4D Fluids, Native 3D, 4K Latents, and Reward Modeling

NVIDIA Advances Human-Aligned Reward Modeling for Visual Control

NVIDIA's latest research focuses on aligning models with human preference beyond simple success criteria, a critical step for practical robotics control and simulation where rigid success/failure metrics are often insufficient.

Generating Native 3D Assets without Camera Conditioning

Native 3D generation remains a huge bottleneck, and this paper introduces a method to generate high-fidelity objects from single images without relying on camera conditioning or wasting capacity on the generator.

Text-to-4D: Language-Driven Fluid Simulation

Language-driven fluid simulation is notoriously difficult; this new architecture allows precise controllable control over fluid dynamics using semantic text rather than just physical parameters.

4K Facades: Compressed Latents for High-Res Diffusion

Compressing high-resolution images without sacrificing generative quality is key for scaling diffusion models. This paper presents a latent compression approach that makes training on 4K images feasible without massive compute costs.

Ensuring Subject-Consistent Storytelling in Latent Video Models

Achieving subject consistency across multiple video clips is vital for storytelling. This framework offers a robust solution for maintaining character identity without the need for expensive fine-tuning.