Academy Software Foundation Technical Advisory Council (TAC) Meeting - June 24, 2025

Join the meeting at https://zoom-lfx.platform.linuxfoundation.org/meeting/97880950229?password=81d2940e-c055-43b9-9b5a-6cd7d7090feb

Voting Representative Attendees

Premier Member Representatives

Project Representatives

Carol Payne - Diversity & Inclusion Working Group Representative, OpenColorIO Representative
Cary Phillips - OpenEXR Representative
Chris Kulla - Open Shading Language Representative
Diego Tavares Da Silva - OpenCue Representative
Jonathan Stone - MaterialX Representative
Ken Museth - OpenVDB Representative
Nick Porcino - Universal Scene Description Working Group Representative
Rachel Rose - Diversity & Inclusion Working Group Representative

Industry Representatives

Jean-Francois Panisset - Visual Effects Society

Non-Voting Attendees

Non-Voting Project and Working Group Representatives

LF Staff

David Morin - Academy Software Foundation
Emily Olin - Academy Software Foundation
John Mertic - The Linux Foundation
Michelle Roth - The Linux Foundation
Yarille Ortiz - The Linux Foundation

Other Attendees

John McCarten - Weta / RMTC
Jim Helman - MovieLabs / Zero Trust WG
Doug Walker - Autodesk / OCIO
John Paul Smith - BorisFX
JT Nelson - Pasadena Open Source consortium / SoCal Blender group
Spencer Stephens - MovieLabs / Zero Trust WG
Stephen Revel
Bill Ballew - Dreamworks
Cory Ormand - Walt Disney Studios
Cottalango Leon - Sony Imageworks / OpenCue
Lorna Dumba - Framestore

Antitrust Policy Notice

Linux Foundation meetings involve participation by industry competitors, and it is the intention of the Linux Foundation to conduct all of its activities in accordance with applicable antitrust and competition laws. It is therefore extremely important that attendees adhere to meeting agendas, and be aware of, and not participate in, any activities that are prohibited under applicable US state, federal or foreign antitrust and competition laws.

Examples of types of actions that are prohibited at Linux Foundation meetings and in connection with Linux Foundation activities are described in the Linux Foundation Antitrust Policy available at linuxfoundation.org/antitrust-policy. If you have questions about these matters, please contact your company counsel, or if you are a member of the Linux Foundation, feel free to contact Andrew Updegrove of the firm of Gesmer Updegrove LLP, which provides legal counsel to the Linux Foundation.

Agenda

General Updates
- OpenQMC #434
- Open Source Days 2025 #1005
New Project/Working Group Proposal: RMTC (Rongotai Model Train Club) #1075
Annual Review: CI Working Group #511

Notes

Are we OK recording today? Larry at Open Source North America conference
Emily: Open Source Days
- Virtual Down Halls: Aug 4-8
- Open Source Days Main Program: August 10
- Bofs: Aug 11
- 153 people registered already
- Please make sure your descriptions are “as final as they can be”
- OSD schedule
  - Strong list of CFPs, very high quality this year, amazing program committee
  - Keynote: 2 people from Flow animated movie
  - Leika will be joining foundation, keynote in the afternoon
  - ACES talk not on schedule yet
New Project/Working Group proposal: RMTC
- Presentation Slides
- John McCarten, WETA
- Rongotai Model Train Club: ML artifact management system for VFX
- Rongotai: a location in Welligton
- Placeholder name, can be changed
- Lots of challenges! Including from the audience and creative community, tooling issues, copyright, lack of standards
- 4 categories of issues
  - Legal
  - Scale
  - People
  - Tooling
  - RMTC proposes a framework to track and address these 4 issues
  - Want to make this VFX first, not research first
- Mission statement
- Phase 1: Formalize
- Phase 2: Optimize
- Phase 3: Democratize
- Where it fits in pipeline
  - Aim to formalize model to train image to image pipelines in Nuke
  - Wrap, Refine -> Visualise, Publish, Infer
  - Target Nuke, but want to work with other DCCs
- Have been working on a prototype of the model
  - How do we manage priorities from different facilities?
  - Each component designed with dependency injection in mind
- Independently usable tracking layer, which defers to storage system (database, file)
- Can integrate into your own system
- Next tier up: Primary API, Basic ML Operations, On disk products - UUID. Also hope to add a GUI implementation
- Top layer: RMTC core (Reference Implementation, PyTorch, AGE, PIP). Licensed under Apache 2
- Very top layer: facility specific extensions. Create mechanisms for publishing, training…
- System block diagram
  - Red: training strategy, scheduler, tracker -> represents a solution
    - Runs tie data sets, models and weights, product artifacts and store metrics
    - Trainer class holds hyper parameters, have a simple regression trainer for now
    - Can also just push models without having to go through a training step
  - Yellow: checkpoint, weights, model, dataset, asset, converter
    - Can do I/O to disk, artefacts
    - Can use Tensorflow model, ONNX models
    - Converter can convert standard model types
    - Checkpoints: can store optimizer and trainer state
  - Green: Store, entity
    - Can be used independently
    - Deferred storage system, can query parts
    - Can get a shell to operate
  - Blue: Inference, Inferable
    - A DCC plugin can be an implementation of an Inferable
    - Inference connects to weights and models
- Graph which represents a production example
  - Used for segmentation
- Steps to do
  - Scalability
    - Shared Service
    - Standards
    - AI Platform
  - Extensions
    - Schedulers
    - Trainers
    - Inferers
    - Artefact wrapping
  - Validation
    - API
    - Testing
    - Stakeholders
    - Design
    - Use Cases
  - Interfaces
    - Explorer
    - Jobber
    - Recipes
    - Inference
- Shared Service?
  - RMTC as a Service
    - Persistence at the DOM level
    - Job integration
    - Service stack implementation
  - Shared artefact repository?
    - Do we consider a ASWF hosted artefact repository?
    - Impacts - facility agnostic asset UUIDs
- Standards
  - Asset tensors?
    - HWC or CHW images? PyTorch or cache coherency?
    - Training agnostic tensors?
    - Safe tensors?
  - Inference time assets?
    - DCC memory mappable assets
    - Runtime converters
  - Other standards to consider?
    - C2PA provenance embedding?
- AI Platform?
  - Do we extend VFX Platform?
    - AI projects are currently “off structure”
    - Separate or as part of VFX Platform
  - Training, inference & utilities
    - PyTorch - training
    - ..
- Extensions
  - Additional training strategies
    - More than 1:1 correlation models
  - Renderwall training scheduler
    - Concurrent w/ weights fusion
    - OpenCue / Plow
  - Inference plugins for DCCs
    - Nuke
    - Maya
  - Wrapping
    - Common foundational models
    - ..
- Explorer
  - GUI for exploring artefacts
    - Track provenance
    - Datasets, models & inferences
    - Ingested model repository
  - Visual tracing
    - Licensing
    - Inference -> Weights -> Run
    - Create reports
  - Invocation system to trigger tools
    - Dataset curation
    - …
- Lots of open questions
  - How do we collaborate?
  - How to futureproof during such rapid changes?
  - Implications of scale on design?
  - Depth of tracking?
  - Do we want to consider inference tooling?
- Q&A
  - Carol: there are a couple of other efforts in this place, are you thinking of connecting, like C2PA? Kimball: C2PA is about tracking the authenticity all the way to delivery, you could track a hash through training, so for instance if an actor doesn’t want their image included. So we can eliminate portion of dataset and retrain. A large data problem, but tractable. We’ve been doing asset tracking for a long time. John: also relates to GDPR for instance.
  - Eric: may also be doing synthetic generation.
  - Kimbal: some models trained from output of other models.
  - Jim: there are also efforts to feed data from production and master workflows, some work at SMPTE to get data into MXF, DPX. Not sure if some data captured in EXR? Capturing a lot more of the metadata around provenance of the model, but also human input that went into gen AI tool. Is this something you are looking at, and what the data model would be to support all this. Have been doing some work with ETC to inject this data into ComfyUI. John: this is Inference Time data, can be modeled in Inferable class. Could have a GenAI Inferable, holds data about the prompts. That would tie in with the current schema we have. We are looking at different licenses, agreements, so can track licenses per dataset. What happens when we transform data? Each actor could have their own agreement, but what if you are doing face swaps? We have to track chain of transformations. We are investigating how to represent dynamic graph.
  - Kimbal: trying to connect the data to the model, so we can make sure a version of a model has or has not a specific actor in it. Not just for LLMs, but instead all classes of ML.
  - Alex: who might contribute and/or adopt? Kimbal: hoping everybody will! Hoping we have a framework that all studios producing images that are a blend of images from models and live, so we can keep track of provenance, and make sure we value human creations. Hopefully everyone can use it. Also we provide utilities that provide good utilities to load images from OIIO, models via USD…
  - Larry: scope is so much broader than I expected. The people who may best know how this fits into our plans may not be on this call. May need some digestion time, make sure the slides and any supportive materials are available so we can discuss internally.
  - Larry: we probably need time to “digest”. Facilities will like an opportunity to offer opinion, what components they could contribute.
  - John: am trying to limit the scope. But want to make it open ended. But we have a tight remit. Larry: what is the “MVP” through this project? Minimal proof of concept? John: the pipeline through Nuke.
  - We should be able to share the slides. Kimbal: yet can attach PDF to the issue.
CI Working Group Annual Review
- Presentation Slides
- Q&A
  - Carol: we can also make the CI WG a long term WG, “this isn’t going anywhere”. JF: maybe that’s the vote? Carol: we didn’t know if we wanted to make this a project.
  - Larry: we appreciate this work of the WG, projects would spend a lot more time working on CI.
  - Carol: move to vote to a long term WG, at a future meeting we should discuss strategies to help participation. John: vote to move to long term WG is the proposal? Carol: yes!