad).

 

[LittleMouse1]

Anyway, while we're waiting for that fix, worth commenting that the only way to combat this kind of identity shift would be with traditional deepfake techniques taking over the latter part of any video that transitions out of the provided identity and into standard Stable Diffusion 'hot babe'.

Y.yes... for now. There's so much in flux. AnimateDiff allows you to use LoRAs all the way through, so one thing is to take the existing video and use it as a OpenPose model to redo the video. Or, you could do the deepfakery in SD with inpainting. That's one reason I like the backgrounds in these vids so much. All the inpainting is fairly "easy" compared to getting any good motion backgrounds, at least in my experience

Again, I think I might be your biggest fan at the moment. Yeah, from an ascetic point the vids are kind of weak, but all 100% generative videos are kind of weak right now. Within the framework of what you are doing, these are stellar.

 

[Papyrus1000]

Y.yes... for now. There's so much in flux. AnimateDiff allows you to use LoRAs all the way through, so one thing is to take the existing video and use it as a OpenPose model to redo the video. Or, you could do the deepfakery in SD with inpainting. That's one reason I like the backgrounds in these vids so much. All the inpainting is fairly "easy" compared to getting any good motion backgrounds, at least in my experience

Again, I think I might be your biggest fan at the moment. Yeah, from an ascetic point the vids are kind of weak, but all 100% generative videos are kind of weak right now. Within the framework of what you are doing, these are stellar.

Yeah, I'm trying to install AnimateDiff now. I just know, from experience, the sheer amount of troubleshooting and patience necessary to get one of these GitHub generative frameworks working and producing anything, so I have put it off a long time.

There's a lot of adoption friction with AnimateDiff. The installation process is no picnic, with the

You must be logged-in or registered for see links

non-linear and fragmented. It doesn't encourage anyone to get into it, and if I don't make some headway on the confusing install process soon (the environment frameworks download has currently hidden multiple crucial processes behind the opaque message '... (more hidden ...', with no indication whether or not the whole thing has stalled) I think I am going to let it go.

 

[Papyrus1000]

...one thing is to take the existing video and use it as a OpenPose model to redo the video. Or, you could do the deepfakery in SD with inpainting. That's one reason I like the backgrounds in these vids so much. All the inpainting is fairly "easy" compared to getting any good motion backgrounds, at least in my experience

The problem here is temporal consistency. The only way I could think to not have an SD-inpainted face across multiple frames 'sizzling' is to do keyframes and then use EbSynth to in-between among the keyframes, and then overlay the smoothed-out face motion across the original video, in After Effects. Or use DeepFaceLab or FaceSwap, which means weeks or even months of model training, 24/7.

For those that don't know (I know that doesn't include you, LittleMouse1), every image created in Stable Diffusion is created 'from scratch', with no consideration for temporal continuity, or what was going on in the previous image/frame. I'm not aware of any system that has solved this, except EbSynth workarounds, which are really kludgey and high effort.

 

[Papyrus1000]

I'll put some Luma gens that aren't total disasters in this post. Here's a pretty static Catherine Bach:

 

[dbcoop13]

So as you might have noticed, I created a

You must be logged-in or registered for see links

(use 0.5 strength, and optionally use 'retropornamateurs' as trigger word, but it will probably work without it), with no smartphones. Here's the first attempt with Carrie Fisher.

View attachment 2355579 View attachment 2355586 View attachment 2355911
View attachment 2355990

These are great. Has anyone made this style for 80's Molly Ringwald?

 

[Papyrus1000]

As ever, adding analog noise helps (though it doesn't help hands)...

 

[Papyrus1000]

Late-stage Phyllis Davis. With no prompt, it did a better job. Prompts seem to be mostly remedial, ie, 'Don't do that!'.

 

[Papyrus1000]

I have just figured out that Imgur is auto-deleting the pretty tame Nicola Bryant Luma vid every time I upload it. Therefore the only way to pass it on is at MediaFire

 

[FilthyPervUK]

First off, @Papyrus1000, let me just say that these are great and I really appreciate you sharing your work.

The Carrie Fisher slave bikini fakes are very good, though I know what you mean about once she moves her head that her face doesn't really look like Carrie anymore.

The Helen Mirran one is very good though. That's really almost there. Ditto with the Catherine Bach and Phyllis Davis ones too.

Anyway, great work. Keep 'em coming (especially the Slave Leia ones).

 

[SarahTurner]

I have just figured out that Imgur is auto-deleting the pretty tame Nicola Bryant Luma vid every time I upload it. Therefore the only way to pass it on is at MediaFire -

You must be logged-in or registered for see links

.

It's really not that great, though. Grab is below, but it is practically identical to the source pic (very limited movement in this vid).

View attachment 2534844

Yet Imgur had no problem with this video:

You must be logged-in or registered for see medias

Good job, buddy

 

[Papyrus1000]

This project is now over, for me. Have fun with the models.

 

[Shoestring]

So long and thanks for all the flesh

This project is now over, for me. Have fun with the models.

 

[Sigmatorn]

This project is now over, for me. Have fun with the models.

Thanks for your contributions!!! Your models are amazing!

 

[Sigmatorn]

@Papyrus1000 are you planning on doing something with Flux? Would be really interesting to see your LoRa results and your experiments using it.