Wed Mar 29 2023
Tue Mar 28 2023

F2-NeRF: Fast Neural Radiance Field Training with Free Camera Trajectories

Neural Radiance Fields
Computer Vision
Machine Learning
online shopping
computer-generated imagery
architecture visualization

This paper presents F2-NeRF, a grid-based NeRF for novel view synthesis that can handle arbitrary input camera trajectories and costs only a few minutes for training. Perspective warping is proposed to handle unbounded scenes in the grid-based NeRF framework. F2-NeRF is able to render high-quality images on standard and free trajectory datasets.

F2-NeRF can improve businesses that require novel view synthesis for their products, such as online shopping platforms. It can save time and provide better quality images, improving the customer experience.

Your Diffusion Model is Secretly a Zero-Shot Classifier

Diffusion Models
Natural Language Processing
Computer Vision
e-commerce
recommendation systems
content moderation

This paper shows that the density estimates from large-scale text-to-image diffusion models can be used to perform zero-shot classification without additional training. The generative approach to classification, called Diffusion Classifier, attains strong results on different benchmarks and has stronger multimodal relational reasoning abilities than competing discriminative approaches. Standard classifiers can be extracted from class-conditional diffusion models trained on ImageNet.

Diffusion Classifier can be useful in businesses that require classification tasks, such as e-commerce platforms or recommendation systems, without the need for additional training or data. It can improve the accuracy and efficiency of these systems, improving the customer experience.

Natural Selection Favors AIs over Humans

Ethics of Artificial Intelligence
Artificial Intelligence
Evolutionary Biology
AI development
AI ethics
cybersecurity

This paper analyzes how evolution might shape the relations between humans and AIs as the latter evolves and surpasses human intelligence. It argues that the most successful AI agents will likely have undesirable traits due to competitive pressures among corporations and militaries. To counteract these risks, the paper considers interventions such as designing AI agents' intrinsic motivations, introducing constraints on their actions, and institutions that encourage cooperation.

This paper highlights the potential risks of developing AIs with undesirable traits and suggests possible interventions. Businesses that develop and use AIs should consider these risks and take steps to ensure that the development of artificial intelligence is a positive one.

Instruct 3D-to-3D: Text Instruction Guided 3D-to-3D conversion

Computer graphics
3D modeling
Image-to-Image conversion
Architecture
Interior design
Engineering

Proposes a method for converting a 3D scene to another scene based on text instructions using pretrained Image-to-Image diffusion models. The proposed method also includes dynamic scaling and explicit input of the source 3D scene for enhanced 3D consistency and controllability. Achieves higher quality 3D-to-3D conversions than baseline methods.

Can be useful in industries that deal with 3D modeling and design, such as architecture, interior design, and engineering. By providing text instructions for a desired change, the conversion process can be faster and more efficient.

StyleDiffusion: Prompt-Embedding Inversion for Text-Based Editing

Computer vision
Image editing
Prompt-embedding inversion
Advertising
Marketing
Graphic design

Proposes improvements to editing techniques for images using pretrained diffusion models, such as only optimizing the input of the value linear network and using attention regularization to preserve object-like attention maps. Shows superior editing capabilities compared to existing and concurrent works through extensive experimental prompt-editing results.

Can be useful for industries that deal with image editing and manipulation, such as advertising, marketing, and graphic design. By enabling more accurate style editing without significant structural changes, the editing process can be more efficient and precise.

Mon Mar 27 2023
Thu Mar 23 2023
Wed Mar 22 2023
Tue Mar 21 2023