Huawei and Six Universities Release Boogu-Image-0.1: Open-Source 2K Text-to-Image Model Trained for $400K

The Boogu team believes user needs have shifted from simple descriptive prompts to complex intents, implicit constraints, and cross-modal context. To address this, they decompose 'understanding' into every system component: using a stronger instruction encoder (Qwen3-VL-8B), experiments confirm that scaling the frozen encoder from 1.7B to 14B monotonically improves generation quality without saturation; introducing Agentic prompt rewriting with identity transformation as a lower bound, intervening only when intent is ambiguous, with stronger rewriting models yielding greater gains (near log-linear improvement from 0.8B to 397B); and using model routing to dynamically distribute requests based on difficulty, reserving heavy models for complex tasks.

In their self-built Arena blind test (with Spearman correlation of 1.0 and Pearson correlation of 0.986 with the public LMArena ranking), Boogu-Image-0.1-Turbo-Thinking achieves an Elo of 1048, ranking first among open-source models, behind only GPT-Image-2 and Nano-Banana-Pro. On the newly released Qwen-Image-Bench, its Base-Thinking variant achieves the best total score among open-source models in both Chinese and English; the editing model Boogu-Image-0.1-Edit-Thinking scores the highest overall 4.64 on ImgEdit-Bench.

The team found that models trained on 187 million open-source data clearly hit a performance ceiling; however, post-training with a small amount of human-prior-organized 'Boogu Syllabus' data (only 21.62 million unique samples) comprehensively improves performance. The principle is systematic fine-grained decomposition, full capability coverage, and sufficient data volume per category—'If the model hasn't seen a cat during training, it defaults to not being able to draw a cat.'

The team also disclosed multiple reusable quantitative findings: stable rendering of a Chinese character requires ~300 exposures, and covering 3,500 modern common characters suffices for daily Chinese rendering; making the model remember an unseen person's identity requires ~4,500 language-matched exposures (using 170 daily photos of one paper author for full ablation). Additionally, 'bad data' such as watermarks, overexposure, and motion blur need not be discarded; if defects are described in detail in captions, the model can learn to controllably avoid or reproduce these visual elements.

The team compared six recent models and found a clear inversion between LMArena human preference rankings and GenEval/DPG-Bench scores: GPT-Image-2, which has the strongest human preference, ranks only mid-tier on both academic benchmarks. The team directly announced they will no longer use these benchmarks as primary evaluations and called on the community to reassess their applicability—even frankly noting that while their model scores high on ImgEdit-Bench, it still underperforms Nano-Banana-Pro in human evaluation.

Regarding reinforcement learning, the team found that heavy aesthetic RL leads to output distribution collapse; for example, when asked to draw 'a sixty-year-old Chinese woman,' the heavy RL model tends to output an idealized image with refined makeup. Boogu chooses to incorporate human preferences into the understanding system rather than bake them into the generator, using RL only on issues like limb structure and text rendering that do not harm diversity.

As one of the earliest open-source models trained at native 2K resolution, the team discovered that the standard dynamic time-shift strategy produces an 'over-squeezing' effect at 2K resolution, slowing convergence, and provided a simple and effective truncation correction. The appendix also proposes a training-free inference enhancement method called BOG (Boosted Orthogonal Guidance), which treats DiT predictions as a 2D matrix for structured normalization, strengthening information orthogonal to the flow field, thereby improving photo-realism and artistic quality to some extent.

Model weights, code, and training recipes are fully open-sourced under Apache 2.0, supporting Huawei Ascend NPU, vLLM-Omni framework, etc., and are available on ModelScope and ComfyUI.