Skip to content

Boogu-Image-0.1: Boosting Open-Source Unified Multimodal Understanding and Generation

🕒 Published (v1): 2026-07-14 00:00 UTC · Source: HuggingFace · link

Why this paper was selected

Open-source unified multimodal model family; strong VLM agent backbone baseline

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Boogu-Image-0.1 is an open-source unified multimodal understanding and generation model family (Base, Turbo, Edit, Edit-Turbo) that treats language understanding as a first-class component by wrapping T2I inference inside an agent that rewrites prompts, routes between model variants, and applies reflection. Trained from scratch on 208.62M images at approximately $400K, it leads the open-source tier on human-preference benchmarks and approaches closed-source frontier systems.

Problem

Open-source T2I models lag behind closed-source systems (GPT-Image-2, Nano-Banana-Pro) whose internal training recipes remain undisclosed. Standard academic benchmarks (GenEval, DPG-Bench) exhibit rank inversions relative to human preference, saturation at top scores, and data contamination—making them unreliable discriminators of real capability. No systematic open study has analyzed how text-encoder strength, prompt rewriting design, caption style, and inference-time agentic orchestration interact to drive generation quality under constrained compute.

Method

Core pipeline — Requirement-to-Image Generation: 1. Instruction Encoder: Qwen3-VL-8B adopted as the text encoder after systematically confirming that stronger LLMs yield better text-encoding across scales. 2. Agentic Image Generation: At inference time, an agent (a) performs agentic prompt rewriting—treating the rewriter as a faithful translator of user intent rather than a generic enhancer; (b) routes the request to the appropriate model variant (Base vs. Turbo) based on quality/latency requirements; and (c) applies inference-time Reflection to iteratively improve outputs. This yields a Thinking variant (Turbo-Thinking) that scales inference compute for quality gains. 3. Data curation: Quality-over-quantity pipeline capping unique training images at 208.62M; ablations show targeted data quality outweighs raw scale. 4. Evaluation: Boogu Arena—an in-house human-preference arena emulating LMArena's blind pairwise protocol—uses 1,200 bilingual (ZH/EN) prompts spanning three categories (photorealistic/cinematic, text rendering, stylized art), 27 user personas across novice/intermediate/professional tiers, and Elo aggregation via a Bradley–Terry model.

Key Contributions

  • Open-source model family (Boogu-Image-0.1 Base, Turbo, Edit, Edit-Turbo) with weights, code, and training recipes under Apache 2.0.
  • Agentic inference-time scaling framework: prompt rewriting + model routing + Reflection, demonstrated to progressively raise Elo with added inference compute.
  • Systematic study showing text-encoder LLM scale directly improves T2I instruction following; adoption of Qwen3-VL-8B as encoder.
  • Boogu Arena benchmark (Pearson \(r = 0.986\), Spearman \(\rho = 1.0\) agreement with LMArena) released to the community.
  • Empirical critique of GenEval and DPG-Bench saturation, contamination, and rank-inversion relative to human preference.
  • Training recipe achieving competitive performance at \(\approx\$400\text{K}\) compute cost from scratch.

Results

  • Boogu Arena (Overall Elo): Turbo-Thinking = 1048, Turbo = 1021 — lead all open-source models; vs. GPT-Image-2 (1196), Nano-Banana-Pro (1087) closed-source.
  • Qwen-Image-Bench (Chinese): Base-Thinking overall = 53.57, best open-source; vs. Qwen-Image-2512 (52.06), HunyuanImage-3.0 (50.81). Thinking variant gains >8 points on Creativity dimension (48.62 → 56.74).
  • Qwen-Image-Bench (English): Base-Thinking overall = 53.73, best open-source; vs. HunyuanImage-3.0 (51.35), Qwen-Image-2512 (51.32).
  • LongText-Bench (AVG): Turbo-Thinking = 0.971 (2nd overall, behind Seedream-4.5 at 0.988); ZH score = 0.985 (highest open-source).
  • Boogu Arena vs. LMArena agreement: Pearson \(r = 0.986\), Spearman \(\rho = 1.0\) across six evaluated models.
  • Benchmark inversion confirmed: GPT-Image-2 (strongest by human preference) scores only mid-pack on GenEval and DPG-Bench.

Limitations

  • Turbo model introduces visible typographic artifacts in dense text layouts despite scoring competitively on LongText-Bench, revealing a gap between OCR-level metrics and perceptual quality.
  • LongText-Bench assessed on sequences <100 characters; does not capture dense-layout fidelity.
  • Benchmark evaluations do not cover all available open-source baselines due to time constraints.
  • Boogu Arena is internally constructed; while highly correlated with LMArena, it relies on in-house annotators and may not generalize identically.
  • Training limited to 208.62M images; the paper acknowledges data ceiling effects without fully resolving the optimal quantity.
  • No evaluation of video generation or fine-grained spatial reasoning beyond text rendering.

Relevance to Agentic AI / LLM Agents

The paper's Agentic Image Generation framework is a direct instance of agent-wrapped model inference: a controller interprets user intent, rewrites the input prompt, selects among model variants (tool/model routing), and applies iterative self-correction via Reflection—mirroring the tool-use and planning patterns studied in LLM agent research. It provides concrete empirical evidence that inference-time agentic orchestration (compute scaling through agent loops) yields monotonically improving output quality, a finding transferable to any generation task where understanding and execution are separable. The model-routing component also instantiates mixture-of-experts agent dispatch, an increasingly important design pattern in multi-agent LLM systems.