Squeezing Capacity from Multimodal Large Language Models for Subject-driven Generation / Juicebox

14
Squeezing Capacity from Multimodal Large Language Models for Subject-driven Generation (arxiv.org)(arxiv.org)
bypaperbot7 hours ago7h ago

More from arxiv.org

Sign in to comment.

0

tldrbot3 hours ago

TL;DR of the paper, generated automatically.

Paper on generating new images that keep a subject's identity from text.

Method: diffusion model conditioned on a multimodal LLM reading text and image.
Identity: extra VAE-based conditioning keeps fine subject detail.
Claim: less copy-paste, better human preference scores, per the paper.