110 comments

[ 4.2 ms ] story [ 176 ms ] thread
Sure graphics are looking amazing...but I still think we are in the uncanny valley. Especially when it comes to faces.
I agree. Rogue one was a great example. I found it very disturbing. Both Tarkin's and Leia's face
Right and even when it was just about perfect it's unsettling because you get taken out of the moment wondering, "Am I looking at a real person or CGI"?
It isn't so much the faces look either, but how the faces move. Movement is often too slow, too long, too regular...and ill-timed.
Yes! Motion is key. It's not enough to be photorealistic -- you need to be "videorealistic" as well.
Is that something deep learning could tackle? (asking as a complete non-expert in DL)

It feels like a similar problem to realistic speech synthesis, and there have been some really impressive TTS demos recently.

Athough on that note, how come Google can produce these amazing speech synthesis demos, but Google Maps still sounds robotic when giving directions? Is there some big problem around generalizing or productionizing the technology? Is there some hidden trickery behind the demos?

(comment deleted)
I went in to the movie not knowing that Tarkin was CGI - I assumed he was a stand-in and makeup - and came away none the wiser. After I was told he was CG, it was obvious and weird...

Wonder how much of the uncanny valley stuff left is in our heads?

I know a few people had that reaction, but I went in blind too and the CGI was painfully obvious. They had close-ups of Tarkin which looked really terrible.

CGI is superb at creating indistinguishable still images, but animation is still riddled with problems.

animation was too "perfect" in my eyes.
he looked like a wax figure from madame tussauds that was animate
At least wax figurines get lit realistically, the one place most CGI seems to fail.
It is easy to A-B test. Maybe someone has done that already?
What would you ask the test subjects though? Seems to me as soon as you say anything regarding realism a lot percentage would notice things they wouldn't have before. I'm sure a clever experiment designer could come up with something though.
Show a mix of video clips. Swipe left for real. Swipe right for CGI. After a few clips, show how often you were right. Hit the HN frontpage.
Had the opposite reaction. Went in not knowing, then when I saw Tarkin and Leia I was kind of thinking, wtf? They digitally rendered their faces? When I talked to friends afterwards they didn't notice it. This must be pushing some sort of mental limit of our brains - not everyone recognizes faces/cg the same as others.
I was blown away by Tarkin, I kept thinking how in the heck did they bring him back exactly as he was? Leia was another matter, something was not quite right with her.
It's funny, while both felt pretty plasticky; Leia looked much better to my eyes than Tarkin. The brevity and minimal movement helped, I think.
Out of curiosity, did you see it in 3D? Most of my friends that saw it in 3D were complaining about how bad the renderd Leia and Tarkin were, but very few of my friends who saw it in 2D had any complaints about it.
I almost find this hard to believe, unless you seen it somewhere with poor resolution or lighting. Sure they did a good job, but they were so noticeable it broke the spell of the movie for me. They are the best proof that we are still in the uncanny valley.
I saw the film on one of those tiny in-flight screens (maybe 9 inches diagonal?) and even there it was painfully obvious.
Me too, but a friend of mine hadn't noticed that certain characters were CGI, so it must work for some of us.
Agreed. And I always associated the uncanny valley with representations of humans, not terrain, robots, etc.
I urge you to watch Blade Runner 2049. Minor spoiler: there is a scene involving Rachael. They brought in the original actress and de-aged her face digitally to look like she did in her 20s and the effect was... mind-bogglingly close to perfect. Easily the best facial CG I've seen so far.
> They brought in the original actress and de-aged her face digitally

No, Sean Young (the original Rachael) was brought in to consult, but the CGI face was superimposed on a stand-in, Loren Peta [1].

I saw the movie on Monday and wasn't impressed with the quality there; worse than Rogue One Tarkin and on a par with Rogue One Leia. As many others have opined in this thread, the technology's fine for stills but really, really not there for animation.

YMMV, obviously. I have a friend who made it about a third of the way through Zemeckis' 2007 Beowulf before asking "hey, why did it all just go CGI?", and that movie wasn't remotely aiming for photorealism.

[1] https://en.wikipedia.org/wiki/Blade_Runner_2049#Cast

It may be a while. We have a large amount of evolutionary neural instinct when it comes to human facial recognition. Realize that camouflage in nature is very common, and as a result our ability to distinguish a real human face from a predator imposter has our entire evolution and survival baked in.
I've discovered something akin to the uncanny valley - 4K video. It looks so sharp and clear it feels artificial. Not sure how long it will take to get used to.
Agree. 4K video + high FPS makes real life look 3D graphics.
I think we just need more "real life" events in 4k video and high fps. Right now the only thing associated with that is 3d graphics in video games
Some sports are broadcast in high FPS IIRC, but not in 4K (which would be great).
Actually, Sky broadcasts all Bundesliga and world cup games in 4K at 60fps.
60FPS movies do this to me - everything tends to look like a stage play.
Actual movies in 60fps, or movies in ~24fps on a TV that is using interpolation to produce 60fps output? The interpolated/faked experience is often known as Soap Opera Effect [1]. I really dislike interpolated video, but really do enjoy 'real' 60fps video.

[1] https://en.wikipedia.org/wiki/Motion_interpolation#Soap_oper...

I mean 60fps movies - like when I went to see The Hobbit in theatres. I like it too, don't get me wrong, but it takes some getting used to and gives a different feeling. Like a painting versus a photograph.

But yes motion interpolation is awful.

The 48fps version of the Hobbit was jarring for a number of reasons. I'd be interested in seeing a movie with more real actors and sets at a high frame rate, just to have a better point of comparison.
I've noticed it too with a friend's TV. Some shows / movies become unwatchable because it's so distracting. -- I think part of it is that many of the artificialities of the screen become too obvious: the lighting, the makeup, the fact that many elements on a set a plastic / CG, oddities in perspective, etc... all of these elements are "fuzzed up" by a lower resolution and FPS, forcing your brain to fill in the gaps -- ironically more convincingly.

Sometimes in art, less really is more. Many modern Sci-Fi flicks fail to achieve that feeling of wonder precisely because they are shooting for too much realism in their special effects, I think. The problem is not CGI itself, but the technology does seem to make it too tempting for the filmmakers to fill in blanks that are perhaps better left up to the viewers' imagination.

I think you've hit on it. I feel like I'm on a movie set watching the actors when watching 4K, as opposed to getting into the story like normal (can't remember the term for that...)
No. The landscape and nature in that video didn't even look real--never mind people.
The pulldown may have helped in that impression, as pretty much every camera motion was jerky ...
For those interested and unfamiliar, “pulldown” here refers to converting frame rates between different video standards by doubling up or blending frames at regularly spaced intervals, which can introduce a visual artifact called judder which is apparent in those clips.

https://en.m.wikipedia.org/wiki/Three-two_pull_down

For some reason the video was uploaded at 25fps which almost never plays right on computers, and frame-by-frame playback shows the skipping is baked into the video. I got bored counting frames to figure out what the source framerate might have been, if there was a consistent source rate.
In my experience the moment people know that something is CGI they get into some kind of "extra scrutinize" mode, which moves the goal. A study showing people videos and asking them to point out CGI vs real would be far more interesting to me.
Nah. The landscape and plants were too shiny and high contrast. Everything looked plastic.

There were some convincing ones in the video though, such as Avengers and the truck demo.

Once you get realistic "rendering" you also need to stick to familiar cinematic styles. There's no motion blur or depth-of-field like there is when shooting with a film camera, so even if the landscape were simulated perfectly, it would still look weird.
Yeah,no.

We're not even close to passing the uncanny valley in movies or games

We might ve passed it in pictures, but that's kinda old news at this point.

05:56> as computers graphics get better, we believe all images less

This nugget in the video struck me most. I remember when I began trusting images less once I understood the capabilities modern photo software provided. We've now achieved this with most physical forms in motion picture, short of humans. Which is not to say we won't soon. Once that threshold is crossed, then what's next? Eventually virtual reality, augmented reality, and finally reality reality. We'll trust everything we see (and eventually hear/taste/smell/feel) far less than we do now.

Indeed! I remember this being discussed as early as 1993, in the film "Rising Sun" http://www.imdb.com/title/tt0107969/

Another good example is https://www.cbsnews.com/news/bp-and-the-gulf-oil-spill-misad... ...which I remember being a wake-up call to those who didn't know it had already been going on for a couple decades.

I truly wonder how much of what is presented in the MSM these days is actually not real and manages to slip by...

On another note, did Michael Crichton ever apologize for that book? It was alarmist xenophobia targeting the Japanese, verging on racism, and one of the reasons I stopped reading him. (Watching history unfold in a way totally unlike his dark, paranoid vision has been a great pleasure.)
(comment deleted)
(comment deleted)
That was a quarter of a century ago. Maybe you should let go?
> I truly wonder how much of what is presented in the MSM these days is actually not real and manages to slip by...

Quite a lot presumably. Ever since the idea of a media started, there have been people trying to hoax it for whatever reason or not:

http://hoaxes.org/weblog/categories/category/Journalism

If a story is mainstream enough (or remotely political), hoaxes will probably get identified at some point or another. If it's in a specialist subject area? Forget it. Who knows how much questionable stuff slips through when neither the journalist nor their audience know what's plausible and what isn't?

It still requires large enough means that you should know from the context whether a video is likely to have been manipulated or not (based on whether it would be worth it). Unlike a photo where pretty much anyone who can use a computer can photoshop it.
50 years ago faking a picture was hard. It won't be long for video to be as easy as still is today.
Of course, no one now can know for sure what the consequences will be.... but I'm intrigued by the possibility that this might be a good thing. Could this help make people more savvy about evidence, and less focused on the concrete and superficial? Like I said, I -- we -- don't know.

Perhaps the effect of spam and phishing on the general population provides some insight into this. Have they increased the degree of healthy scepticism and delayed judgement in the population overall? I don't think there's any knee-jerk obvious answer to that, but perhaps there has been some research into that.

People are situationally sceptical. We tend to be sceptical of things we already don't believe, but not of things we do. We can expect people now and in the future to question an images authenticity if it questions something they want to believe.
And more dangerously, if they already agree with the majority of the image's content, and there is an introduction of fictional content: that false information gets baked into the observer's mind as belonging to their "trusted information". And that is a major perception problem with humans.
No, we won't trust things less. It's wired into us to believe what we see. Rather we'll get fooled more often.
eh, i think technology can erase some of the need for human trust. Checksums, public/private keys, blockchain, etc. should all be able to handle verification of authenticity.

As a result, i'm sure we will soon have CAs for more than just https.

That mentality right there is dangerous. Yes, it is possible, but in practical execution it will be insecure. I see an active mentality to delegate responsibilities to software that should not be. Case in point: using facial recognition for authentication of services with access to one's finances. If anything, we need to realize the reliance on human trust increases with technology.
I had at one time thought something similar, but in review, I think the weakest point is that such things become background noise to the initial reaction and join "References" as the newest proof of validity that never get properly vetted, as we often see with think-pieces that want to take on an air of authority. It's very simple to convince people right now with a reference to an article every other word, regardless of how reliable or even relevant each of the references actually is. The effect though is that the piece attains a state of legitimacy because "boy look at all those references", and disproving such a piece is far more difficult than writing one, with the research to break it down taking far longer, and with attention spans being more interested in whatever the piece is trying to push.

I fear that when we try to use thinks like Checksums, Pub/Private Keys, Blockchain, they're not going to be applied in a reasonable measure, nor will they be explained to audiences in an honest way, but instead simply be tacked on as a badge of legitimacy for the content, not as simply a validation of the source/transaction history.

>we often see with think-pieces that want to take on an air of authority. It's very simple to convince people right now with a reference to an article every other word, regardless of how reliable or even relevant each of the references actually is. The effect though is that the piece attains a state of legitimacy

[case in point this Youtube Video that's been making the rounds.](https://www.youtube.com/watch?v=KqFNbCcyFkk)

(comment deleted)
This is how I know the moon landings were fake.

/jk

Kant and the rest of non-realist philosophers are going to have a field day.
Overall a really solid video about the modern state of computer graphics, with a comprehensive tour of the cutting edge realism of light and fluid, the current tricks and limitations used in the modern cutting edge, and then an exploration of the artists and formats that strive to be on the cutting edge of computer graphics while also not aiming for realism.
Related video about why people think CGI sucks or is still in the uncanny valley: https://www.youtube.com/watch?v=bL6hp8BKB24

Basically we consider CGI bad because we only see bad CGI. If it's used properly and the director/creator understands its limitations we either don't notice it or use our suspension of disbelief and don't care.

We have gotten very good at nearly perfectly rendering images that accurately capture the behavior of light and materials, but I feel like to produce truly convincing CGI we need to improve the imperfections that we have come to expect from more traditional media.

As other commenters pointed out here, the opening nature valley scene was unconvincing. I think this is partly due to the author not having modeled the limitations that affect physically filming a similar scene.

- a helicopter or airplane could not plausibly move along the track shown in the opening shot

- camera rigs have imperfect tracking control, sometimes biased along certain axes

- high res 4k sensors have physical limits and defects. the scene is rendered as if captured by a totally perfect sensor

Simulating the motion of a helicopter that doesn't even appear in the shot seems well beyond the effort that most digital artists are willing to go to, but I think these sort of things are key to producing something truly convincing.

That's the point. The opening scene is the uncanny valley and supposed to be not quite realistic
That was my read as well. He even intentionally dropped the frame rate.
In addition to that list, there are the physical issues of film and optical devices. CGI can just work around them, but with film and lenses, physics still gets yah. For reference (http://www.quadibloc.com/science/opt0505.htm) the first order optical issues with all systems are:

> Spherical Aberration: this is the aberration affecting rays from a point on the optical axis; because rays from this point going out in different directions pass through different parts of the lens, then, if the lens is spherical, or otherwise not the exact shape needed to bring them all to a focus, then these rays will not all be focused at the same point on the other side of the lens. Essentially, lenses are spherical, they need to be parabolic but that is very expensive.

> Coma: this aberration affects rays from points off the optical axis. If spherical aberration is eliminated, different parts of the lens bring rays from the axis to the same focus. But the place where the image of an off-axis point is formed may still change when different parts of the lens are considered. Essentially, your lens may not be flat compared to the film.

> Astigmatism: this is another aberration affecting rays from a point off the optical axis. These rays, as they head through the lens to the point in the image where they will be focused, pass through a lens that is, from their perspective, tilted. Even if neither spherical aberration nor coma prevents them from coming to a sharp focus, if we consider the rays of light that are in the plane of the tilt, and the rays of light that are in the plane perpendicular to that, these rays pass through a part of the lens with a different profile. So they may not be focused at the same distance from the lens, even if they do come to a focus in each case. Essentially, lenses aren't perfectly spherical and may have a more cylindrical portion.

> Curvature of Field: even when light from every point in the object is brought to a sharp focus, the points at which they are brought into focus might lie on a curved surface instead of a flat plane. Essentially, lenses are round, and film is flat.

> Chromatic aberration: Essentially, red light focuses at a different distance than blue. This is the reason for rainbows.

I have to say, the most recent Planet of the Apes was seriously impressive. It's the closest I've ever been to admitting that something looks indistinguishable from reality.

However, they are animals and arguably I don't spend nearly as much time looking at apes as I do at human beings. Still renders of humans may look very close but anything in motion still looks completely fake.

Totally agree though the concept of the uncanny valley is based on computer generated or robotic representations of humans specifically -- not other forms of life or inanimate objects. Meaning, not just realism in general.
The one that did it for me was Disney's remake of The Jungle Book.
Some environments pass the uncanny valley in video, namely vehicles, architecture, nature - but complex things like smoke from a large fire are still ridiculously non-realistic.
Except half of the modern popular video content discussed in this clip represents evermore uncanny depths of said surreal valley. Overly glossy blockbuster movie FX sequences, subversive glitch gif style viral media, vaporwave, and so on.

So, maybe the title is something like:

  Warm Hugs For the Uncanny Valley And Other CGI Oddities
There's still a huge tell in contemporary CGI, which is pretty obvious when compared side by side to older, practical effects: smoke and light scattering. Given the amount of destructive action we typically get on screen, the picture should quickly devolve into a murky, misty mess that takes ages to settle down. But of course it doesn't, because the audience needs to be able to see the spectacle.

If the scene is composited digitally out of differently rendered layers, the light won't scatter and occlude correctly through the mist. There's a reason why e.g. the original Blade Runner still looks so timelessly atmospheric compared to your run-of-the-mill dystopian cyberpunk city of today.

In that sense, we've been in the age of post-real CGI for much longer than the quest for photorealism suggests. We're not actually seeking photorealism, we're seeking the hyperstimulus version of it that obeys cartoon physics more than real physics. Simulating fire, smoke and fluids is one thing, but simulating realistic combustion, fogging and drainage is quite another.

Then again, you could make exactly the same observation about cars that explode upon flipping over, or the jumping bus scene in Speed. In real life it made a tiny hop and then smashed its front wheels. Magicians don't need CGI either.

(comment deleted)
I don't know about that. CGI effects are everywhere in modern film and TV, but I rarely hear people talking about them except in the case of action scenes or "obviously CGI" (fantasy/fictional) creatures. Take a look at the VFX highlights reel from season 1 of Game of Thrones [1]. Even 8 years ago they were using CGI throughout the show's environments.

[1]: https://www.youtube.com/watch?v=YpjdpDx8_l8

The bus scene in Speed used a 50-foot practical jump from a ramp. CGI was used to remove the road that beneath the jump, not to lengthen the jump itself. I guess you could call that a “tiny hop” if you're measuring vertical distance from ground; not if you're measuring distance from road, or length.

Sources: I remember this from Cinefex, which is not online; Wikipedia repeats the description https://en.wikipedia.org/wiki/Speed_(1994_film)#Filming

also, that jump went way further than expected IIRC
I remember it was classically famous for how it was much higher than expected.

So much so I thought it looked very fake in the movie, but the special effects people kinda shrugged and emperor new clothes it.

Fake physics is the most distracting thing about character CGI in particular. Any time a character is leaping, landing, swinging on a rope, etc. - it all just looks incredibly wrong and jarring.
For me the most painful and infuriating example of this was The Hobbit trilogy.
When they are inside barrels on rapids, Jesus... Anyway the shots are so short that you cannot understand anything about what is happening. You just have to wait patiently until the end of the cinematic scene so that the game^Wmovie starts again.
> We're not actually seeking photorealism, we're seeking the hyperstimulus version of it that obeys cartoon physics more than real physics.

On a much higher level, this is what art has done since the dawn of time.

Presenting a version of reality that's subtly altered to make a point and convey a message.

Now we can finally do it easily at the low purely graphic level for motion pictures.

Lasers don't make sound in space.

The uncanny valley isn't about making things realistic, it is about making things feel natural. Things flying by make sounds, so it's natural to assume lasers in space zipping by to make sounds. Same with all sorts of visuals rendering situations we are highly unfamiliar with.

Dear videographers, when you pick a "photorealistic" opening scene, make sure the first major object you see (the palm tree) doesn't look like its made of plastic. It also helps if you don't try and mix close up objects on far away objects (the transition from grass to what appears to be a white-sand delta or a glacier).

Also, if it is a glacier, snow and running water reflect light very differently.

Just three points in the first 10 seconds which puts "the battle for photo-real CGI has been won" on awful shaky ground.

It seems, to me, like photorealism is currently possible for static images, but the moment you add motion, it all goes to hell.

It's a shame you responded only to the title, because the content of the video is much better.
Time constraints, a rather visceral disagreement with the premise, and bouncing ogres on chairs (so avant-garde!) have postponed completion of the video for me.
So, watched the whole video finally (such a salty response to having to postpone watching more than the intro). I still think the premise is fundamentally flawed. The author tries to get around this by initially defining the uncanny valley as "how we got to today's CGI", and later by calling it "bad CGI" - but by either definition, there was no "good" cgi being shown; no truly post-uncanny valley content.

Admiral Tarken? Princess Leia? Faked Obama? All examples of "beyond", and all really, well, uncanny to watch.

I can't even honestly agree that the commercialized software has started to plateau. The software to create CGI - especially the software being used by the big studios - is evolving constantly. Even Blender changes on a monthly (or more frequent) basis.

For example, in How to Train Your Dragon, one person's time was solely devoted to improving the water splash dynamics. In Frozen, they devoted decades of processing time to rendering one 15 second scene with the ice castle. In Mad Max, the CGI was praised because it wasn't the primary method of creating the effects, because it didn't get in the way of the practical effects.

Can we eventually get there? Yeah, I think so. But we're not there yet. The software isn't finished; the hardware isn't fast enough. Perhaps it will come about in tandem with general AI.

Totally agreed with this. Faces have further to go. We're not out of the uncanny valley yet.

We're also unable to create emergent materials right now, because we don't right now (and may never) have the horsepower to do molecular-level simulation to find out how light interacts with the surface. For similar reasons, breaking something open like an egg, a log, or slicing open flesh is never going to yield truly emergent results. To some degree, they'll always have to be artistically approximated on a case-by-case basis. Those approximations may have increasingly sophisticated interactions as time goes on, but I am reasonably confident they'll never truly simulate things at the molecular level.

> Dear videographers, when you pick a "photorealistic" opening scene

I kind of think that it was supposed to be an "uncanny valley" opening scene. The rest of the video is quite good; it's hard to imagine that they'd flub the opening so badly, unless it was on purpose.

In my opinion, none of the CGI they showed was really beyond the uncanny valley; most scenes which which weren't still photos or landscapes made predominantly of rock were still quite fake looking. Especially the "beyond" examples of Tarkin and Obama.
Awesome video! Thanks for creating it!
Unfortunately, a lot of this video essay consists of ideological political speculation from a very narrow perspective. His history of the industry was very handwavey as well, the requisite prelude before political grandstanding.
> prelude before political grandstanding

What are you talking about, and did you watch the whole thing? It seemed to mostly be about the potential future of CGI as art and its place in contemporary/future culture.

We are still not there. Press pause button on a CGI heavy movie like "2012". The effects work if they are in motion. But looking at them in "pause mode" on 1080p it looks like shit hits the fan. The same for Transformer movies. 90% is real world material with green screen background CGI effects. There are still way too many physical effects going on. And many directors went back to do more physical stunts, because the end CGI-only result was easily spotted and universally hated by the audience.

Jurassic Park 1993 had only 18 minutes of dinosaur screen footage of - 95% of them were miniatures, costumes and hydraulic mechanical robots. Only a very short T-Rex running sequence was rendered with CGI with the predecessor of OpenGL rendered on Silicon Graphics hardware. The good thing is, even today the movie (shot in analog) can be viewed in 2k or even 4k. They will have to re-render "2012" and the Transformer movies for 4k and 8k in near future.

Also the real time rendering is behind, video games still can't render 2005 era movie quality in real time, let alone 2017 era quality.

So a bit pre-mature to call it "goodbye". The progress is great, but there is a road ahead for sure.

Only at a distance. Failure to extend research into vector textures over bit textures in the 80s now means 3D objects look blurred when close up. A big deal in VR and AR applications.
I have yet to see a movie that tries to mimic what our eyes really see (with all the weird diffraction effect we have with our pupil and eyebrows, and some particles moving in the liquid in front of our eyes)
When I started watching the video, I thought "yep, this clip is a perfect example of uncanny valley" and kept waiting for them to switch to a clip that was actually photorealistic.

Nope, turns out they seem to think the splash video is realistic. The whole video has tons of obvious flaws and overall feels wrong.

I don't know if I have superman vision or what here...

I was really blown away by one of the video clips this video uses [0]. It's a tech demo for a "zones" effect in VR. Really interesting effect that I could see being a lot of fun in a Sci-Fi game (wormhole travel).

[0] https://www.youtube.com/watch?v=6D-8kkH1KPE

The author doesn't seem aware that "The Uncanny Valley" refers to the inability to make simulated humans appear natural, not making CGI in general look natural. Very annoying seeing someone use a term with a very specific meaning and not understanding it
I agree. The author's claim on the video that it's "common" now to use the term more widely is strange to me. I've never heard of it used in the more wide sense since the term became widespread, and I've worked in computer science research since the '90s. No citation given ...