I presented a poster at SIGGRAPH Asia 2025 in Hong Kong. The most valuable thing I did was read everyone else's.
The poster program ran December 15-18, 2025. 127 poster articles across 10 sessions, from Animation and Motion to Machine Learning. 178 accepted out of 869 submissions. 20% acceptance rate.
I spent three days going through them and cross-checking against the ACM proceedings afterward. I wasn't looking for breakthroughs. I was looking for patterns.
Three things stood out.
1. Diffusion is becoming plumbing.
Everyone knows diffusion models are everywhere. That's not interesting anymore. What caught my attention was that people stopped treating diffusion as the main contribution and started treating it as a stage you plug into a bigger system.
Zenith (Sayyad, Rothman, Zhai) couples geometry processing with diffusion to generate multi-layered top-down maps from 3D scenes. The input is World of Warcraft dungeon content. Procedural step first, diffusion as downstream refinement.
Floorplan Generation with Graph Beta Diffusion (Takeuchi, An, Yamashita) goes the opposite direction. Single-stage, end-to-end graph diffusion for residential floorplans, a simpler alternative to pipelines where errors stack up across stages.
Our own work, A Framework for Architectural Geometry Synthesis Using Cellular Automata and Conditioned Diffusion Models (Gokmen, Vince, Oh), combined a 3D cellular automata engine for voxel massing with diffusion conditioned on the geometry for higher fidelity output. Procedural rules handle structure and controllability. Diffusion handles visual and geometric richness. Separation of concerns.
Diffusion also showed up outside the Machine Learning session entirely. In Creativity and Digital Art, PolyArt is explicitly diffusion-based for multilingual movie poster generation. That's the signal. It's not just ML researchers using it. It's leaking into every session.
The pattern: diffusion is becoming a reusable component. Bolt it onto a floorplan graph, a voxel massing, a processed game level. The interesting work isn't the diffusion model itself. It's the contract between the structure and the diffusion stage.
2. Physics-informed ML is where the real edge is.
The best simulation posters weren't replacing physics with ML. They were combining them.
TryIto (Oppenheimer, Marino) is a hybrid cloth simulation framework. Classical physics simulation for the backbone. ML to learn the residual, everything the physics model misses. They didn't train a network to approximate the whole simulation. They let physics do what physics does well and used learning to close the gap.
Simulating 3D Thermal Fluid Dynamics in Data Centers (Wang, Ng, See, Wang, Guan) applied a physics-informed neural network to 3D thermal fluid simulation, a domain that normally requires expensive simulation software. Same idea: keep the model based on real physics, cut the computational cost.
If you build anything that touches simulation, rendering, robotics, or infrastructure, pay attention to this. Pure ML won't replace physics anytime soon. Hybrids will.
3. One poster pointed at something bigger.
I'll be honest. Two posters don't make a trend. But sometimes one example is enough to make you think about where things are going.
AniME (Zhang, Xu, Yang, Yin, Liu, Xu, et al.) is a director-oriented multi-agent system for automated long-form animation generation. A director agent maintains global memory and coordinates specialized sub-agents across the workflow, from story to final video. Not prompt engineering. A stateful, coordinated system that maintains context across steps.
The Theater Machine (Gallist, Abasolo, Pham, Hagler) took a different angle. An interactive generative AI installation in a theater lobby, where the audience interacts with the content through a camera.
Neither is a breakthrough on its own. But they point in the same direction: AI in creative contexts moving from "generate an output" to "orchestrate a production." Single-shot generation to systems that maintain state over time. That shift hasn't fully landed yet. But the early work is showing up.
What I took away from all of it.
The "Machine Learning" session label means less than you'd think. Diffusion appeared in Animation, in Creativity and Digital Art, and in ML. The boundaries between sessions are blurring because the methods are becoming shared infrastructure.
Multiple posters were framed as pipelines, not isolated models. The interesting contribution wasn't the model. It was how stages talk to each other, what assumptions they share, and where they break.
The most convincing results came from hybrid architectures. Physics plus residual learning. Procedural geometry plus diffusion. Stateful agent coordination. Not pure ML. Not pure classical methods. The combination.
The conference theme was "Generative Renaissance." It captured something real. Generative models aren't a separate research area anymore. They're treated like infrastructure components.
If you're building tools, pipelines, or products, the practical question stopped being "will ML be involved." It's now about what interfaces, constraints, and checks make an ML component actually safe and useful inside a real system.
Andrei Vince is a Stanford Design Challenge finalist (top 8 of 249 teams, 33 countries), co-author of research presented at SIGGRAPH Asia 2025 in Hong Kong, and a Software Engineer building AWS infrastructure that processes 30,000+ insurance documents at SuperKey Insurance.
Top comments (0)