BlendFusion - Scalable Synthetic Data Generation for Diffusion Model Training

Thejas Venkatesh, Suguna Varshini Velury · 2026

Read BlendFusion - Scalable Synthetic Data Generation for Diffusion Model Training by Thejas Venkatesh, Suguna Varshini Velury (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-09022.txt.

Description

Venkatesh and Velury propose BlendFusion, a scalable framework for making synthetic training data for diffusion models from three-dimensional scenes. Images made purely by diffusion models often show visual inconsistencies, and training on such data can feed an autophagous loop that ends in model collapse, known as Model Autophagy Disorder. BlendFusion builds FineBLEND, an image-caption dataset drawn from diverse 3D scenes, and its object-centric camera placement beats object-agnostic sampling. The open-source framework is highly configurable, letting the community build its own datasets from 3D scenes.

Catalog record

CREATORThejas Venkatesh, Suguna Varshini Velury
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37082
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.09022

Browse the full library · About the Tower of Babel