Data-based emulation of long discontinued film stocks

[Florian Wischnewsky]

Hi there, with all the recent speculation about Genesis, PixelTools announcing a DCTL, with a 2393 profile and the general hype about film emulation, I finally decided to throw in my 2 cents.

This is Eastman Color Negative 125T 5247, negative only, entirely replicated out of spec sheet data:




Same 5247, combined with Resolves 2383:



5247 with a custom 2383:




For comparison – Eastman Color Negative 125T 5247 with 2383 in Genesis:




So, it’s doable, and all of you who have speculated that’s what they did for Genesis are probably right. This is not entirely new, or particularly unique, and a few have already done something similar. Also, I'm dead sure a lot of colour scientists know all this perfectly well and just never talk about it, or even feel the need to.

Besides, Genesis is of course way more refined and sophisticated than what I've thrown together here, and I believe that Mitch Bogdanowicz would have access to more complex data sets, than the publicly available (and he probably created a good chunk of them in the first place). But it may hit the same note on a basic level.

The future of film emulation will be pretty exciting, I guess. And the market will be even more saturated, than it already is.

 

[Oswaldo Rubal]

It’s close to Genesis. I’m really curious about how you approached the replication process to achieve that level of accuracy.

It’s also great to see a revival of these classic film stocks that hardly anyone is talking about. I’ve experimented with Genesis using ECN 5254 and got results strikingly similar to The Mirror (1975).

 

[Rajneesh Kassin]

It’s close to Genesis. I’m really curious about how you approached the replication process to achieve that level of accuracy.

It’s also great to see a revival of these classic film stocks that hardly anyone is talking about. I’ve experimented with Genesis using ECN 5254 and got results strikingly similar to The Mirror (1975).

I don't doubt you believe that, but there were literally thousands of films shot on 5254, and trust me, they all looked different. There was no single "5254" look. There were dozens and dozens of looks. It doesn't mean anything.

At the same time, I don't doubt that Genesis can provide a reasonable approximation of 5219 and 2383 used together. Everything else I think is going to be a guess. Mitch Bogdanowicz is a bright guy, and I'm sure he knows the math and all the density readings of different film stocks going back 50 years.

 

[Florian Wischnewsky]

It’s close to Genesis. I’m really curious about how you approached the replication process to achieve that level of accuracy.

It’s also great to see a revival of these classic film stocks that hardly anyone is talking about. I’ve experimented with Genesis using ECN 5254 and got results strikingly similar to The Mirror (1975).

Accuracy is a bold term here. It's basically a big, nasty pile of WIPs, patchworked together.

I'm still kind of surprised, how close it comes, by basically just implementing the data.

At least, I can say for sure, that Genesis is really accurate. I'm not so sure, if accuracy is the right premise in general, though. As you probably know, real film always has an element of interpretation and there is much more going on. Or rather not going on here.

 

[Firat Konuslu]

Amid all these film emulation discussions, I’ve noticed that Filmlab AI is almost never mentioned.
Dado has been sharing his thoughts and explaining his process on his YouTube channel for a while now, offering insights that I personally find quite valuable.
Maybe you’d like to check it out.

 

[Jason Bowdach]

Keep up the great work! The future of film emulation is very bright

 

[Rajneesh Kassin]

Keep up the great work! The future of film emulation is very bright

What if we want it to be dark and moody?

 

[Florian Wischnewsky]

Keep up the great work! The future of film emulation is very bright

Thank you, Jason. I appreciate the kindness.

 

[Cyriel Roumen]

That's definitely very close, nice!
What is the process like to take a spec sheet data to create this emulation?

 

[Rajneesh Kassin]

BTW, I don't know if anybody is watching it, but the two best film emulation TV shows I've seen for the entire year are It: Welcome to Derry (on HBOMax) and The Runarounds on Amazon Prime. Phenomenal color on both shows. I was absolutely convinced Runarounds was shot on film (35 / 16 / Super 8) but no.

I'm not sure who did the color on Runarounds, but I see that Dave Cole and his crew at Fotokem/Burbank are doing Welcome to Derry. Really, really fine work. In particular, Welcome to Derry very much has an early 1960s aesthetic that I totally believe.

I have no idea what process or workflow they're using, but whatever it is, it's brilliant.

 

[Florian Wischnewsky]

After a longer hiatus, here’s an update with a couple more examples.

Initially this was about reviving the looks of deprecated stocks, and it still is for me. But for the lack of well… deprecated stocks, to shoot tests with I mostly focused on Kodak Vision3 and to a degree on Fuji ETERNA, for now.

The process is basically spectral integration creating a profile that’s, in theory, fully invertible. As a sidenote, you can use such a profile to normalise film scans (depends on how they’re scanned, though). If you invert it, you get film emulation.

This works for negative, print and for reversal film. Anything you have a spec sheet on.

I began working on this a couple of years ago, rooted in the wish for a handful of reliable and unbiased negative LUTs that would go well with typical FPE LUTs and didn’t rely on camera matching at all. That would also give me the freedom to use whatever grain, halation and MTF plugin or DCTL I want.

For reference, I did a bazillion of test shots with all Vision3 stocks and used Filmbox Lite, the unloved FilmConvert Nitrate and since it came out, Genesis 1.0.5.

This is still work in progress, next step would be to replace the ACES Gamut Compression I’m utilizing temporarily. But it pretty much works for now, and I would carefully call it production ready. I want to have a few more things changed though.

Enough talking. Here are some screenshots:

1.

Custom 5207 – with 2383 from Resolve


5207 with 2383 from Genesis


5207 with 2383 from Filmbox


2.

Custom 5207 – with 2383 from Resolve


5207 with 2383 from Genesis


5207 with 2383 from Filmbox


3.

Custom 5207 – with 2383 from Resolve


5207 with 2383 from Genesis


5207 with 2383 from Filmbox


4.

Custom 5207 – with 2383 from Resolve


5207 with 2383 from Genesis


5207 with 2383 from Filmbox


cont.

 

[Florian Wischnewsky]

5.

Custom 5207 – with 2383 from Resolve


5207 with 2383 from Genesis


5207 with 2383 from Filmbox


6.

Custom 5207 – with 2383 from Resolve


5207 with 2383 from Genesis


5207 with 2383 from Filmbox


At one point I was seriously considering making my own little plugin. Though, since the market for film emulation became so saturated and the whole spectral thing became such a hot topic recently, I don't really see a point in commoditising this anymore.

Anyway, I will make this available pretty soon. In the state it is now, and probably for free.

So far for now.

Cheers

 

[Silvan Schnelli]

The results look great. So if I’m not mistaken you are using the spectral data only for simulate the negative part and then applying Davincis print film LUT or a tweaked version of it.

I’ve been trying to get a FPE LUT for a bit now with spectral data, but I feel like the printer lights part is messing up my output. Recently though I’ve been thinking that I should just emulate the neg part first and then apply an available print LUT. Your results show that it can also lead to pleasing results.

Is there a reason you didn’t emulate the print directly?

 

[Silvan Schnelli]

Also since you mentioned the avoidance of camera matching and ACES compression, does this mean your digital actor is operating in ACES?

 

[Firat Konuslu]

This has probably been answered somewhere before, but I’m asking from the perspective of someone observing from the outside and trying to understand the issue.

Here we’re looking at the results from three different film emulation plugins/attempts, all applied to the same negative and all three differ in some fairly fundamental ways.

Yes, we know that when doing film emulation, even if it’s the same stock, each physical roll or batch and process has its own uniqueness. But in this case, the differences seem too significant to be explained merely by variations between different samples of the same negative. Or is it so?

So my question is:
Are the natural deviations between film stocks really that large?
Or are we looking at fundamentally different interpretations of what that film stock is supposed to look like?

 

[Florian Wischnewsky]

The results look great. So if I’m not mistaken you are using the spectral data only for simulate the negative part and then applying Davincis print film LUT or a tweaked version of it.

I’ve been trying to get a FPE LUT for a bit now with spectral data, but I feel like the printer lights part is messing up my output. Recently though I’ve been thinking that I should just emulate the neg part first and then apply an available print LUT. Your results show that it can also lead to pleasing results.

Is there a reason you didn’t emulate the print directly?

Thank you, and that’s exactly what it is. I started out a couple of years ago with what I was personally missing. Solid negative emulations, that go well with the FPE LUTs from Resolve. They are very good and especially the FPEs you can export from LightSpace/ColourSpace are magnificent. So I didn’t really have an urge to recreate FPEs besides the ones, you can’t get anywhere else of course. In the screenshots it’s the one from Resolve. Despite its errors, it gives out a pleasing finish indeed, which was both surprising and satisfactory for me, initially.

 

[Florian Wischnewsky]

Also since you mentioned the avoidance of camera matching and ACES compression, does this mean your digital actor is operating in ACES?

It’s operating in XYZ, as spectral integration is working in XYZ. From there you can go anywhere. The ACES Gamut Compression is incorporated later in the finishing and pretty much a placeholder for something more elegant, I've yet to deliver.

The final LUT expects DWG/I and the primaries of the profile are also set to DWG/I before the finishing.

It’s really just a temporary solution. There's no ACES gimmick.

 

[Florian Wischnewsky]

This has probably been answered somewhere before, but I’m asking from the perspective of someone observing from the outside and trying to understand the issue.

Here we’re looking at the results from three different film emulation plugins/attempts, all applied to the same negative and all three differ in some fairly fundamental ways.

Yes, we know that when doing film emulation, even if it’s the same stock, each physical roll or batch and process has its own uniqueness. But in this case, the differences seem too significant to be explained merely by variations between different samples of the same negative. Or is it so?

So my question is:
Are the natural deviations between film stocks really that large?
Or are we looking at fundamentally different interpretations of what that film stock is supposed to look like?

Three important things to point out, before I answer the question.

1.
Nobody knows, how Genesis was build, besides the people involved. I'm perfectly fine with Cullen not disclosing it, and I certainly won’t ask him that question, just to know. Especially not, if I'm about to release a "product", he could in any way view as competition. That would be wildly inappropriate. So, You have to go with my more or less educated guess.

2.
I'm almost dead centre certain they’re doing something similar, just more complex and refined. Mitch Bogdanowicz would have access to way more information than what’s in the official sheets. He pretty much delivered a good chunk of it by himself, when he was at Kodak. Though, it seems obvious that they also model something out of spectral data points, when Kodak et al. are measuring and publishing that data for decades. Also, something with Genesis felt instantly familiar, when I looked at it first. Or not, maybe I'm completely mistaken. Who knows.

3.
Filmbox works differently and is dependent on actual scans and how they come out. They would have developed and evaluated many negatives and not just one, to get an average of the ideal outcome per stock. So would every film manufacturer in their sheets. But you would still get two different averages and therefore two different results.

With that out of the way, even with two identical data sets, you can get wildly different results, depending on how you interpret them. E.g. a real negative retains way more saturation in the highlights, than any film emulation, including mine. That’s especially true for the reds. You just have to make compromises in emulation (and even in scanning also). The tricky part for me personally is to still give out a jarringly bright red in the highlights, without breaking anything or compromising skin tones in any way.

And that’s just one aspect of many. As mentioned above, scanners are another one, and a complete can of worms. It’s chaos.

 

[Curtis Peacock]

Yes, we know that when doing film emulation, even if it’s the same stock, each physical roll or batch and process has its own uniqueness. But in this case, the differences seem too significant to be explained merely by variations between different samples of the same negative. Or is it so?

So my question is:
Are the natural deviations between film stocks really that large?
Or are we looking at fundamentally different interpretations of what that film stock is supposed to look like?

One of many reasons why various film emulation solutions don't match is that they have to start with RGB data instead of full spectral data.

The cameras we use reduce a spectrum of light to just three values. It is impossible to determine the exact original spectrum from RGB values alone, as there are infinite possibilities. Two distinct spectra might result in the same RGB digital exposures, but different RGB film exposures, because the spectral response of the digital camera differs from that of the negative film.

Whether we try to match the appearance of film by doing a real-world test with colour charts or by using a spectral model, we still have to assign exactly one spectrum to any given digital RGB input.

 

[Rajneesh Kassin]

Three important things to point out, before I answer the question.

1.
Nobody knows, how Genesis was build, besides the people involved. I'm perfectly fine with Cullen not disclosing it, and I certainly won’t ask him that question, just to know. Especially not, if I'm about to release a "product", he could in any way view as competition. That would be wildly inappropriate. So, You have to go with my more or less educated guess.

2.
I'm almost dead centre certain they’re doing something similar, just more complex and refined. Mitch Bogdanowicz would have access to way more information than what’s in the official sheets. He pretty much delivered a good chunk of it by himself, when he was at Kodak. Though, it seems obvious that they also model something out of spectral data points, when Kodak et al. are measuring and publishing that data for decades. Also, something with Genesis felt instantly familiar, when I looked at it first. Or not, maybe I'm completely mistaken. Who knows.

I worked with Mitch and his right-hand person Peter Postma every day for 2 years at Kodak's CInesite division, and they straight up told me, "we developed the film emulation LUTs for our proprietary WildStar LUT system by shooting dependable charts and lots of trial and error." They told me it took years and years before they were able to predictably see reliable results when looking at digital material in the DI theater, then shooting it out to a Cineon film recorder and making an answer print based on that, which we'd screen the next day. It did not work well when they did O Brother Where Art Thou -- Roger Deakins was absolutely irate about the amount of time it took -- but it worked by the time I was there from 2002-2004.

Here's a one-hour video interview/discussion between Cullen Kelly and Mitch Bogdanowicz as to how Genesis was developed:

No "great secrets" are revealed, but they reveal the gist of it.