2

I've read this question and its answers, but they didn't solve my problem, so I am re-asking this question under the context of my setup.

I am rendering some planes under HDRI lighting. These planes' material is a mix of transparent BSDF and diffuse BSDF (with 0 roughness!).

To avoid the grainy renderings (or fireflies, as some other people may call them), I tried the methods mentioned on the internet. Specifically,

  • Sample number is et to 128
  • Light paths are set to "full global illumination," where maximum numbers of bounces are all 128 (but this also means caustics are on. Maybe I should try turning them off. Thought they contribute to photorealism, the graininess I current have is unbearable...)
  • Filter glossy is set to 0.5
  • world.cycles.sample_as_light set to True, as I read from a article by Andrew Price

Despite all these efforts, the results still look pretty bad. Worse still, I have those weird white dots that seem like missing pixels...

enter image description here

Is this completely hopeless? What else can I attempt?

Community
  • 1
Sibbs Gambling
  • 719
  • 2
  • 15
  • 30

2 Answers2

6

Sample number is set to 128

Unless you have the "square samples" option enabled (which turns it into 128^2 samples, aka 16384 samples), 128 is not high at all. Set it higher.

Light paths are set to "full global illumination," where maximum numbers of bounces are all 128 (but this also means caustics are on. Maybe I should try turning them off. Thought they contribute to photorealism, the graininess I current have is unbearable...)

I'd advise never using the Full GI setting unless you specifically want to wait around for awhile to get every last lighting detail. The benefit of extra diffuse bounces drops of FAST after 2-4. Glossy and transmission often need more to avoid dark reflections/refractions, but 128 is overkill. Try 4-6 for glossy and 10-16 for transmission.

Filter glossy is set to 0.5

Try 5.0. 0.5 is barely enough to even do anything.

world.cycles.sample_as_light set to True, as I read from a article by Andrew Price

Do you know what that setting actually does? If not, give this a read: What is multiple importance sample option While that option is usually the right call for scenes with HDRIs, you should understand WHY you are enabling options, not just flipping them on because the tutorial said to.

Generally, here's my steps for cutting down render time and noise, in order of increasing desperation:

  • Turn on filter glossy
  • Disable refractive caustics
  • Turn off diffuse visiblity on mesh lights, add a lamp object to cover the missing lamp.
  • Lower diffuse bounces
  • Disable reflective caustics (they're generally softer than refraction caustics, so less noise and more contribution)
  • Turn on clamp indirect. Remember that it clamps sample values and not final pixel colors. Mind what white point you're using if you have enabled exposure options, post-processing, the filmic-blender LUT kit, etc.
  • Turn on clamp direct (absolute last resort, avoid this setting if you can).
Community
  • 1
JtheNinja
  • 9,122
  • 1
  • 31
  • 49
  • Can you go into more detail about "Turn off diffuse visiblity on mesh lights, add a lamp object to cover the missing lamp."? I never heard about this trick. How do you do it? – WannaKnow Dec 31 '16 at 12:01
  • Required reading: http://blender.stackexchange.com/a/58570/1853 read this answer and the comments. –  Dec 31 '16 at 16:28
  • I'm not sure whether cegaton's answer comment was meant as an answer to my question, but in the linked answer and comments I see nothing about "Turn off diffuse visiblity on mesh lights, add a lamp object to cover the missing lamp.". My best guess is to uncheck diffuse ray visibility on a mesh with emissive material, and add a lamp which has only diffuse ray visibility. – WannaKnow Dec 31 '16 at 17:27
  • @WannaKnow He basically meant that mesh lights being area light behave in a way that introduces noise to the scene. So by turning off diffuse visibility exclude it from lighting emission calculation. But which means there is no more light emission but the mesh will still look bright white. Hence to complete the illusion, you will need to create a light object to "fake" those missing emission you just turned off. – hawkenfox Dec 31 '16 at 17:41
  • The comment was to make others aware of the information on the link, more detailed information on the same subject –  Dec 31 '16 at 17:42
5

In my personal experienced with cycles ...

Because of the law of conservation of energy in the real world, light hitting a surface should never reflect more energy than what was initially available from the source. Some energy should have been diffused in other direction or lost during the interactions with surrounding environment and/or objects.

Fireflies or bright white speckles is a sign that law of conservation of energy has not been accurately observed within the engine's calculation. Hence the white speckles are light sources reflecting a greater amount of energy when compared to the initial input. So imagine 1 Lux of light is now 2 Lux after reflecting off the surface of the object... this is definitely a problem and noise can be introduced easily ... imagine the light source gets to bounce around and it becomes stronger every bounce and create inaccurate solutions.

The really effective solution is to hinder this scenario from happening.

  • Use Clamp indirect. Make sure that the indirect light can only get up to a certain intensity helps. Starts with 4.0 and Lower the number until the noises disappear. This would usually solve most of the noise issues.

  • Using light portals to direct light sources into hard to reach area. Engine's are not smart enough to know where to aim all the samples towards a certain critical area of your model. If your model has a lot of surfaces overlapping and light exposure are not direct, it would often help to add some light portals to focus on those areas where samples may have missed them and noise were inevitably introduced. This is often true for an interior scene with a small window opening.

hawkenfox
  • 4,266
  • 1
  • 15
  • 28