TLDR AI 2023-03-30

TECH LEADERS CALL FOR 6 MONTH PAUSE ON AI

A GPU

STYLEDIFFUSION: PROMPT-EMBEDDING INVERSION FOR TEXT-BASED EDITING

Researchers address limitations in pretrained diffusion models for image editing by optimizing the input of the value linear network in cross-attention layers and introducing attention regularization to maintain object-like attention maps. These improvements result in superior editing capabilities, demonstrated through extensive experimental prompt-editing on various images.

INTRODUCING GPT4ALL

GPT4All is a 7B parameter language model, trained on high-quality assistant data and accompanied by 800K data samples, training code, and model weights for open research and reproducibility. A 4-bit quantized version that can run on a CPU is also provided.

ADDING CONSISTENCY TO TEXT-TO-3D MODELS

Since the advent of powerful text-to-image models last year, people have been trying to move from 2d to 3d generation. One issue is the “multi-view consistency” problem, where if you only see part of an object, how do you infer what the other parts look like. Many naive methods will copy and paste the 2d image around the object which results in very odd images. By using a consistency injection technique, you can leverage the semantic understanding from 2d models to create fairly robust 3d models. The text to 3d printing pipeline is coming!

The authors present VIVE3D, a new technique that extends the capabilities of image-based 3D GANs to video editing. It uses a novel GAN inversion technique optimized for camera parameters and introduces new edits that show novel views of the head using optical flow-guided compositing to combine the head with the background video, resulting in high-fidelity, consistent quality face edits from various camera viewpoints.

INSTRUCT 3D-TO-3D: TEXT INSTRUCTION GUIDED 3D-TO-3D CONVERSION

Instruct 3D-to-3D, a high-quality 3D-to-3D conversion method, leverages pretrained Image-to-Image diffusion models to transform 3D scenes based on text instructions while maintaining consistency and controllability. Quantitative and qualitative evaluations demonstrate superior conversion quality compared to baseline methods, with the addition of dynamic scaling for adjustable geometry transformation intensity.

ATTENTION IS ALL YOU NEED FROM NOT BORING BY PACKY MCCORMICK

AI RISK DOES NOT EQUAL AGI RISK

MICROSOFT RESTRICTING BING ACCESS FOR CHATBOT COMPETITORS

GPT4ALL

GPT4All provides demos, data, and code to train an assistant-style LLM.

LIT-LLAMA

Lit-LLaMA is a fully open source model of LLaMA.