Visual AI Digest: Scaling Models and Practical 4D Video

Paper

Scaling text-to-image generation with MiMo

This multimodal hierarchical diffusion transformer achieves state-of-the-art performance on text-to-image benchmarks. The focus on a hierarchical architecture suggests a push toward more efficient and scalable models.

Fast and high-fidelity text-to-video diffusion with progressive distillation

Researchers have developed a diffusion model that generates high-fidelity text-to-video content. The focus on progressive distillation points to a method for accelerating inference, making high-quality video generation more practical.

Achieving character consistency in text-to-video

The research introduces a method for achieving character consistency across frames in AI-generated videos. This addresses a key pain point in content creation, making it easier to produce videos with coherent characters.

News

Stability AI launches Stable Video 4D

Stability AI has launched Stable Video 4D, a generative model for creating dynamic 3D videos from a single image. This signifies a major step towards practical applications of AI-generated content in areas like gaming and immersive media.

Tools

Genmo releases Mochi-1, a 10B-parameter video model

Genmo has open-sourced Mochi-1, a high-performance, 10B parameter video generation model. The release on Hugging Face provides the community with a powerful new tool for creating high-quality, coherent videos from text prompts.

Analysis

Training diffusion models on unpaired data

This paper proposes a novel framework for training diffusion models on unpaired data, a significant challenge in visual generative tasks. The approach could unlock new training paradigms and improve model robustness.