How $400K and 208 Million Images Powered Boogu-Image-0.1 to Rival Closed‑Source Models

Huawei’s Hong Kong Leibniz Institute and six partner universities released Boogu-Image-0.1, a unified multimodal model trained on 208 million distinct images at an estimated cost of $400 k. The model, licensed under Apache 2.0, supports native 2K resolution, Ascend NPU, the vLLM‑Omni framework, ModelScope and ComfyUI.

The authors argue that text‑to‑image generation is evolving toward “Requirement‑to‑Image”, where models must understand complex intents, constraints, and multimodal context. To address this, they elevate “understanding” to a first‑class citizen, refining it across the entire system.

Stronger instruction encoder: Freezing a 1.7 B encoder and scaling it to 14 B yields monotonic quality gains without saturation. The final choice, Qwen3‑VL‑8B, balances capability and deployability.

Agentic prompt rewriting: The paper proposes a design principle that the rewrite module should act as an identity function unless the user prompt is ambiguous. Experiments show that larger rewrite models deliver larger gains (approximately logarithmic improvement from 0.8 B to 397 B parameters) and provide several‑fold advantages on counting, reasoning, and scene‑text tasks.

Model routing: A lightweight “Turbo” model and a larger “Base” model produce nearly identical outputs on simple requests, yet Turbo is ~50× cheaper to run. Boogu estimates request difficulty with an agent and dynamically dispatches heavy models only for truly complex tasks.

In a self‑built Arena blind test (Spearman = 1.0, Pearson = 0.986 against public LMArena rankings), Boogu‑Image‑0.1‑Turbo‑Thinking achieved an Elo of 1048, ranking first among open‑source models and second only to GPT‑Image‑2 and Nano‑Banana‑Pro. On the newly released Qwen‑Image‑Bench, the Base‑Thinking variant attained the highest total score among open‑source models in both Chinese and English. The Edit‑Thinking variant topped ImgEdit‑Bench with a score of 4.64.

Key data‑efficiency findings include:

Training on 1.87 billion open‑source images quickly hits a performance ceiling.

Fine‑tuning on a curated “Boogu Syllabus” of 21.62 million human‑prior‑organized samples substantially lifts performance, demonstrating that data quality outweighs sheer quantity.

Roughly 300 exposures per Chinese character are needed for stable rendering; covering 3,500 common characters satisfies everyday Chinese generation.

About 4,500 language‑image pair exposures are required for a model to reliably reproduce a previously unseen person’s identity.

The report also advises against indiscriminately filtering “bad data” such as watermarked, over‑exposed, or motion‑blurred images; instead, explicitly describing these defects in captions enables the model to learn controlled handling of such visual artifacts.

Regarding reinforcement learning, heavy aesthetic RL collapses output diversity, causing the model to generate idealized depictions (e.g., overly polished portraits of a 60‑year‑old Chinese woman). Boogu therefore incorporates human preference into the understanding system rather than the generator, applying RL only to aspects like pose and text rendering that do not harm diversity.

Additional technical contributions include a simple truncation correction for the standard dynamic time‑shift schedule at 2K resolution, which mitigates “over‑compression” and speeds convergence, and the Boosted Orthogonal Guidance (BOG) inference enhancement that treats DiT predictions as 2‑D matrices for structured normalization, modestly improving photographic quality.

Limitations are acknowledged: world knowledge lags behind leading closed‑source systems, text rendering is limited to Chinese and English, and complex multi‑person interactions can still produce skeletal distortions. Nonetheless, under a $400 k budget the authors provide a reproducible, verifiable pathway and openly share the engineering insights that are usually hidden inside industrial teams.

For further details, see the arXiv paper (https://arxiv.org/abs/2607.13125), the HuggingFace paper page (https://huggingface.co/papers/2607.13125), and the code repository (https://github.com/Boogu-Project/Boogu-Image). An online demo is available at https://boogu.org/.