SIGGRAPH 2025 Workshop on 3D Generative AI

Event: SIGGRAPH 2025 Workshop on 3D Generative AI · Duration: 196 min · ▶ Watch on YouTube

Abstract

This workshop session on 3D Generative AI explores the evolving landscape of human-AI collaboration in creative processes. Speakers from Adobe Research, Snap, UCL, and HKUST present novel tools and frameworks that integrate generative AI into diverse applications, from video editing and personalized image generation to animated text design and data storytelling. The session highlights the critical need for predictable and controllable generative models that empower users with ownership and creative agency. Discussions delve into the challenges of defining and preserving identity, managing the trade-off between automation and customization, and fostering effective communication within collaborative workflows. The overarching theme emphasizes the potential of generative AI to augment human creativity, provided that tools are designed with user intent, ethical considerations, and a deep understanding of the iterative nature of the creative process.

Speakers

  • Mira Dontcheva — Adobe Research
  • Jackson Wang — Snap
  • Niloy Mitra — UCL / Adobe
  • Huamin Qu — HKUST
  • Maneesh Agrawala — Stanford University / Roblox
  • Apolinario — Independent Researcher
  • Organizer 1 — SIGGRAPH 2025 Workshop
  • Organizer 2 — SIGGRAPH 2025 Workshop

Talks (6)

  • 00:00:00 — Organizer 1: Welcome and Workshop Overview
    • The organizers welcome attendees, thank sponsors, explain the workshop’s goal to foster diverse communities in generative AI, and outline the schedule for talks and panel discussions.
  • 00:32:50 — Mira Dontcheva: Building AI-Driven Tools to Better the Creative Process of Video Editing
    • Mira Dontcheva discusses Adobe Research’s focus on integrating generative AI into video/audio editing workflows, presenting tools for generating podcast teasers, comparing video variations, and pre-editing interview footage, emphasizing user control and collaboration.
  • 02:09:00 — Jackson Wang: Connecting Through Generations: Identity, Relationships, and Diffusion Models
    • Jackson Wang introduces Omni-ID for holistic identity representation and MoA for subject-context disentanglement in personalized image generation, detailing their training and demonstrating applications like multi-subject generation and subject swap.
  • 03:42:00 — Niloy Mitra: Towards Iterative and Predictable Workflows for Creative Authoring
    • Niloy Mitra explores the creative process, highlighting the need for predictable and controllable generative AI tools, showcasing ControlNet for depth-based generation, Loosen Control for abstract outputs, and ModelGPT for image retouching, and emphasizing the importance of user intent and problem design for LLMs.
  • 04:21:00 — Huamin Qu: Human-AI Collaboration for Data Storytelling
    • Huamin Qu discusses human-AI collaboration in data storytelling, presenting tools like Camel for animated text design and ButterFly for music generation from text, emphasizing AI’s role in augmenting human creativity for effective communication of insights.
  • 04:55:00 — Panel Discussion: Panel Discussion: Challenges and Opportunities in 3D Generative AI
    • A panel discussion with all presenters and organizers covering the balance between control and creativity, open-source vs. proprietary models, defining identity, and the future of human-AI collaboration in generative AI.

Key Takeaways

  • Generative AI tools should prioritize user control and customization within creative workflows, moving beyond fully automated solutions.
  • Effective human-AI collaboration requires predictable and interpretable models, enabling users to understand and refine generated content.
  • Identity preservation and subject-context disentanglement are crucial for personalized generation, allowing for flexible and consistent content creation.
  • The creative process is inherently iterative, and generative AI tools should support exploration, refinement, and feedback loops, potentially leveraging procedural workflows and multi-turn interactions.
  • Ethical considerations, including data bias, intellectual property, and the societal impact of AI-generated content, must be addressed proactively in the development and deployment of these technologies.

Methods / Models / Datasets Mentioned

  • PodReels
  • Video Diff
  • ChunkyEdit
  • Omni-ID
  • MoA (Mixture-of-Attention)
  • ControlNet
  • Loosen Control
  • ModelGPT
  • Photoshop
  • Stable Diffusion
  • CLIP
  • Camel
  • ButterFly
  • GPT-3
  • Layer Diffuse
  • Icy Light
  • ComfyUI
  • VQGAN+CLIP
  • DALL-E
  • DALL-E 2
  • Midjourney
  • GLIDE
  • Few-to-Many Encoder Training
  • Omni-ID Encoder
  • Stage-wise Training
  • Auto-encoding
  • Motion Graphics Verification Language

Topics

Generative AI · Human-AI Collaboration · Creative Workflows · Personalization · Identity Preservation · Controllable Generation · Data Storytelling · Diffusion Models · Ethical AI · Open-source AI

Notes

Open for commentary — connections to other work, critiques, follow-up reading.