September 29, 2020

Host: Michael Dolan

Attendees:

  • [X] Mark Boorer (TSC) - Industrial Light & Magic
  • [X] Michael Dolan (TSC Chair) - Epic Games
  • [X] Carol Payne (TSC) - Netflix
  • [X] Doug Walker (TSC Chief Architect) - Autodesk
  • [X] Matthias Scharfenberg - Industrial Light & Magic
  • [X] Dennis Adams - Sony
  • [X] Thomas Mansencal - Weta Digital
  • [X] Christophe Brejon

Apologies:

  • Kevin Wheatley

OCIO Configs Working Group Meeting Notes

  • Update on config skeleton generation work:
    • Thomas: For practical purposes it's in a state where it should work with builtin transforms. Most concerning issue right now is the ACEScc builtin transform is not serialized. The transform is created correctly, but when the config is written, nothing is in the config. Good that it doesn't support non-existent style. Also CDL transform methods not working.
    • Doug: We have unit tests for CDLTransform, so not sure. May be untested issue.
    • Thomas: Serialization issue is just this one transform. Others are fine.
    • Michael: Should we auto generate builtin transform cpp code or do the work manually?
    • Doug: Depends which transforms, but if we at least agree on names, that's a good step forward. If we agree to take them from CTL, as Thomas is doing, that makes sense. Going to come down to manually writing the code.
    • Michael: Agree on manual approach, to have comments, etc.
    • Thomas: More thinking to generate stub files to be filled, since the data is there.
    • Mark: From current config, FileTransform in there now. Will there be external files?
    • Thomas: Just a placeholder until builtin transform styles are supported.
    • Michael: Thomas raised a question about direction in PR. As to whether we should follow typical transform forward/inverse behavior, or have separate named transform per direction?
    • Thomas: Either approach is fine with me, but prefer small things which are easier to test and maintain.
    • Doug: all transform have direction attribute, so needed, but nothing prevents us from having both directions as transform names.
    • Mark: Prefer log-to-lin since the inverse order has precision loss in a LUT, but for the facade that the builtin transform has it shouldn't matter. Use most optimal and have both in interface. If people don't need to read docs and just use transforms, having both directions would make it easier.
    • Matthias: From config author standpoint, good to have both.
    • Thomas: Make it clearer when going from reference. More explicit is better.
    • Mark: not sure how helpful to have every transform in both directions. Do we need to do output transforms in both directions For example?
    • Doug: We don't have to explicitly put them in config if no reason to. If OCIO can generate the other direction. Was that the questions, needing both from_ref and to_ref?
    • Thomas: Since it's from CTL, will get both sides from some transforms.
    • Doug: In general there is a need to combine two CTL files in repo to generate a single color space. Need some logic for that. Sounds like we agree that builtin transform should define strings in both directions, but then there's the question of the aces repo that has CTL in both directions (no inverse argument). Thomas' generator looks at CTL and builds color spaces. Most of the time there are two CTLs but don't want two color spaces.
    • Thomas: In aces-dev if both directions are defined, you’ll get from and to reference transform in color space in the config.
    • Mark: Think that’s fine. OCIO generated in reverse not as good as manmade inverse. Don't think thats a problem.
    • Thomas: Think it should stay as close to metal as possible, to aces-dev. The studio config can be massaged differently, but this one as a base should be close to CTL. Could be good for Scott and Alex to be able to see it as a test bed.
    • Mark: Only new feature in OCIO v2 in this regard is 3D LUT inversion and don't want to rely on that vs. using the actual inverse.
    • Doug: shouldn't be 3D LUTs in aces config unless there is camera vendor spaces. Under the hood the builtin transforms we will just write one and the other will be implemented with direction flag. CTL not ideal in this regard. OCIO generated inverse will be at least as good as CTL.
    • Carol: Think an explicit definition for any 3D LUT in CTL would be a case for that. If we all trust and use the inverse calculated by OCIO, that should be what we use in the config.
    • Mark: The builtin transform won't show you that they need a from_reference. Each has it's own implicit from_reference. If you use builtin transform you can assume an inverse exists. Should guarantee that all builtin transforms are invertible.
    • Carol: Should not need to define both from or to reference. Since it is mathematically high precision and invertible, defining both not needed.
    • Thomas: An interesting case is the IDT going to reference and ODT from reference. We could do one or other, but could do it the way CTL is implemented. If we want to impose a direction we can, but think there is value for Scott and Alex and AMPAS people to represent how they are implemented on CTL side as faithfully as possible.
    • All: Agree that is fine.
    • Matthias: Could also serve as a guideline for how to build an OCIO config. Helps see where it's good practice to have transform from and/or to reference, that don't need both always defined.
    • Thomas: Yes, it's logical, that's how it flows in ACES, so good to follow that.
    • Mark: Might end up having to use builtin transform with opposite name as a result, but that's fine.
    • Michael: Any interest in starting work on builtin transform implementation on the cpp side?
    • Mark: I am interested, but hoping to replace it eventually.
    • Michael: Good to see current implementation to understand the needs, even if it changes later.
    • Mark: Will look at what Thomas has and give one of them a try.
    • Doug: Recommend not starting with ODTs. Would be better to use spline operator for tonecurve, so I should provide an example of an output transform first. Good to start with input transforms. Already some examples in there for reference. A basic starter set, all of the ACES color space conversions are done.
    • Mark: Is a LUT built at runtime?
    • Doug: In some cases no LUT is needed, when CameraLog transform can be used, but we did include some examples where Lut1D is needed. A lot of the curves use the same formula with different constants, so that can be used when the formula is similar. ACEScc transform needs an update, but there are examples of 1D LUTs in there.
    • Mark: Skeptical that they match. Had problems where formula didn't match official spec at bottom of significant digits, so mixing official LUT and parametric formula, not sure they will match and roundtrip.
    • Doug: Fair amount of discussion in CLF working group around this. Scott and Alex signed off on this approach. Sometimes manufacturers don't do good documenting for implementation because of included significant digits.
    • Matthias: Has anyone come across anything that doesn't work with this?
    • Doug: There are some examples. ACEScc needs Lut1D and some camera examples (Cannon) wen their own way.
    • Michael: OCIO follows the CLF implementation, so ops should adhere to that.
    • Doug: Yes, but CTL doesn't.
    • Mark: Disappointed we're not generating these at build time, but probably not expensive to build at runtime.
    • Doug: Faster than parsing LUT file.
    • Matthias: Reasoning behind maintaining color space naming from previous ACES config? Where family name duplicates color space name and makes long color space names. Can break combo boxes.
    • Carol: It’s the academy’s names, matching CTL transforms.
    • Matthias: Do we want to keep family name in front of everything?
    • Thomas: Can trim it.
    • Matthias: Already condensing names in builtin transforms. Would prefer those names to the color space names.
    • Thomas: The current naming trying to stick to original naming, but easy to change. Would be good for people to pull down the code and start to play with it. Might give some ideas. One of next steps was going to do was dockerize the generation. Will work with OCIO containers.
    • Michael: Aloys is working on ASWF docker images for production. Beyond the current CI images, which will have full builds with tools, etc.
    • Thomas: Anything we can do to streamline the work for people would be good. Docker is a good solution.
    • Doug: Agree that sounds useful.
    • Thomas: Will work on that to make the setup one click.
    • Michael: Do we have a decision on color space naming changes?
    • Mark: Once we can get a config to play with we can play with and suggest changes.
    • Thomas: Can be parameterized too.
    • Mark: Could get in touch with Academy to get their blessing on improved naming.
    • Thomas: Yes, need to follow up with Scott on that. Would like to have official conversation with them. Work for consistent names in the future. Should start official conversation.
    • Mark: Once we get something we can generate and play with that would be a good point for that conversation.
    • Thomas: They can build it to look at on ACES side too.
    • Mark: Would argue away from using docker. Just a python script to run would work.
    • Thomas: If it was pure python no problem, but with OCIO, etc. it gets complicated.
    • Mark: Should just need pure python to build graph. Do we need all these dependencies? NetworkX, etc?
    • Thomas: You could, but good to utilize libraries that are there. You can replace, but more maintenance on our side.
    • Mark: Every dependency has a cost. Gets passed on to users. Having to download lots of libraries to build a config not ideal.
    • Thomas: Matter of balancing cost. If you have something available that works well, why reinvent the wheel? Just use that.
    • Mark: Gets hard for build engineer in studio.
    • Thomas: Is work, but helpful.
  • No labels