Open Image Model Narrows Gap Under Tight Compute

Open image-generation systems are under pressure from two directions: closed systems are increasingly agentic, while open models are judged mostly by static text-to-image benchmarks that hide inference cost and system design. Boogu-Image-0.1 addresses that gap with an open unified model family for image generation, bilingual text rendering, and instruction-based editing. The paper’s central claim is practical rather than architectural maximalism: better understanding, cleaner data, and selective inference-time computation can move an open model closer to closed systems without frontier-scale training budgets.

Core Contribution

The genuinely useful part of Boogu-Image-0.1 is the way it treats generation as a requirement-following problem, not only a prompt-to-image mapping. The authors argue that the field is shifting from Text-to-Image toward Requirement-to-Image, where the model must interpret implicit constraints, style requirements, layout demands, and cross-modal context. Their answer is a family of variants — Base, Turbo, Edit, and Edit-Turbo — plus an agentic layer that rewrites prompts and routes requests to the cheaper or stronger variant depending on difficulty.

Figure 2 shows the intended operating range: photorealistic portraits, Chinese ink-and-gold painting, Chinese calligraphy, a 3D miniature scene, and clean English title rendering inside one released model family.

Technical Approach

Boogu-Image-0.1 builds on a diffusion-transformer-style image generator with a multimodal understanding stack around it. The appendix lists a DiT with 32 single-stream layers and 8 dual-stream layers, 3D RoPE, hidden size 3360, 28 attention heads, and 10,292.56M trainable parameters. With the optional prompt-tuning transformer, the total rises to 11,022.55M parameters. The system uses FLUX.1-dev as the VAE and Qwen3-VL-8B-Instruct as the instruction encoder.

The training pipeline is staged by resolution and task. The paper describes progressive training from low-resolution to higher-resolution images, followed by fine-tuning from text-to-image into text-and-image-to-image editing. For the T2I phase, the authors report roughly 22M samples: 11.29M pure T2I samples drawn from a 47.19M Boogu Syllabus dataset, plus 10.70M editing samples augmented from an initially constructed 2.50M T2I dataset. That sits inside the paper’s larger budget claim: 208.62M unique images and a theoretical base-model training cost of about $400K.

A second component is inference-time scaling. Instead of always running the most expensive variant, Boogu routes requests across the family. The paper explicitly notes that Base and Turbo outputs can be nearly indistinguishable in many cases while inference costs can differ by 50 to 100 times. This is a sensible engineering choice: spend latency on dense text, editing, or hard composition, but keep ordinary prompts on the faster path.

Results and Analysis

The evaluation mixes standard benchmark comparisons, Boogu Arena human comparisons, and many qualitative pages. The authors compare against open models such as Z-Image-Turbo and HiDream-O1-Image, and against closed systems including Qwen-Image variants, Seedream-5-Lite, Nano-Banana-Pro, and GPT-Image-2. The qualitative comparisons emphasize photorealistic and cinematic images, stylized art, and bilingual text rendering. The text-rendering pages are a useful signal because they stress character-level fidelity rather than only aesthetic preference.

The paper reports that Turbo handles short-to-medium text rendering up to about 100 Chinese characters or 100 English words, while Base is recommended for dense text over that range under 2K resolution. Figure 3 illustrates the paper’s strongest qualitative case: English and Chinese text rendered as posters, product images, signs, and layout-heavy designs rather than isolated words.

The evidence supports a narrower conclusion than the abstract’s broad comparison to leading closed systems. Boogu appears competitive among open models, especially when the task requires bilingual typography or instruction following, and the cost story is credible because the paper discloses model size, data scale, and training budget. The comparison to closed systems is less clean. The authors themselves point out that GPT-Image-2 can take up to 100 times longer than Z-Image-Turbo for a simple prompt, and that public leaderboards often omit inference-time cost. That makes a single quality ranking incomplete unless latency and test-time computation are fixed.

Limitations

The paper is unusually direct about its own limits. Boogu still trails leading closed systems on tasks requiring broad common-sense and domain knowledge, such as artistic styles, landmarks, public figures, and commercial products. Text rendering is optimized only for Chinese and English. Multi-person interactions, heavy occlusion, and unusual viewpoints can still distort hands and limbs. The use of the open FLUX.1 VAE also leaves reconstruction errors in fine details such as small faces, limbs, and text, with 2K generation only partly reducing the issue.