Susanna Reid (and friends) Facefusion Fakes

sausages

Well-Known Member
My first attempt at faking with this software. I'm using Facefusion 3.0.0, modified to remove the NSFW restrictions.

I'm not sure if this is classed as deepfake, AI, or a mixture of both? Stable diffusion is AI, but it's not generating new images from scratch. You need to supply the headshot and the target, like you would with a deepfake.

Anyway, Susanna Reid.
Links, technical bollocks and how to patch in the spoiler.
Main page is here,
Installation looked really complex so I bit the bullet and paid $20 to the developer (buy me a coffee) for the . It's hassle free and installed everything I needed with a couple of clicks.

It's got a GUI that opens in your browser, but it looks like a steep learning curve to understand all the different options.

Nothing worked for me until I found out it was blocking NSFW content.

The way I got around it without having to start from scratch was:
  1. Download the master zip from here
  2. Delete the folder located at "C:\FaceFusion\3.0.0\facefusion\__pycache__"
  3. Delete the file "C:\FaceFusion\3.0.0\facefusion\content_analyser.py"
  4. Unzip the facefusion-unlocked file, and copy/paste the files directly into the installation folder.
You need to make sure you're replacing the files directly.
I've checked the scripts and they look fine, just amended python instructions telling the NSFW model to always report back as false.

I didn't realise it restricted nudes at first, and it took me 3 days of countless reinstalls and reading every document I could find to work out why the output and the GUI didn't work like it said it should. Found the answer on Reddit, patched it, and it fixed everything with no need to reinstall.
There are other ways to install the whole package, Pinokio is one I tried when I was trying to find out what was going wrong, but I kept going back to the original Windows installer because it was faster and works without needing to anything else.
960x540 face swap with no audio. It rendered 1 frame every 10 seconds on average, which is slow as fuck and worse performance than I expected.
Good result though, first frame is a bit off with her big eye, but the rest seem okay. Once I'm familiar with how each option works it should speed up.
View attachment srjw_sausages.mp4

Some bonus pics. I need to fix the pixel settings and test different options, but it looks promising.
sr01.jpgsr02.jpgsr03.jpgsr04.jpg
 

baronvonhappypants

Active Member
This one has audio. I haven't looked into voice swapping yet, but the lip sync in Facefusion works really well.
A couple of glitches when the head comes back into frame, but processing time is getting better.
Susanna has a birthday treat for you.
View attachment 2705701
Voice swapping is something I looked into a while back but there didn't seem to be a quick and easy solution yet, there may be now though. I've used Roop Unleashed for videos and it produces similar quality but is considerably quicker.
 

sausages

Well-Known Member
I've had a few crashes trying to make videos recently. There's an inbuilt face enhancer tool that makes the quality amazing, but it needs more processing power than my laptop can handle, so I've had to stick to images for now. I might have to buy myself something better for Christmas.
I'll edit the thread title because I've tried a few fakes with other celebs.

Nigella Lawson.
These made her look younger than the input. It got the model's expression perfect, but I think the features need more detail.
This is the source image I used, with results underneath.
23091826crop.jpg
nigellanad001.jpg
nigellanad003.jpg
nigellanad005.jpg
nigellanad007.jpg
 

sausages

Well-Known Member
Good stuff. You should try Nigella as a blonde, could pass as Victoria Cohen
If I can find a good model I'll try both of them. They come out better the more their head shape matches, so it's finding a good target that takes a while.
 

sausages

Well-Known Member
And a few images. Still experimenting to find out what settings and type of pics work best.
The celeb face seems less important than the target, if the target isn't a great match it doesn't work well.
It always matches different target expressions from a single pic. That's the clever bit.
Anyway, some more Susanna. Click the pics for full size.
Source:
sr-source2.jpg

srcowhq.jpg srhq.jpg

Cara. Mixed results, last pic is good with her smoking.
hhgf.jpg
carad001.jpg carad002.jpg
carad003.jpg cdhq.jpg
 

sausages

Well-Known Member
A bit off-topic, but it's all about face swaping so...
When I was admin of DeathAddict, someone suggested a fake celeb section when the deep learning apps first came out, but it was so hard to find decent content I put this together. I called the bad videos Derpfakes because they always look derpy.
You must be registered for see images


Must be 7 years ago I made this, things have improved a lot.

Derpfakes Part 1: The Good, The Bad & The Fugly
View attachment DerpFakes Part 1 - The Good, The Bad & The Fugly.mp4
 
Top