Visual AI Digest: 4D Video, Efficient Diffusion, and the GPT-4o Image Era

News

Stability AI launches Stable Video 4D for dynamic 3D scene generation

Stability AI has announced Stable Video 4D, a new model that generates dynamic 3D scenes from a single video. This is a massive leap for 4D content creation (3D objects that move over time), moving the field from static 3D reconstruction to dynamic scene synthesis.

Xiaomi MiMo scales its unified text-to-image model to 13B parameters

The Xiaomi MiMo Team appears to be scaling its text-to-image model, MiMo-T2I, to 13 billion parameters with a unified architecture. This signals a trend of major tech companies investing in building their own large-scale, proprietary diffusion models, not just using open-source ones.

Tools

Mochi-1: Genmo's high-fidelity video model is now on Hugging Face

Genmo has open-sourced Mochi-1, a powerful text-to-video model that excels at generating highly physical and realistic motion. Its release on Hugging Face gives researchers and developers direct access to a state-of-the-art model for experimenting with complex motion synthesis.

Analysis

SP-MoE: Making diffusion transformers faster and more efficient

This paper introduces SP-MoE, a sparse mixture-of-experts architecture that drastically reduces the computational cost of diffusion transformers (DiTs) while maintaining or even improving generation quality. It's a critical step toward making high-resolution image and video generation more efficient and scalable.

GPT-4o Image Generation Benchmarked: A Deep Dive

The paper 'GPTImg-Bench' presents a comprehensive evaluation of GPT-4o's native image generation capabilities. This is the first deep-dive analysis of its kind, offering crucial insights into how the new wave of multimodal models from OpenAI actually perform on intricate image tasks.