Radiance and Autodesk Insight and Sefaira


#1

I have started a discussion thread in the past about a comparison between Radiance and other rendering engines, but this time, I wondered if anybody had any specific experience with Autodesk Insight and Sefaira? These two differ from artistic rendering engines in that they claim to use the same lighting model as Radiance (with a simplified daylight model) in the case of Insight, and perhaps they just use Radiance in the case of Sefaira.

In particular, this is the introductory webpage about Insight and their daylighting services:

https://blogs.autodesk.com/insight/daylighting-as-a-service-raas-illuminance-radiance/

I have copied some of the more intriguing claims:

The Rendering setup uses the same daylighting and electric lighting models that are used for the visual rendering for the time being. While the electric light model is the same as Radiance and other industry standard tools, and uses the fixture and lamp photometric models, if included, the sun light (daylight) model is slightly simplified. To be consistent for visual renderings, the sun light (daylighting) distribution is modeled using the Preetham sky model, which assumes a clear sky and no aerosol effects like fog or haze. The underlying algorithm for this model is similar to the more advanced Perez model, but the heuristics assume standard solar response and therefore do not use solar radiation or illuminance values from a weather file.

In another article coming to this site in a couple weeks, we will talk about the technical validation work that has been done on the illuminance output from the Autodesk 360 Rendering engine. For the moment, we can say that the engine has been validated with the help of one of the top daylighting firms in the country by comparing our cloud rendering output using the same model rendered in the industry standard lighting simulation tool Radiance, as well as to the actual space measured in the real world.

Results of the validation tests are strikingly similar in all ways but one…the time it takes to create the rendering. When we asked about the level of quality used in the Radiance rendering and how much time it took, the answer was: 5 bounces and it took about 4 hours to render. When we asked the cloud rendering technician how many bounces, the answer was ‘all of them’, and the rendering took about 10 minutes .

As for Sefaira, here is some related reading I found:

Any thoughts and the above two products and how they compare to a usage “vanilla” Radiance?


#2

Hi Dion,

I have no particular experience with those tools. However, I wanted to comment that I am surprised by the calculation times they mention. I do not think I have ever made a routine analysis (e.g. DF, DA, or similar) that took 4 hours using Radiance (except, perhaps, if you include my extremely slow Ruby scripts that sometimes process my data).

For instance, I have been using the (fairly complex) model below for making some optimization tests in Groundhog (which now uses highly optimized routines, I must say). It takes about 6-7 minutes to calculate UDI and DA using 2-phases in my late 2012 MacBook Pro (including meshing of workplanes, and post-processing data). If I add ASE or use Direct+Daylight coefficients (i.e. requires calculating sharp direct suns) it takes 33 minutes in the same computer. It should take much less time in the cloud.

I am using -ab 7, lw 1e-5, ad 5000, aa 0.1.

They say they made a render, which has never been part of my routine analysis… I guess that might take 4 hours.


#3

I also have no experience with these tools, but I’ve talked to the people who make them.

Sefaira does indeed use Radiance. They presented their approach last year at SimBuild/BPACS under the same title as the article you linked. The settings they use are a secret, but from the artifacts in their visualizations, its fairly easy to tell that they use a low -ad value to achieve their speedup. (For the record, the “specialist suggested” settings in their graph came from one of my conference papers.)

For the Autodesk renderer, keep in mind the date of that article (2013). In 2009, Christoph Reinhart published that Autodesk 3ds Max gave similar daylighting results to DAYSIM and faster, but a 2015 follow-up study by Bellia that considered more cases found 3ds Max to be less accurate. This highlights the point some people made in your previous thread that changes to proprietary rendering engines tend not to be documented, which is an advantage of Radiance.

I’m not sure why using the Preetham sky model instead of the Perez sky model is considered a shortcut. They’re both very simple calculations that shouldn’t have an effect on simulation time. Probably, Autodesk was trying to limit the effect of weather on lighting in order to make results look more uniform from day to day. However, nowadays we typically produce climate-based metrics using matrix-based methods on a full year of simulated data. This “shortcut” is no longer relevant.

I’m not sure who Emile Kfouri was quoting about the 4-hour Radiance rendering. Alstan Jakubiec listed 4 hours as a typical rendering time for daylight glare probability renderings in his PhD thesis. However, it’s not necessary to use extremely high accuracy settings for this because DGP is insensitive to most rendering artifacts. In any case, I’m not aware of Autodesk having a glare analysis product.

I heard the story about rendering “all of the bounces” back in 2011. It’s kind of meaningless, though. Basically, what Emile is saying is that their rendering engine uses a method (probably Monte Carlo) that doesn’t count the number of bounces. It likely just extinguishes rays with some probability at each intersection. If you were to count how many bounces occurred, you would find that most ray paths had very few bounces, and only a few go through many bounces. You can achieve a similar behavior in Radiance with negative -lr settings.

Autodesk does have a small rendering team that develops their own engine. I’m pretty sure that it’s changed substantially since that article was written, though.


#4

Thanks for the explanations @German_Molina and @Nathaniel_Jones. I notice you talk about Autodesk 3DS Max, which I assume was rebranded into Insight 360, but is the same product underneath.

Thanks also for the explanation about the Preetham sky model. It makes sense.

I guess if Autodesk is developing their own engine, the usual arguments in the previous thread about undocumented changes and secret simulation options in proprietary engines apply.