Blender 4.2 - EEVEE-Next Feedback

[Lawrence_Teng]

We have to maintain two versions for now (actually three because the shader setup changed from 3.6 to 4.1), anyway we had to save an Eevee next version of our main assets to keep our stuff in the loop even if we are not planning on using Next for now. We will miss out on other improvements though so we get your point. But 4.1 has most of what we need for the time being while we wait to make the upgrade to 4.3 or beyond and by that time, we will have all our files already up to date and ready to go. It’s extra work to maintain multiple versions, but it is necessary for us. We still work with Maya for some of our projects and we have to do the same thing, we have multiple versions of the software and the corresponding file versions just in case. I think this is same for most software.

There is a case of “diminishing returns” when it comes to Legacy vs Next renders. Next took three times as long or more compared to Legacy and the visual improvement does not correspond to this; it is certainly not three or more times better visually. Furthermore, all we are simply doing is trying to make it visually the same; we are not even trying to make it better.

We also noticed online Eevee reviews don’t mention this slowdown, but we assumed that these users don’t really use Eevee as their main render engine. They use Cycles and Eevee seems to be treated like a material previewer that is “faster than Cycles”, i.e. Cycles-lite. We don’t quite agree with this mentality but this is just us.

[thorn-neverwake]

@Lawrence_Teng

Furthermore, all we are simply doing is trying to make it visually the same; we are not even trying to make it better.

This is what is completely keeping me on 4.1. Just getting the same (or let’s call, it close enough?) result to 4.1 is impossible. Add this to the speed factor (tweak something, go have a cigarette, come back) - I’m not finding just a diminishing return, but a no-return.

@Lawrence_Teng

We also noticed online Eevee reviews don’t mention this slowdown, but we assumed that these users don’t really use Eevee as their main render engine.

That is my assumption, as well. I watched most of one of them from a rather popular channel, and it felt like they were going through the bullet of list of new stuff and just going with “Wow, that’s amazing!” one by one. Some of the things that aren’t working well are not hard to miss, so clearly they either a) weren’t looking for them b) never really use Eevee or c) didn’t want to point out the problems, for whatever reason.

I’m perfectly ok with a new rendering engine not being feature complete, work in progress, whatever term one wishes to use. That’s the process. But it’s a bit mind-blowing that Next has replaced Legacy, at this phase of maturity.

(Which isn’t the first occurence of “mind blown”; that was when learning that Next wasn’t tested much/developed on a system with an nVidia card until rather late in the game.)

[Illasera]

Yes sir, something new came out and it ended up being slow (Even after testing and delaying release) a shocker, anyway.

any leads as to what can cause the slowdown that may help the developers?

Yes we can see that it takes more time to render, with the snippet of “Benchmarking” provided above; however the question is why? what causes the bottleneck?

is there a way we can help testing it? i mean, obviously beside manual debugging; I was thinking of a more verbose output for each step of the rendering process.

As for : Was it released prematurely? the answer to this will ALWAYS BE : Yes and No, No if the developers think its ready but with wider audience you will find new problems.

Ace_Dragon Nowadays, it should be the responsibility of software makers in general to stop relying on Moore’s Law to keep their apps usable and instead have more of a focus on optimization to help people stay on their current hardware for longer periods

Things got more complicated than that , its not just moore’s law , its also a question if : “Do i want to buy new hardware”

Intel is now at the phase of recalling CPUs (Probably due to outsourcing their work and cheap labor, won’t go into that); also testing the water with “subscription based unlock-able CPU features”

nvidia’s software bundle (nvidia experience) is getting beyond ridicules per update and logitech tells me that i need to pay a fee every month to use their mouse.

we are at a point in which people will refuse to upgrade even if they have the means, blender devs should account for it, it will only get worse.

Most users fail to realize how powerful computers are and i am not talking about modern 2024 computers with RTX nvidia cards but 2007 computers as well.

Any software can always be optimized further, the question is ; how much time and money are we willing to throw at it.

A word of gratitude to eevee-next developer and maintainers , Thank you.

[rogper]

Youtubers mostly are youtubers and not professional 8 hours by day blender artists. Anyone can pick up some models, make an impressive scene that doesn’t even pushes eevee to 50%…

On my current cinematic trailer, I got 3 weeks to make it, so I full comited to make it on 4.2, no fail safe to 4.1.
It’s an interior scene, so Next light bounces really become handy. It’s all maxed out, if I add one more lamp it crashes. But even on Legacy this would be simillar (didn’t tested it).
I’m at 3 days to complete it and, if you really don’t look back, Next is ofering solutions… one just can’t think in legacy terms and instead embrace the Next ways

[Illasera]

On the note of youtube hype, i am glad that people are bringing this up .

*This comment by no means is against blender developers

Youtubers hyping and not disclosing critical issues (And i am not talking about eevee-next but many other aspects as well and not just in blender); are lying by omission (if they are aware of such issues and fail to disclose them, generating hype trains);

@thorn-neverwake
Many Youtubers share the goal of creating positive vibe and being able to name their videos as: “Why blender is awesome” and what not, to quickly catch the viewer’s eyes and pump the view numbers up.

Since version 3.3.1 i have been waiting to update, decided to update last week to 4.1.1 (NOT 4.2, because i speculated that EEVEE-next will have issues to iron up);

If i were to follow youtubers, i may have ended up in a bad situation (Not that downloading previous version is hard but there are compatibility issues).

[Lawrence_Teng]

Both you and Illasera brought up some really great points.

We have no desire to go back of course, but we have the luxury to rest and pause, until things settle. We tend to not upgrade right away and do it after a few versions; we stayed with 3.6 for a while and just upgraded to 4.1 in March (bypassing 4.0). We are thankful that previous versions still work very well and Blender foundation is not forcing us to upgrade in a hostage subscription situation.

In contrast to this, we are reminded of when our main software Softimage was killed in front of our helpless eyes, and we are coerced into going back to Maya, and then later on Maya’s main renderer at that time, Mental Ray, was killed also and we had to use Arnold. I mean they are all good software so it’s not a functionality issue, but it’s a strange feeling to be held hostage in an ecosystem of this sort and be funneled around helplessly. The irony of this too is that Blender is not charging us a cent, but Autodesk…whatever don’t want to go into a bitter rabbit hole. This is also why we have decided to start gradually move things over to Blender, and once we weaned ourselves off a good chunk of our Autodesk licenses, we will use the budget to support the Blender foundation. It has been mentioned in this thread a few times already that the limited resources made it impractical to keep legacy and next together, we figure that we will help in whatever small ways we can.

[Steve_K2400]

I’m working for an automotive design studio where we started using Blender 3 years ago for visualization and concept modelling. Visualisation both for real time reviews with Eevee as well as renderings and animations with Cycles.
For this purpose I wrote an Add-On to help bring Alias data (via FBX) to Blender and thereby automating material conversion and providing environments and stage sets with a Library approach. (In such a way, that Blender novices can achieve quick visualisation results without actually knowing Blender. This is a very different use case compared to use cases in entertainment studios where the Blender users are experienced.)
The first thing we noticed with 4.2 was that the scenes and libraries (essentially template files) opened in 4.2 produced unrealistic results (the shadows under the car for example were not dark enough) and so real time reviews were no longer possible with the same dataset in 4.2.
Tweaking lights/materials/environments have not resolved these problems and even if it would work, keeping two sets of libraries requires far too much work and time.
In the 3 years we never experienced such “show stoppers” in the development of Blender. We could deal with breaking changes in new versions quickly and the “ride” was quite smooth until now.
It’s clear that development has to continue and sometimes one needs new ground to build upon, but in this case, for us, it brakes too many things.
The name change (Eevee Next to Eevee) makes things even worse as casual and beginner users who have not followed the development process assume that they get the same look from the same Renderer with the same name.
So, for our visualization use case (which is certainly a non standard one) the new version is unfortunately not usable. We will stay on 4.1 and see how things proceed in the coming months.
This is not intended as a complaint but rather to tell our experience from a different perspective.

[ThomasDinges]

May I suggest to either

a) Post clear information, a screenshot, example renders etc. so others can more easily follow what’s going on and help

or

b) Make a complete bug report using Help -> Report a bug from within Blender if you’re sure there is a bug.

Personal fights don’t help here and will be moderated.

[Illasera]

Yes but once again, i would like to re-suggest to post the blend file of both versions (Replying to the person in question @Steve_K2400), as the problem discussed here might be a settings problem.

[Steve_K2400]

Wow, gone for two hours and everything exploding… Sorry for answering late but I also have to work.
Providing files or images from our actual data is impossible as we are a car design studio and I’m not allowed to upload any data.
I quickly built a simple sample scene on my private machine that shows our problem:

How can I upload Blender files here? (Currently I can only upload images.)

[Illasera]

Nothing to apologize for, there was no time limit on replying, sorry if it felt urgent.

as for uploading.

i have seen people using google drive or other hosting websites and posting links to them here.

If what you are working in is a company secret, ask for permission from your supervisor, if non granted; a mockup like the above will do but make sure it highlights the actual problems you are experiencing.

[Steve_K2400]

Sorry, I thought I can upload hem here. I put them on my server:
Blend File 4.1
Blend File 4.2

In this industry, we are generally not allowed to upload anything. I guess it is not much different from Entertainment or game studios working on their next release…

Personally I’m a huge fan of Blender (after working with Alias for 33 years) and the adoption of Blender in the automotive industry increasing rapidly. That’s the reason why I would like help solving issues which we are currently facing with the new release. It is entirely possible (highly likely) that it’s my user error that is causing this problem but I tried a lot of suggestions and were not able to get the same look as with 4.1.

Even if it’s my mistake, it means that there might be more information needed for the transition to Eevee Next as currently provided.

There is for example a page in the current documentation about the ambient occlusion render settings… but in 4.2 these settings are gone:
Ambient_occlusion in 4.2 documentation

Just to be clear as I forget to put markers on my images:
The top image is 4.1 and the bottom one is 4.2. The Problem on the images is the missing dark shadow underneath the car on the 4.2 image.

[fclem]

You scene seems to exhibit artifact from huge coordinates. If you scale everything to 1:100 it looks correct.

This is undeniably a bug an should be reported.

[Steve_K2400]

Yes, we are using 1:1 scaling as the data from the different CAD systems are coming in with real size scaling.
Also, we are using millimeter as the unit with 0.001 unit scaling, but I guess that should not be a problem.

I scaled down the scene to 0.01 and the visible artifacts (when Raytracing - Max Roughness is < 1.0) are gone as you say.
The dark shadow underneath the car is still missing so, for that, the size seems to be not relevant.
BTW: If I turn off Ambient Occlusion in 4.1 in the Render Settings, I get roughly the same bright shadow as in 4.2.

[kram1032]

I’m not sure much can be done there: This is a precision problem.
What you could try is to, rather than adjusting the total scene scale, adjust the near and far clip of the camera (by the same 100x). That might also work to resolve this. Especially increasing the near-clip.

If that doesn’t work though, it’s just floating point precision limitations

[Steve_K2400]

If it’s a precision problem, then Eevee legacy and Cycles would show similar artefacts but we use both render engines without problems for years now. (Ever since we started with our workflow, we are using 1:1 scale models.)

These shading inconsistencies appeared only in Eevee Next and only if the Raytracing → Max Roughness value is less than 1.
In Eevee Legacy and Cycles only when the near clipping plane is set to a very low value (under 1 mm) there are visible artefacts and I think that shows the limits of the single precision calculation.

I think, the missing dark shadows in Eevee Next under the car are a result of a light calculation problem.
I just made a photo outside:

It’s 8 am, low standing sun, super bright day with clear blue sky and still, underneath the car it is pitch black.
Even if AO was a fake calculation method, it captured reality (at least in this case) quite well.
Eevee Next should give the same realism but I was not able to achieve this. (This could be entirely my fault, but then please show me how to do it correctly.)

In my simple model above I changed the bottom of the car to a black diffuse material and even then there is no change in the way how the shadows are calculated.

[Steve_K2400]

I made another test with the Light turned off and only using the new sun from the HDRI feature:

This image was done with Max Roughness = 1, means Fast GI Approximation disabled.
It shows a perfectly even shadow color depicting the light facing contour of the object but no interaction of objects in close proximity.

If I turn on Fast GI Approximation (by lowering the value below 1) the result is better:

But then the artifacts due to the large scale start to appear.

Is Fast GI Approximation a similar process like the old AO?
Disabling it should mean that a more physically correct calculation should take place, or am I wrong?

[rogper]

This image was done with Max Roughness = 1, means Fast GI Approximation disabled.

Thanks for mentioing this I tought Max Roughness = 0 was Fast GI Approximation disabled (because there’s many other cases in these blender sliders that zero is the thing turned off)… man… This is very unclear… why the term “Max Roughness” to turn off GI!?

Why not just “Fast GI” or something like that Then Fast GI value of 1 would make sense it was Fast GI turned on.

[mhjang]

Based on the test results, it seems advisable to submit a bug report. When the scale increases, nothing seems to work properly. Even when the scale is reduced, occlusion does not appear as well as it did in legacy versions. Until everything is improved, it would be wise to stick with version 4.1.

[Steve_K2400]

I’m not experienced with bug reports, therefore my question: should I treat these as one problem (artifacts at large scale + missing AO) or should I file two separate reports?