I am talking about lens flares, vignetting and chromatic aberration, depth of field, bloom effects and film noise/grain. In photography or film they are considered as visual artifacts caused by imperfections and properties of film and lenses and mainly - the physics behind optics. Most of photo/cinemato-graphers do their best to avoid them by using better lenses and lens hoods to reduce vignetting, lens flares and chromatic aberration and use different sensitivity films for appropriate scenes to reduce image noisiness so in the end they get as clear image as possible that is usable for editing - tracking, compositing or whatever they do.. In the gaming industry it is pretty opposite - renderer outputs a perfectly clear image and artists do whatever they can to make it imperfect.
Few years ago these effects were done with some moving view-aligned textures, but now with clever shader tricks we can fairly accurate simulate them as an image post-process effects.
I have re-created most of these effects already and posted results in my blog. Links:
lens distortion
lens blur with bokeh v1
dof with bokeh v2.1
lens flare
And here are very nice results from other guys
Lens Flare by Max Planck Institut Informatik
Lens Flare by John Chapman
DoF with bokeh by Epic Games
DoF with bokeh by Matt Pettineo (MJP)
Film Grain..
..is the reason why I am writing this post here right now.
Film grain is a texture on the photographic film caused by an emulsion containing photon sensitive silver halide salt crystals. They are sort of pixels of the photographic film. Depending of the size of the crystals varies the resolution and sensitivity of the film. Bigger the particle (higher ISO number), higher the light sensitivity, but the image is less detailed. Unlike the digital image sensor where light sensing pixels are arranged in a regular grid, film crystals are jittered randomly over the film which gives an image more pleasing for the human eye. The possible reason for that might be the fact that the film grain does resemble the pattern how the photoreceptors are arranged in the retina of the human eye. Awesome, right?
photoreceptors in human eye (source http://www.sciencecodex.com):
an extreme case of color film grain (source Wikipedia.com):
Now in the digital era of cinematography most movies are shot digitally, that`s why often the film grain effect is artificially added on digital image. Many of the Blu-ray movies have a very distinct graininess, which actually gives a nice high-def cinematic feeling.
So exactly the same should apply to computer games.
Film Grain does seem to be the least difficult to simulate compared to other lens effects, but surprisingly I have not yet seen any convincing real-time implementation that does resemble, for example, the nice granularity look of 35mm film. Well the best examples of film grain effect I found used in games are by Crytek and Valve software - Crysis Warhead/Crysis2 and L4D series to be specific. Unfortunately the effect can hardly be called "film grain", it is rather an overlay noise that makes the appearance of slight jitter. What they did right is the elusiveness of the effect. Increase the grain amount a little bit more and it gets annoying. For many game titles film grain effect bothered me so much I disabled it completely (if there was an option at all). Mass Effect was one of them, in fact in darker scenes it looked nice, but in highlights made the image look dull and dirty. It did a good job covering up some less detailed textures and models though.
The possible reason why film grain does not seem to have advanced over the years like other effects might be that the grain filter is so subtle the developers do not bring much attention to it. And why waste time, money and precious milliseconds of computing time to create an effect most gamers would never notice. Well, here exactly lies the problem - overdone effects. Most notorious are Bloom, SSAO, DoF and yeah.. Film Grain and they are certainly very easy to overdo.
Personally I love subtle details. For me the best effects are the ones that enhances the visual quality while not bothering me with its presence during gameplay.
And there is another point - high ISO film photograph can get very grainy but the image looks nice and natural, while a bit overdone film grain effect in game ruins any viewing pleasure.. So the problem lies in the technique which generates the grain pattern and mixes it with the scene image.
Approaches
The common methods of simulating it in games are either using a real grain texture or computing a noise pattern procedurally in runtime and then mixing them with the rendered image.
Pre-computed texture approach might make the best results as you can use actual grain texture taken on real film (filming against a grey background) and then overlaying this image to game scene. the downside is that it needs to be different in every frame. You can offset it a little and tile it, but a trained eye will always notice the tiling and this can make the effect very annoying. The approach does work best for still shots.
Procedural approach will ensure the grain will never tile, but the result is actually a noise and does not resemble the granularity of film grain at all. It does look more like a digital sensor noise. One might actually misinterpret it for an interference of the video signal to the monitor.
My take on Film Grain effect
All the time I have been using the procedural noise approach which is fast and gives nice results if the noise amount is very little, but as I mentioned it looks more like an interference in video signal.
After doing some research in this matter I collected bunch of reference images taken with film camera. I extracted the noise pattern from them and compared it to the noise shader I have been using. As you see it is lacking the graininess of the real thing.
I opened the image of my noise shader in Photoshop and tried to replicate the real grain texture. Blurring a little and adding 3 passes of sharpening did actually the trick. Unfortunately doing that for full-screen texture in shader would significantly slow the process and make it unusable in real-time.
So now I had to find a procedural or semi-procedural approach to simulate the same pattern in the shader without having to blur or rendering it in multiple passes.
Few years ago I did a vertex displacement mapping experiments with a procedural Simplex noise algorithm which created a nice grainy pattern. I resurrected the shader and applied it to a grey color and here is the result:
That is a really pleasant uniform grain. It is lacking a bit of the randomness but otherwise I call this a success!
Comparison shots of noise and grain applied to a gradient.
noise:
simplex grain:
Results
There is yet long way to go for more accurate results. It could be extended to include the real science behind it like varying graininess based on camera`s relative aperture. But I am only an artist so I go by whatever looks the best :)
This whole thing is a work in process, but I feel that even at this stage it looks already better than many other real-time film grain shaders. As you see the grain is not making the image look dull. It is less visible at bright areas of the image, but more noticeable in darker shades. An option to change the grain size is yet to be done.
Perlin noise grain applied in CryEngine (big images):
screenshot 1
screenshot 2
screenshot 3
100% crops from the shots above:
Continuing
today during further grain shader development I tried to replicate the grain in the image with the rally car. I had to de-noise the original image with resulted in some loss of detail. Then I applied the new grain shader on it. Here is the result:
original:
de-noised and artificially added film grain shader:
real film grain compared to new film grain shader:
There are some significant changes from the original shader:
- user variable grain size
- added varying coordinate orientation for noise pass to eliminate any directional artifacts
- option to reduce grain based on luminance value
- added more tweakables for color noise
downloads
v1.0 (old one):
HERE you can find the GLSL shader file.
v1.1:
HERE
References:
- noise algorithm I copied from HERE by toneburst, but original implementation comes from Stefan Gustavson found HERE