Hallo, developed by Fudan University and collaborators, introduces a hierarchical audio-driven approach for animating portrait images, leveraging a blend of pretrained models and custom algorithms. Paper
Can language models serve as text-based world simulators?
Language models, like GPT-4, struggle to reliably simulate text-based worlds, indicating a gap between current capabilities and the potential for autonomous virtual environment creation.
NumPy 2.0.0
NumPy 2.0.0 has been released on June 16, 2024, marking a significant update to the fundamental package for array computing in Python.
OpenAI and Microsoft Azure to deprecate GPT-4 32K
OpenAI has silently removed mention of GPT-4 32K from its documentation, and Azure plans to deprecate this model in September.
[R] CFG++ : A simple fix for addressing the flaws of CFG in diffusion models
CFG++ addresses the inherent design flaws of the original classifier-free guidance (CFG) in diffusion models, offering a simpler guidance scale and enhanced invertibility.
[P] An interesting way to minimize tilted losses
Tilted empirical risk minimization offers a more equitable approach to training models by adjusting sensitivity towards challenging samples, as detailed in a JMLR paper.
[D] 1D CNN on Waveforms and Spectrograms vs. 2D CNN Performance
1D CNNs struggle with waveform inputs in audio processing tasks, often failing to converge, unlike their 2D counterparts when applied to spectrograms.