Blender 4.x Cycles Photorealism Improvements...still not quite there for me

As much as I kinda dread reopening this whole topic of photorealism in Cycles, I think it can still lead to a constructive conversation (and perhaps I’ll learn a thing or two).

I had high hopes that Blender 4 would bring new revised shaders which are more physically accurate to Cycles and hence address the “cartoony/CG” issues I’ve been having with it. With the addition of AgX I thought that we’d finally start getting into the type of photorealism that I’m seeing on other renderers (most notably Octane, V-Ray and even Renderman).

Now that Blender 4 has been out for a few months, and I’m seeing some of the results that people whom I consider “masters” of photorealism are achieving, I’m still left feeling like we’re not really there yet.

Most recently, Paul Chadeisson who in the past used Octane exclusively, switched to Cycles for Solstice-5 (which is still an absolute stunner of a piece), but IMHO is visibly less realistic (however you want to define that for high-concept sci/fi) than his breathtaking Migrants. The sense of scale and detail IMHO isn’t as “concrete and real” feeling in Solstice-5 as it was in Migrants.

After really pouring many hours studying a ton of images, what I have narrowed it down to is the light falloff difference between Cycles and other render engines. There is a scattering of shadows and light at the edges of objects that implies detail that Cycles seems to lack. It’s very subtle and nuanced, but it’s present and it affects the perception of distance, scale, and detail (IMHO). Cycles still feels digital and overly sharp in those regions of the luminance and spectrum, and I am convinced this is what I’m seeing that it’s not quite doing it for me.

I have no idea if the Spectral branch of Cycles would close that gap, or if the issue is more deeply endemic to just the way Cycles (or perhaps the Cycles cameras) render the path tracing results.

Anyway, just wanted to share my thoughts. Not a criticism of Blender or Cycles necessarily as I know many here just love the way it looks. Just an observation that surely will find a lot of opposition but perhaps also some agreement.

P.S.
Strangely, I have found EEVEE Next to yield more natural light falloff results than Cycles, and had really high hopes that it would be ready for the upcoming release. Unfortunately it looks like more time is needed.

9 Likes

Having used BlenderOctane for a while now and having converted several of my Cycles scenes to Octane for comparison, I can confirm that the Octane results look more realistic to me. It’s hard to put a finger on what exactly it is, but they do. The Cycles renderings look fine too, but they always had a bit of a technical and clean look to them. So for those cases, Cycles is a great option. But if you want to go more into the direction of a film look, I much prefer Octane. And with my scenes now sometimes taking over a minute to compile in Cycles and not really being able to cancel renderings without having to wait around 30 seconds for the denoise to finish a cancelled rendering (the devs think that this is OK), Octane offers a better experience. No matter what - it’s interesting to see where developments will go.

2 Likes

Can you verify if these subtle differences are also due to differences in how samples and edges are filtered?

Cycles does not soften samples like Luxrender and other engines did back in the day, and the Blackwell-Harris filter is known to be quite sharp compared to Gaussian and other algorithms. There is also the possibility that denoising is playing a role in the ‘technical look’, use the version in the compositor and sharpen the denoising normal pass before feeding it in to get maximum detail (though you will then need more samples).

3 Likes

wish they would do one of those shorts in both cycles and another engine to get a real 1:1 comparison

4 Likes

I don’t think this has necessarily to do with sharpness per say. There’s more to it than that. I did some comparisons between Cycles and Octane and the pixel-peeping level, and Cycles seems to produce marginally sharper images. But that’s not what causes the differences - this is a merely technical observation.

1 Like

I walk beside you man!

6 Likes

Acknowledging post-processing differences is crucial when comparing the perceived realism of two projects. It’s possible that differences in post-processing techniques could significantly impact the final look and feel of the renders, regardless of the rendering engine used. So, attributing all perceived shortcomings solely to the rendering engine without considering post-production factors might not provide a complete understanding of the situation especially for this instance.

Cheers

9 Likes

I think that is more related to geometry and camera angles, than shading.
Clearly, the last movie is more ambitious in terms of geometry to show.

There is a will to create tiny details that makes the whole looking like a miniature.
Even with real shots of insanely detailed landscape, you may feel that looks like a miniature. It is normal that falloff of shadow of a small detail is not perceivable.
Small detail on picture → Small shadow → Sharp falloff or invisible shadow.
In Migrants, camera makes fly-by closer to geometry. Details of geometry are bigger.

In space, shadows are supposed to be sharp. There is no GI from atmosphere.
Only bounces from geometry are supposed to attenuate darkness of shadows, but not their edges.
In Solstice-5, shadows are sharp probably because scene is lightened from a Sun lamp with a small angle. There is no will to create velvetry, sheen on surfaces.
There is one shot or two where clouds are softening the shadows.
That is probably requiring post-processing to mix with another lighting with softer shadows.
If there is no coherence between an HDRI lighting and Sun lamp, that is not the software that should be incriminated.

The pattern of small details is probably too regular. There is a lack of intermediate details.
The use of fractal patterns may also reinforce the feel of CG work.
With such complex geometry, bounces of lights are more costly.

So, I am not sure that with another renderer, same feeling will not be present.
That is not obvious that Cycles has to be blamed, here. When there is a technical gap to jump from a lesser difficult to realize project to a more ambitious one.

3 Likes

To be honest, I think that your observations are biased.
Let’s take this shot, that is best to compare, as they quite similar. So, please mark where are those subtle changes in shadows. Honestly, I don’t see any of those nuances.

What I see is that there’s a different artistic choice for post production that I tried to mimic (added fake bloom and lowered shadows tones) and compare to Migrant one.

Another thing: does Paul use newest Blender 4.1 with new Principle BSDF V2 in his project? From 80.lv article I see that he use Blender 3.0 but it could be old screencap, from the middle of production, as it could be few month/years long.


Newest principle shader has better diffuse than V1. And with Oren-Nayar diffuse, when it will be done, it will also better than anything that is available now. Also, as I remember FStorm has in his principle material roughness for diffuse, that has 0.8 value as default and when you change it, it is drastically different perceived. Maybe it could be a good idea to have this parameter visible in V2, dunno.


Two big factors that are a little bit ruining the “realism” aspect are, in my opinion, low resolution smoke and some repetitiveness (scene with many of these diggers from top view, and those fighters on the desert; they should have more variations/rotations, some should be missing/empty spaces).
(besides this, overall, whole movie is gorgeous; thank you Paul for making this, as it makes good marketing for Blender!)

To summarise this, with newest and better shader and color managment, I don’t think that Cycles could have worser final look than other renderers, as all those engines use similar equations for generating shadows, GI and etc., also they have only one reference, that they are mimicking - real world, so they can’t have big differences.
What I can agree that Blender/Cycles is weaker in obtaining quick, realistic results in post pro, e.g. Corona has physical based and better Bloom and Flares, also tone mapper has nice and fast options that makes realistic post production on default! It would be great if Blender would get Compositor Principle Node like this one mockup that I made some time ago.

12 Likes

What I forgot to mention that what is needed is new, simple comparison of engines; real case scenario will tell everything. There were scenes in past, personally I don’t have time for this, but maybe someone has? I guess it would be better to compare Cycles, EEVEE and Octane; Corona is not available for Blender (ok, there’s community made exporter), Luxcore is unfortunately dead.

I think there’s more to it than that. As I said, I’ve narrowed it down to the light falloff characteristics of Cycles vs. other render engines (that is not to say that every single render engine out there is better than Cycles). It seems to be a bit harder for Cycles to mimic the organic falloff of, say, film, than for another render engine. This, at least to me, is visible even in smaller settings, like this one for instance which, is fantastic work for sure, but where IMHO the light falloff seems harsh and “digital” for lack of a better term.

But IMHO where this becomes more perceivable is in large settings. Here’s another example, this is from BlenderMarket, this developer offers a variety of landscapes. I am sure his goal of his examples was to offer as much photorealism and sense of scale as possible, and yet, personally I find all the examples to have the feel of a small model. That is to say there is nothing wrong with the collection itself, I’m sure the geometry is sound. I think it has more to do with the light falloff of Cycles which contributes to some of this lack of scale (among other things).

As a “baseline” I often reference this World Creator/Octane video as a good comparison of similar types of landscapes. At least to me, the difference is quite palpable. I think it’s not in the more obvious parts of the render, as much as some of the nuances in how the light falloff curve is handled by Cycles. As a professional colorist, and one who has spent countless hours trying to adjust these nuances in post, I am not convinced that all of these examples are strictly the result of a different compositing/post production process.

3 Likes

I don’t know. The term narrowed down followed by something as esotheric and handwavy as the organic falloff of, say, film doesn’t convince me.

Furthermore, precisely if the issue has been narrowed down to harshness of light falloff, if I understand correctly, should it not immediately follow from this the issue can be reproduced on the simplest of geometries (e.g. a sphere or cylinder)?

If that is the case, I agree with @jacek007g:

greetings, Kologe

9 Likes

First of all I’d like to make this clear that I still use blender 3.6. Don’t have plans to upgrade to 4 anytime soon unless there’s a significant boost in cycles rendering speed.

Now the harsh falloff you see was my own choice on the renders. I “wanted” the renders to be very contrasty. But I do understand your point about cycles being a bit weird looking(or digital looking). I’ve seen this difference many times when youtubers compare it to Octane, Arnold etc. it’s very subtle. So usually I have to compensate as much as possible in post processing.

6 Likes

This stuff is so incredibly difficult to quantify…just ask any cinematographer about why they prefer film vs. digital and you’ll get equally esoteric answers – yet there must be a scientific reason behind even that, right? Why does the ARRI Alexa look so filmic while the Sony A7s feels so cold and digital?

I guess, going back to my original post, I’m a bit disappointed that, as more and more users adopt Blender 4, that I’m not starting to see the loss of this CG look from any of the latest renders presented here or elsewhere.

3 Likes

Even with the work on the Principled Shader, there are still a number of things you still need the building blocks for to do in the most optimal and high-quality manner. I keep emphasizing that a major component of the ‘Cycles’ look is due to the majority of people not venturing out of the Principled family of shaders when making materials. You mention Oren-Nayer, you can get that now with the individual blocks to mention an example.

Yes, I understand that the building blocks are harder to use at first, but you become independent of waiting for the developers to add specific features (which in the event of uncommon cases is unlikely to happen anytime soon).

Then there are the tricks you can do in combination with other nodes that many are not aware can be done, for instance, the White Noise node being surprisingly powerful when used as a map to drive shading values.

5 Likes

Now that you mention it. I noticed Chris Jones work looks very realistic and he uses Blender Cycles. I remember checking his thread he has here- and he doesn’t use the Principled Shader at all.

3 Likes

Also need a team of people for each engine. Seen too many comparisons where the one person is only truly familiar with one of the engines.

3 Likes

Identical, as much as I could. For example, the first thing that catches your eye is the bump. At first it was even worse in Cycles, but with the Distance option I was able to tweak it. The bump uses a texture. For lighting, hdri is used. Octane 1024 pass 150% of full hd, 3m11s, Cycles render time limit.




изображение
изображение
изображение
изображение

7 Likes

Octane looks a lot better in this case- it feels like it has more dimensionality. To make comparison easier, I’ve put your images in a comparison slider:

a8dd7caf56ffd22b05d927051b4fde93f4f11679_2_690x388
1e8fa1b1620625c8ee4e2699cf663047f22aa529_2_690x388

Octane on the left, Cycles on the right

7 Likes

How much of this is RGB vs. spectral (those balls look pretty saturated, which spectral rendering makes a difference with) and how are we supposed to translate the differences into useful feedback for the developers (as to simply say ‘Cycles is still not realistic’ does not give them anything to work with)?

4 Likes