I tried a Cycles render for a 3D icon. I ticked Denoise and it was a bit too smooth, foggy like. Very unnatural texture now. Not sure what is going wrong. Any tips would be great. Thanks.
You probably just need to up the render samples a little. The more samples, the better the denoise works.
Thanks. How many samples should I use? I tried 32. The default is 4096.
1028 works.
It also helps to maintain detail by rendering at higher resolution and scaling it down with sharpness afterward.
Thanks. I used 1080p for rendering.
and scaling it down with sharpness afterward.
Not sure how to do that
Animation or still?
You should adjust light paths to minimum where it matters how image look. More bounces, more noise. Also disable caustics if you don’t need them.
In still image if you need perfection, you can crank it up to some 4096 - 65536 disable denoising and adjust noise threshold where you can render it in overnight. Just estimate time that there is enough hours to render. Noise threshold can be something like 0.01 - 0.0001, keep it enough low that you don’t see noise.
If you are doing animation, 256 - 1024 are good values, keep denoising enabled and adjust noise threshold where you can render single frame enough fast like 10min per frame. Probably range is something like 0.1… 0.01
Thanks. It is a still image, a simple 3D icon.
I render out in 4k to preserve details, I may also use cryptomatte to refine how denoising is done. Then I load up in Krita/Photoshop/Whatever, apply unsharp mask with a “high” radius (iirc, didn’t fire it up to see my templates), then scale it down using Lancoz3, then apply unsharp mask again with lesser radius and effect. If you see noticeable halos from unsharp mask, you’re applying too much.
There is also high pass based sharpening and more of a three step approach to sharpening, but I don’t use these for rendering. The two main things that creates a softer image in Blender (Cycles, I don’t use Eevee) is the pixel filter size (you need it) and denoising (you need it). Meaning you will always end up with a softer than ideal image which you will have to sharpen up in post to get back. Similar things happen in photography although for different reasons; the raw image needs treatment.
I disagree about need of denoising. I prefer to use enough samples to get noiseless result. Image is sharp and light noise will add high frequencies that may be missing from image.
Sure I understand the process that render in double resolution with denoising, then Lancoz3 to resolution you are targeting and adjust unsharp mask to get desired sharpness. That may need to add some grain to image. Not only for add high frequencies but also avoid banding.
I’ve had real projects prior to denoising became a thing that still looked like garbage after 20000 samples. With about 24 hours per image to render out, that really hurts adequate time usage. So to me, 10-100 for previews or 400 for production renders (x4 for double res is 1600) is a fraction of the time needed to complete. So for my projects, denoising is absolutely essential.
I’ve experiencing that complex architectural renders at night, multiple light sources may cause weird sample distribution and denoiser thinks that in flat surface have some texture and that look terrible.
Before Blender 3, I limit samples to 16384. Without adaptive sampling, I just adjust samples lower if I didn’t have enough time to wait. At Blender 2.83-293, adaptive sampling helped a lot but that doesn’t work well so I often had to adjust noise threshold and limit samples to something like 4096-16384.
But in Blender 3, Cycles stop shooting rays more wisely when noise threshold is reached, so I use samples high as 65536 and still have faster render time than in Blender 2.9x. It makes things easy when I only need to adjust noise threshold.
Well, we obviously didn’t have adaptive sampling back then either. I need times I can work with, so low sample count (5 min) at 4k - while some details are still lost - is still far better than a 2k pumping for an hour as far as previews and discussion material are concerned. At these low sample counts and denoising, adaptive sampling does more harm than good for me so I choose not to use it.