Avatarify lets users run realtime deepfakes on live video calls
inputmag.com
inputmag.com
Train a model locally ahead of time and upload it to a server, then whenever you have a call scheduled the model is downloaded in advance by the other participants.
Now, instead of having to send video data, you only have to send a representation of the facial movements so that the recipients can render it on their end. When the tech is a little further along, it should be possible to get good quality video using only a fraction of the bandwidth.
Downside could be some uncanny valley if the models are not very high quality.
But if I had to make a prediction, I'd expect we'll get much more value from higher bandwidth, ultra high definition streaming and features like 3d cameras / virtual reality. I think we have a tendency to really underestimate how important high definition is for human communication.
Low latency is probably more important to me.
Recently I seem to have a 3 second delay on many VC calls at work (and just for me it seems), and I either end up interrupting people or feeling reluctant to talk at all since it becomes impossible to time gaps and conversations right.
Despite that I get a crystal clear HD picture for all participants, but I'd happily sacrifice video quality (in fact I'd accept audio only in some cases) to get a more real time experience (disabling video doesn't seem to have any effect).
If you're really willing to sacrifice video completely, at least for Zoom, and probably for lots of other videoconferencing solutions, you can call into meetings with your phone. In fact, I think Zoom allows you to join with the computer for video and the phone for audio, which might be the best of both worlds.
Slight issue in Toronto has been cellular system overloading and calls not being completed. But once on, no problem.
I can’t blame the providers though. How could they have predicted that people would use the service they’re paying for?
That can be controlled, since these compression algorithms usually work by making a prediction and sending the difference between the prediction and the actual value.
That works both for lossless compression - where the difference is sent in full - and lossy as well - where only the most important part of the difference is sent.
https://people.xiph.org/~jm/demo/lpcnet_codec/
Sure enough, it was discussed on HN when it came out last year. I think I missed it then.
For those who didn't catch this from the URL, this is by Jean-Marc Valin, of Speex and Opus fame.
An utterly brilliant great idea!
One of the premises of the novel's universe is that computational power is generally absurdly plentiful, but communications bandwidth over interstellar distances is not. Most communications are in plain text (modeled after USENET) but in some cases, "evocations" are used to extrapolate video and audio from an ultra-compressed data stream.
The trouble, of course, is that it's not very obvious what aspects of the image you're seeing are real, and what aspects were dreamed up by the system doing the extrapolating.
It would be quite obvious unless the raw data before extrapolating is destroyed, for which there are no reason nor is it possible to stop others in the vincinity from receiving this raw data.
But saying more would be spoilery...
> WaveNetEQ is a generative model, based on DeepMind’s WaveRNN technology, that is trained using a large corpus of speech data to realistically continue short speech segments enabling it to fully synthesize the raw waveform of missing speech.
I don't think you need to train for each person specifically, you can just train a model for all heads, then maybe transmit a few high quality pics when the call starts, and interpolate from that afterward.
To a human, for example, the exact pattern of every strand of hair isn't important at all -- all that matters is that the hairstyle and hair color stays the same.
The algorithm can also not worry about encoding and re-constructing skin blemishes because humans would possibly actually enjoy not having to put on makeup for a video call.
The MPEG-4 part 2 actually had something like that, called "face and body animation (FBA)". As far as I know, there are no implementations in widespread use.
https://www.nvidia.com/en-us/geforce/news/nvidia-dlss-2-0-a-...
Original repo code that Avatarify is based on called First Order Model: https://github.com/AliaksandrSiarohin/first-order-model
Short video demonstrating First Order Model: https://www.youtube.com/watch?v=mUfJOQKdtAk
Sometimes it almost works, and then it's just totally absurd. Long pauses, voices that don't sound compelling, total failures on female voices. It's great in theory but it showed me that there's a ton of work to be done with voice cloning.
Props to the author for using UMAP to seperates voices though
Also lol at the demand being so high that there are open issues of people offering to pay others to install this on their machine. Freelancing opportunities show up in the strangest of places...
The author says he's working on a more polished toolkit called Resemble.AI, but I've never tried it. https://www.resemble.ai/
There's certainly a market out there for just beautifying existing repos to making it easier for non-scholars to get going. Even having a Colab Notebook ready to click-and-start is quite powerful -- probably a big reason why First Order Model (source paper to the original story) got so much traction so quickly.
See what you get.
To be fair, I was trying this repo almost a 6 months ago so updates may have improved it.
anyway, on a 4930k at 4.5 ghz i am seeing reasonable performance(~30fps) but minimal cuda utilization(titan pascal). The biggest issue is that you need a well lit, stationary face for it to properly map features of the jpg you are using to substitute for your face. Also the jpg needs clearly identifiable features. Even then, the amount of facial expression is subdued (for example closing your eyes is not properly processed).
I seem to recall software about 10 years ago where you drew line segments of corresponding features on 2 images and the jpg was then mapped onto the video. It was more accurate and expressive than this is but did require time to set up.
It betrays that it's a fake, which I think makes it easier to be a fun joke.
“My internet speed is very low. I know the video is choppy”.
Or perhaps another neural network to distinguish the real from the fake?
So I think any detection algorithms would get into a never-ending arms race.
Probably something that would be built into phones and laptops first.
Of course, we know that messing with such a datastream would be easy for us, but for enterprise users (think liability) and news organizations it could be a real boon.
If a single shoddy manufacturer of cheap webcams somehow leaks a key that should have been on that webcam, then any software can easily encode a video stream with signatures asserting "this is what the camera saw, guaranteed".
Rick speaking to a crowd of hologram characters: Everyone who's first name starts with an "L" who isn't Hispanic, walk in a circle the same number of times as the square root of your age times ten! - https://www.imdb.com/title/tt3333830/
There are a few problems with these kinds of fakery. If you are simply touching up photos to make them more flattering, that takes a lot of work that the average person doesn't want to do. There's AI that can make you look sexier, but said AI is often inconsistent, and isn't that good at arbitrarily modifying features. People who might want to use photos of people who aren't them may have a moral hangup from using someone else's likeness, and that prevents them from going ahead with it.
Deepfake technology solves both of those problems. You can use it to change your eyes, your hair, the shape of your face, consistently at every angle, and possibly I real time. More people would feel comfortable using a likeness that isn't theirs because they don't feel like they are stealing, especially if they have no intention of meeting people in real life.
I'm not saying that all or even most people would use this technology in dating, but it would require only an appreciable portion of an audience to be dishonest to cause many people to throw in the towel.
These are just my crazy theories, so maybe I'm way off base. Soon we will have deepfaking that is easy to use and requires minimal training, and I believe they will be used often and eventually be normalized in society.
You absolutely are not. From my experience I'd say about 30%+ of the pictures that were posted to OkC 2+ years ago were doctored in some major way (I'm a straight man so this anecdotally only applies to women's pictures).
Using your "own" picture that's doctored is easier to justify for these folks vs. just using someone else's. Also, these people are still going on real-world dates, so if they look like an entirely different person they're taking on more risk vs. using weird Instagram filters.
> I'm not saying that all or even most people would use this technology in dating
I think most wouldn't employ this sort of deceit like you posit, as MOST of people I went on dates with were totally operating in good-faith.
BUT - there will be a huge chunk of people that are happy to do the mental gymnastics to justify posting a fake picture of themselves.. and those people are exhausting.
Protip: NEVER meet up at a bar initially because it's expensive when this happens (AND IT DOES!!!). ALWAYS do an informal coffee. It is easy to keep coffee to a 15-20 minute thing and not get on the hook for at-minimum 2x $15-20 drinks. If you chose a coffee place, do so an hour or so before lunch/dinner in walkable distance to a nice restaurant. This gives you both the opportunity to keep the date going if it's a positive experience =)
I can pay someone to attend meetings all over the world, by just impersonating me.
They already distribute CP, encourage people to believe in Russian conspiracies, troll, encourage violence and the list goes on. So fucked up that we think it is normal.
Edit: now done three images, all well lit face, no dramatic movements, and the output is just terrible.
For reference on Likeness Rights, one of Kevin Smith's Jay and Silent Bob movies has an entire sub-plot that revolves around the financial significance of likeness rights.
ModuleNotFoundError: No module named 'conda'
from conda.cli import main
----
EDIT: To get this to work, I had to remove a PYTHONHOME environment variable lingering from an old (but still valid) Python install (see https://github.com/conda/conda/pull/9239), and set CONDA_DLL_SEARCH_MODIFICATION_ENABLE=1 to avoid mkl_intel_thread.dll issues.
Leaving my comment here for anyone else who gets stumped.
- Miniconda installer suggests you don't add it to the path, so i didn't but install_windows.bat assumes you do
- There is something with windows permissions happening with the installer, where it can't access the network and can't run the conda activation process. I didn't spend any time figuring this out, i just execute dthe command from the (short) batch file one by one and it worked fine.
- i had to add 'fomm' to PYTHONPAT in order for it to find sync_batchnorm.
Link: https://9to5mac.com/2019/07/03/facetime-eye-contact-correcti...
https://venturebeat.com/2019/10/03/microsofts-ai-powered-eye...
Seems they've been working on it since 2002:
https://www.microsoft.com/en-us/research/project/eye-gaze-co...
In general I find that adding an oblique reference is more appreciated on HN than simply name-dropping a particular fictional work.
There is also a part where people had to start worrying about their looks on the screen, so that led to a cottage industry of virtual make-up, stand-ins to pose in front of the camera while you could stay away, etc. After everyone realizes that no one is actually in front of the camera anymore, people start going back to regular phone calls.
Perhaps there are other approaches for that particular case such as a button to 'raise your hand', however, many people have friendly relationships with their colleagues and simply enjoy seeing them when conversing.
> and plenty of Americans and Russian's willing to use whatever means necessary to ensure Trump gets re-elected
That is not what the Russian attacks are about.
In fact, the maker of the site seems to have fallen for the attacks himself.
The attacks are about dividing society. Radicalizing people towards the left and right. The "Trump haters" are just the same as the "Trump fanatics".
The author, by implying that the "bad guys" are just one site, is playing into the hands of the Russian trolls. Its exactly what the attacks are trying to achieve.
Non-Russian-aligned intelligence agencies were pretty unanimous in their assessment that Putin was trying to get Trump elected and that they troll both extremes of the political spectrum.
Once the pandemic is over, I'm never using Zoom again anyway.