Open-source image model rivals closed systems on minimal budget

Boogu-Image-0.1 achieves competitive text-to-image generation and editing with only $400K training cost and 208.6 million images.

PaperIndependentcs.CVarXiv:2607.13125v2 ↗

Guoxuan Chen · Chufeng Xiao · Haoran Yang · Siyue Xie · Binxiao Huang · Ming Zhang · +27 more

Research Digest·15 July 2026·1 min read

Read the paper →

The authors introduce Boogu-Image-0.1, a family of open-source multimodal models that unify understanding and generation. By strengthening the system's understanding capability through a better multimodal encoder, agentic prompt rewriting, and inference-time scaling, the model matches or surpasses existing open-source models and approaches closed-source systems, all with a theoretical training cost of roughly $400K.

What they did

The Boogu Team developed four model variants: Base, Turbo, Edit, and Edit-Turbo, based on an autoregressive transformer architecture. They trained these models on 208.62 million unique images using a three-stage pipeline: (1) understanding pre-training to align a multimodal encoder with the language model, (2) joint understanding and generation training with curated data, and (3) agentic inference-time scaling and prompt rewriting for generation tasks. The total theoretical compute for the base model was only about $400K.

The authors evaluated the models on standard benchmarks for text-to-image generation (e.g., GenEval, T2I-CompBench), instruction-based editing (e.g., MagicBrush, EmuEdit), and bilingual text rendering (Chinese-English). They compared against open-source models like Emu3-Gen and Show-o, and closed-source systems like GPT-Image-2 and Nano-Banana-Pro.

Analysis

This paper fits into a broader trend of 'efficient scaling' in generative AI, where improved architectures and data strategies reduce the need for enormous datasets. By emphasizing understanding over raw generation size, Boogu-Image-0.1 challenges the data-centric paradigm of scaling laws. The use of agentic techniques like prompt rewriting and inference-time scaling also aligns with recent work in agentic AI, suggesting that reasoning at inference can substitute for model capacity. The full release of weights and recipes under Apache 2.0 is a significant contribution to open science, allowing others to build upon these practices.