Sun Feb 19 2023
Thu Feb 16 2023

3D-aware Conditional Image Synthesis

Neural radiance fields
Computer vision
Image processing
3D object generation
Image synthesis
Interactive 3D editing

Pix2pix3D synthesizes 3D objects given a 2D label map, such as a segmentation or edge map.

Can be used to generate 3D objects and photorealistic images from 2D label maps, allowing for explicit 3D user control. Provides an interactive 3D editing demo.

LEVER: Learning to Verify Language-to-Code Generation with Execution

Code language models
Natural language processing
Programming languages
Language-to-code generation
Program verification
Code language models

LEVER improves language-to-code generation by learning to verify the generated programs with their execution results.

Improves language-to-code generation by combining CodeLM decoding with verifiers trained to determine whether a program is correct based on its natural language input, program, and execution results.

Efficiency 360: Efficient Vision Transformers

Transformers
Computer vision
Machine learning
Vision transformer models
Image classification
Efficiency

Efficiency 360 compares various vision transformer models based on their performance, number of parameters, and number of floating point operations on multiple datasets.

Introduces an efficient 360 framework for vision transformers to make them more efficient for industrial applications. Compares various vision transformer models based on their performance, number of parameters, and number of floating point operations on multiple datasets.

Text-driven Visual Synthesis with Latent Diffusion Prior

Diffusion models
Image Processing
Machine Learning
text-to-3D
StyleGAN adaptation
layered image editing

Presents a generic approach using latent diffusion models as powerful image priors for various visual synthesis tasks, including text-to-3D.

Provides a more efficient and effective approach to text-to-image synthesis and visual synthesis tasks, which can be useful for businesses looking to improve their image editing and customized generation processes.

Shared Microexponents: A Little Shifting Goes a Long Way

Block Data Representations
Machine Learning
large-scale generative pretraining
inferencing
production-scale recommendation systems

Presents Block Data Representations, a framework for exploring and evaluating a wide spectrum of narrow-precision formats for deep learning.

Introduces an innovative framework for deep learning that enables the comparison of different quantization standards and identifies a new format called shared microexponents, which outperforms other state-of-the-art quantization approaches. This could be useful for businesses looking to optimize their machine learning models and improve their recommendation systems.

Wed Feb 15 2023
Tue Feb 14 2023
Mon Feb 13 2023
Sun Feb 12 2023