Boogu-Image-0.1: Unified Multimodal Generation

• Boogu-Image-0.1 is a unified multimodal system comprising Base, Turbo, Edit, and Edit-Turbo variants that support text-to-image generation and text+image editing with bilingual Chinese-English rendering.
• It leverages a Diffusion Transformer, FLUX.1 VAE, and agentic prompt rewriting to achieve high-resolution outputs, dynamic time shifting, and optimized text rendering under dense layouts.
• The framework balances quality and speed by employing classifier-free guidance, curated datasets, and variant routing to handle tasks ranging from fast general prompts to complex dense-text edits.

Searching arXiv for the Boogu-Image-0.1 paper and closely related entries to ground the article in the current record. Boogu-Image-0.1 is an open-source unified multimodal understanding and generation model family comprising Base, Turbo, Edit, and Edit-Turbo variants. It is designed for text-to-image generation, fast inference, instruction-based editing, and bilingual Chinese-English text rendering, and it couples an upstream Instruction Reasoner with a frozen Instruction Encoder and a Diffusion Transformer in latent space. The system supports both text-only generation and text+image editing, outputs images decoded through the FLUX.1 VAE at up to native 2K resolution, and is reported to achieve results approaching leading closed-source systems while using only 208.62 million unique images; the base model’s theoretical training cost is approximately $400K (Chen et al., 14 Jul 2026).

1. Model family and task scope

Boogu-Image-0.1-Base is intended for high-fidelity text-to-image generation at native 2K resolution, strong compositional understanding, and superior dense text rendering in Chinese and English. Its reported strengths are balanced quality, aesthetics, and alignment, together with better typography under dense layouts. The approximate parameter count is 10,292.56M trainable for the DiT only, and an optional Prompt Tuning Transformer brings the total trainable count to 11,022.55M. It is trained with Max Instruction Tokens = 944 at 2K resolution, 512 tokens at 1024, and 256 at 512; supported output resolutions are 512×512, 1024×1024, and 2048×2048, with Max Output Pixels and Max Input Pixels = 4,194,304 and Max Side Length = 4096 for reference-image inputs.

Boogu-Image-0.1-Turbo is the fast text-to-image variant. It is recommended for short-to-medium text rendering and general prompts, supports up to 2K output, and is built on the same DiT family with a 10B core but deployed for lower inference budgets. The reported latency difference is substantial: inference is often 50–100× faster than Base depending on sampling budget. The paper recommends Turbo for non-dense text and Base for dense text at 2K.

Boogu-Image-0.1-Edit is the instruction-based image editing variant for text-and-image-to-image workflows. Its editing coverage includes addition, removal, replacement, background changes, style, hybrid edits, and action-like transformations, and its reported strengths include strong scene-text edits such as replace, translate, re-organize, and stylize. Training uses mixed-resolution and mixed-task TI2I with 11.29M T2I and 10.70M editing samples. Reference images are processed through the VAE encoder; a 384×384 thumbnail is sent to the Instruction Encoder’s ViT, while the original reference image goes through the DiT image stream in latent space.

Boogu-Image-0.1-Edit-Turbo, referred to in results as Edit-Thinking, is the fast editing variant with agentic prompt rewriting before edit conditioning. It is reported as the best overall on ImgEdit-Bench among the compared models, with robustness on Remove, Hybrid, and Action tasks and improved Extract via reasoning-aware prompt rewriting.

In Boogu, “unified multimodal understanding and generation” refers to a single system covering text-only generation and text+image editing. The Instruction Reasoner chooses a mode based on the presence of a reference image. Inputs are bilingual user text instructions, optionally with one or more reference images, and outputs are generated or edited images. The paper characterizes the coupling between agentic comprehension and fused conditional generation as “Requirement-to-Image,” where the Instruction Reasoner parses, deconstructs, rewrites, sanitizes, and routes the request, the Instruction Encoder produces hidden-state conditions, and the DiT performs dual-stream and then single-stream fusion to generate the final image (Chen et al., 14 Jul 2026).

2. Architecture and optimization objectives

The core generator is a Diffusion Transformer in latent space with the FLUX.1-dev VAE. Reported architectural parameters are latent channels 16, patch size 2, hidden size 3360, attention heads 28 with KV heads 7, FFN multiple of 256, and 3D RoPE on axes dims [40, 40, 40]. The layer stack comprises 8 dual-stream layers followed by 32 single-stream layers. In the dual-stream design, the instruction stream uses holistic multimodal attention with text residual emphasis, while the image stream uses independent self-attention over image tokens plus fusion with multimodal outputs. Multiple independent QKV projections are used for the text modality, the image modality in holistic enhancement, and image self-attention. Sandwich norms and modulation are used to stabilize and control signal flow, with the main residual path ensuring identity fallback and branch paths providing heavier attention and modulation computation. The single-stream layers unify the text and image streams with modality-agnostic processing and shared parameters across token types.

The instruction pathway has three components. The Instruction Reasoner is an agent that uses a moderately large VLM, for example a 32B model, to rewrite and structure instructions, route model variants, and apply reflection or best-of-N. The Instruction Encoder is Qwen3-VL-8B-Instruct in frozen form; it processes text tokens and 384×384 thumbnails of reference images and outputs hidden-state conditions for the DiT. An optional Prompt Tuning Transformer adds 3 lightweight layers, 32 trainable prompt tokens, and a causal mask to globally modulate encoder hidden states so that they better match the DiT semantic space.

Reference images are encoded by the FLUX.1 VAE encoder into latent space and patchified. Separate Embedders and 2-layer lightweight Refiners exist for instruction, reference-image, and noise streams. Instruction hidden states from the encoder’s last layer are refined and injected into the instruction stream, while latent noise tokens are embedded, refined, and injected into the image stream. After the 8 dual-stream layers, the streams are concatenated and processed by the 32 single-stream layers, after which the VAE decoder reconstructs pixels.

The paper frames generation through Flow Matching and the Probability Flow ODE perspective. With $x \sim p_{\mathrm{data}}(x)$ and $x_t = \alpha(t)x + \sigma(t)\epsilon$, where $\epsilon \sim \mathcal{N}(0, I)$, $\sigma(1)=0$, and $\sigma(0)=\sigma_{\max}$, the PF-ODE is written as

$$\mathrm{d}x_t = -\sigma(t)\,\nabla_{x_t}\log p_t(x_t)\,\mathrm{d}\sigma(t) = -\dot{\sigma}(t)\,\sigma(t)\,\nabla_{x_t}\log p_t(x_t)\,\mathrm{d}t.$$

The DiT $D_\theta(x_t, t, c)$ approximates the intractable score, where condition $c$ includes text instruction or unconditional variants for CFG-style guidance. Sampling uses standard first-order Euler or fast ODE solvers such as DPM-Solver++, and the scheduler maps iteration index $i$ to diffusion time $t$.

A distinctive training component is Dynamic Logit-Normal timestep sampling with resolution-dependent time shifting. Base sampling is

$x_t = \alpha(t)x + \sigma(t)\epsilon$0

The linear shift factor is

$x_t = \alpha(t)x + \sigma(t)\epsilon$1

The shifted timestep is

$x_t = \alpha(t)x + \sigma(t)\epsilon$2

Its logit-space equivalence is

$x_t = \alpha(t)x + \sigma(t)\epsilon$3

The paper further introduces Rectified Dynamic Time Shifting, which clamps $x_t = \alpha(t)x + \sigma(t)\epsilon$4 beyond the 1K regime with $x_t = \alpha(t)x + \sigma(t)\epsilon$5, for example $x_t = \alpha(t)x + \sigma(t)\epsilon$6, to prevent over-squeezing at 2K; this is reported to improve early-to-mid training convergence and reduce noisy generations.

Classifier-Free Guidance is used at inference in the standard form

$x_t = \alpha(t)x + \sigma(t)\epsilon$7

where the unconditional condition $x_t = \alpha(t)x + \sigma(t)\epsilon$8 comes from drop-conditioning or negative prompts (Chen et al., 14 Jul 2026).

3. Data curation, bilingual rendering, and training pipeline

The full pipeline uses 208.62 million unique images. Initial open-source corpora total 187M and include COYO, DataComp, PixelProse, BLIP-3o, and large interleaved text-image documents from OmniCorpus. The paper states that massive open-source data alone has a strict ceiling because of semantic noise, duplication, shallow captions, and uneven coverage, and that this is insufficient for frontier performance.

To address this, the work introduces the Boogu Syllabus, a curated dataset with 21.62M unique samples, upsampled to 47.19M by weighting. It is designed with human priors for comprehensive coverage and balanced long-tail capabilities. Its macro and micro distribution emphasizes complex tasks such as graphic design, typography, poster and logo design, composition, spatial relations, counting, attributes, style descriptors, and related long-tail behaviors. Caption construction uses per-aspect VLM benchmarking over Qwen2.5-VL-7B/32B, Qwen3-VL-8B, InternVL3-8B, and Gemini-2.5/3 with aspect-specific system prompts, together with expert annotations for difficult dimensions. These captions describe salient objects, fine-grained attributes, spatial layout, counts, style, and text. Images with flaws such as watermarks, noise, blur, and exposure issues are retained, with captions explicitly labeling defects for controllable generation and robustness.

The TI2I dataset begins with 2.50M constructed editing samples, is augmented to 10.70M, and is mixed with 11.29M T2I samples to form an approximately 22M mixed dataset for TI2I fine-tuning. For edits, the modality switch is determined by the presence of a reference image, which triggers editing mode and a specific editing system prompt. The Instruction Reasoner rewrites the edit request, sanitizes it, decomposes it into explicit sub-instructions such as replacement strings or region specifications, and resolves ambiguity. The same flow-matching and transport strategies as T2I are used, including rectified dynamics at 2K. The reference image dropout probability is 0.01, which the paper states helps generalization.

A central bilingual text-rendering result is explicitly data-centric. The system systematically enumerates Chinese vocabulary and ensures per-character exposure. The key empirical finding is that a minimum of approximately 300 training exposures per glyph are required for reliable character rendering, while unseen characters are virtually unrenderable. Under resource constraints, the paper recommends prioritizing the 3,500 most frequently used modern Chinese characters and ensuring that each appears at least 300 times in training. Character-level supervision is achieved through curated synthetic SFT datasets that scale rare-character exposures, improving reliability and OCR-level correctness. Fonts and glyph handling are not specified; the emphasis is on exposure coverage, per-character frequency control, and bilingual prompt design.

The training schedule is resolution staged. At 512×512 T2I, the setup uses Open-Source 187M, Max Output/Input Pixels 262,144, Max Side Length 800, Max Instruction Tokens 256, Global Batch 1280, Micro Batch 10, learning rate $x_t = \alpha(t)x + \sigma(t)\epsilon$9, Warmup 1200, and a constant-with-warmup scheduler. At 1024×1024 T2I, the setup uses Open-Source 60M plus Boogu Syllabus 47M, Max Output/Input Pixels 1,048,576, Max Side Length 2048, Max Instruction Tokens 512, Global Batch 2048, Micro Batch 8, learning rate $\epsilon \sim \mathcal{N}(0, I)$0, and Warmup 800. At 2048×2048 T2I, the setup uses Boogu Syllabus 47M, Max Output/Input Pixels 4,194,304, Max Side Length 4096, Max Instruction Tokens 944, Global Batch 1024, Micro Batch 2, learning rate $\epsilon \sim \mathcal{N}(0, I)$1, and Warmup 800. At 2048×2048 TI2I, the setup uses 11M T2I plus 11M TI2I, Max Instruction Tokens 944, Global Batch 448, Micro Batch 1, learning rate $\epsilon \sim \mathcal{N}(0, I)$2, Warmup 800, and Ref. Image Dropout 0.01.

Optimization uses Adam with $\epsilon \sim \mathcal{N}(0, I)$3, $\epsilon \sim \mathcal{N}(0, I)$4, weight decay $\epsilon \sim \mathcal{N}(0, I)$5, Max Grad Norm $\epsilon \sim \mathcal{N}(0, I)$6, and constant learning rate with warmup from initial $\epsilon \sim \mathcal{N}(0, I)$7. The SNR strategy is Dynamic Logit Normal, Dynamic Time Shifting is enabled, and Rectified Dynamic Logit Normal is used at 2K. Hardware and precision details include HYBRID_SHARD_ZERO2 at 512 and 1024, HYBRID_SHARD at 2K, bf16 mixed precision, and gradient checkpointing set to True. The theoretical training cost for the base model is approximately $400K (Chen et al., 14 Jul 2026).

4. Agentic inference-time scaling and variant routing

Boogu-Image-0.1 places substantial emphasis on system-level integration at inference time. The agentic strategy includes prompt rewriting with a strong VLM to refine intent, resolve ambiguity, and add explicit constraints such as exact strings for text rendering; the conclusion highlights DeepSeek-V4-Flash in this role. The agent also estimates difficulty and dispatches the request to Turbo, Base, or Edit variants, thereby trading off quality and latency. Reflection and best-of-N sampling are also part of the system, with reranking used to improve prompt alignment and aesthetics at increased compute cost.

The paper states that these procedures produce clear gains in human preference in Boogu Arena and in benchmarks such as Qwen-Image-Bench, particularly in text rendering and creativity for “thinking” variants. The quality-versus-latency relationship is presented as a continuous trade-off curve rather than by per-step numbers. The paper also reports that closed-source GPT-Image-2 can be approximately 100× slower than Z-Image-Turbo for simple prompts, and that Boogu’s Base versus Turbo can differ by 50–100× for similar outputs when heavy inference strategies are applied.

Base and Turbo share the DiT family but are used differently. Base uses native 2K sampling, longer instruction tokens, higher-resolution latent representations, and heavier sampling budgets, and is identified as the best option for dense text. Turbo uses fewer steps and reduced budgets and is described as excellent for general prompts and short-to-medium text. Edit and Edit-Thinking follow the same distinction in the editing domain: Edit is the standard instruction-conditioned editing model, while Edit-Thinking adds agentic rewriting and is recommended for complex edits, especially extract and localized operations.

The paper also describes an optional inference method called Boosted Orthogonal Guidance. BOG applies matrix-structured normalization of DiT per-step predictions, a parallel-orthogonal decomposition relative to the conditional prediction, and rolling-sum momentum. The normalization is expressed as

$\epsilon \sim \mathcal{N}(0, I)$8

and, if set to all ones, as the optimal semi-orthogonal approximation

$\epsilon \sim \mathcal{N}(0, I)$9

The rolling-sum momentum is

$\sigma(1)=0$0

and the final per-step update is

$\sigma(1)=0$1

The paper reports that BOG improves cinematic textures and micro-structures while mitigating oversaturation, but also notes a practical restriction: using BOG every other step with $\sigma(1)=0$2 reduces structural distortions and helps maintain text rendering reliability. This suggests that BOG is intended as a conditional enhancement rather than a default setting (Chen et al., 14 Jul 2026).

5. Evaluation, benchmarks, and ablation findings

The empirical evaluation spans Boogu Arena, Qwen-Image-Bench, LongText-Bench, GenEval, DPG-Bench, ImgEdit-Bench, and multiple ablations. Boogu Arena is an in-house, LMArena-style evaluation with 1,200 bilingual prompts across photorealistic or cinematic content, text rendering, and stylized art, stratified by prompt length and user personas. Random seed is assigned per model and prompt, judgments are collected as blind pairwise battles, and aggregation uses Bradley-Terry Elo. The paper reports that Boogu-Image-0.1-Turbo and Turbo-Thinking dominate the open-source tier, trailing only GPT-Image-2 and Nano-Banana-Pro, and that agreement with LMArena is extremely high, with Pearson $\sigma(1)=0$3 and Spearman $\sigma(1)=0$4.

On Qwen-Image-Bench with Chinese prompts, the Overall scores are 53.57 for Boogu-Image-0.1-Base-Thinking, 53.13 for Turbo-Thinking, 51.53 for Turbo, and 50.96 for Base. The Chinese Base-Thinking breakdown is Quality 50.58, Aesthetics 55.20, Alignment 55.99, Realism 47.35, and Creativity 56.74. Open-source baselines listed are Qwen-Image-2512 at 52.06, HunyuanImage-3.0 at 50.81, and GLM-Image at 48.19, while the closed-source frontier includes GPT-Image-2 at 64.69, Nano-Banana-2.0 at 59.82, and Nano-Banana-Pro at 59.45. On English prompts, Overall scores are 53.73 for Base-Thinking, 53.58 for Turbo-Thinking, 51.61 for Turbo, and 51.00 for Base. The English Base-Thinking breakdown is Quality 51.52, Aesthetics 55.89, Alignment 55.58, Realism 47.23, and Creativity 56.24. The paper notes that thinking variants show the largest gains in Creativity, with +8 points in Chinese and a similar trend in English, and describes the system as robustly bilingual.

LongText-Bench emphasizes OCR-level correctness, primarily on shorter sequences of fewer than 100 characters. Reported averages are 0.971 for Boogu-Image-0.1-Turbo-Thinking, 0.961 for Turbo, and 0.961 for Base. The Turbo-Thinking breakdown is EN 0.957 and ZH 0.985, which places it second overall behind Seedream-4.5 at 0.988. The paper also makes an important distinction: despite similar OCR scores, the Base model exhibits more coherent typography on ultra-dense layouts at 2K, whereas Turbo can show artifacts in dense regions. A targeted SFT optimization on 3,500 frequent Chinese characters increases LongTextBench (ZH) from 0.9055 to 0.9538.

For compositional generation, GenEval reports 0.85 Overall for Boogu-Image-0.1-Base and 0.84 for Turbo. Base achieves Two Obj. 0.95, Counting 0.80, Position 0.85, Attribute 0.68, Single Obj. 0.99, and Colors 0.84; Turbo is reported as strong on Single Obj. at 1.00 and Position at 0.86. On DPG-Bench, Turbo records Overall 88.35 with Global 88.54, Entity 91.67, Attribute 92.19, Relation 93.20, and Other 93.83, while Base records Overall 87.13. The paper explicitly notes that gaps versus human preference still exist.

On ImgEdit-Bench, Boogu-Image-0.1-Edit-Thinking reports Overall 4.64 with Add 4.59, Adjust 4.64, Extract 4.32, Replace 4.69, Remove 4.85, BG 4.60, Style 4.94, Hybrid 4.26, and Action 4.83. Boogu-Image-0.1-Edit reports 4.51, and the Extract gap is singled out as evidence for the benefit of reasoning. Open-source baselines are JoyAI-Image-Edit at 4.57, FireRed-Image-Edit at 4.56, and Qwen-Image-Edit-2511 at 4.51. Closed-source entries listed are Seedream-5.0-Lite at 4.42 and Nano-Banana-Pro at 4.37. The paper also cautions that VLM-based scoring compresses true quality gaps, and that human-side comparisons favor closed-source systems in many real scenarios.

The ablations are used to support several architectural and procedural claims. With a fixed 1B DiT and identical settings, GenEval improves monotonically as the frozen instruction encoder scales from Qwen3-1.7B at 0.6034 to Qwen3-4B at 0.6251 and Qwen3-14B at 0.6477. Larger thinking models improve the advantage score of the prompt rewriter, and skill-specific PE modules for Counting, Infographic, Reasoning, NSFW, and Scene-Text improve their respective dimensions. The advantage score is defined as

$\sigma(1)=0$5

The “translator principle” is that when a prompt is clear, the rewriter leaves it unchanged as an identity lower bound. Under identical model and training settings, the Boogu Syllabus is reported to outperform raw open-source data substantially: on Qwen-Image-Bench CN+EN, 53.65 versus 48.45; on LongText EN, 0.952 versus 0.828; and on LongText ZH, 0.969 versus 0.869. Additional findings include persistent Western-centric cultural bias despite SFT, an estimate that memorizing an unseen human identity robustly requires approximately 4,500 language-matched exposures in SFT, a warning that heavy aesthetic RL can collapse diversity, and the observation that dynamic time shifting rectification at 2K speeds convergence and alleviates early-stage noise artifacts (Chen et al., 14 Jul 2026).

6. Usage patterns, limitations, safety, and licensing

The code repository is released at https://github.com/Boogu-Project/Boogu-Image, and the license is Apache 2.0. The released models are Base, Turbo, Edit, and Edit-Turbo or Thinking variants, with support for bilingual Chinese and English prompts.

For text-to-image generation, the high-level recommendation is to choose Turbo for speed and Base for high-fidelity 2K outputs and dense text. For complex instructions or text rendering, the paper recommends enabling “thinking” prompt rewriting through the agent wrapper. Recommended sampling for Turbo is DPM-Solver++ or Euler with 10–20 steps, moderate CFG such as 5–8, and BOG off by default. For Base, the recommendation is DPM-Solver++ with 20–30 steps, CFG tuned for alignment, and optional BOG with $\sigma(1)=0$6 for cinematic textures while monitoring text fidelity. For bilingual text rendering, the paper advises including the exact string to render, specifying the language, and structuring typographic constraints such as title, subtitle, sections, and footer. For ultra-dense text of more than 100 characters or words, it recommends Base at 2K, moderate CFG, and avoiding BOG or setting $\sigma(1)=0$7–4. Negative prompts can improve cleanliness, but high CFG is discouraged because it can oversaturate tones.

For image editing, the workflow is to provide a source image and an edit instruction and to enable agentic rewriting for complex edits such as extract or localize operations. Sampling is similar to T2I, with conservative CFG recommended for text edits and avoidance of BOG on heavy text-edit jobs. For scene-text edits, the paper recommends specifying the exact replacement string as well as placement and style; bilingual strings are supported.

The limitations are stated directly. Boogu trails leading closed-source systems in deep common-sense and domain knowledge, including styles, landmarks, products, and public figures. Text rendering is currently optimized for Chinese and English only, with other languages not guaranteed. Anatomy failures can occur in multi-person interactions, occlusions, and extreme viewpoints, often affecting hands and limbs. The FLUX.1 VAE imposes a reconstruction bottleneck that limits fine-grained details; native 2K helps but does not eliminate errors in small faces or small text. Cultural bias remains Western-centric; SFT reduces but does not remove this skew, so diversification of the pretraining distribution is still required.

Safety is implemented primarily through the agentic prompt rewriter, which sanitizes unsafe requests such as NSFW content. The identity-lower-bound design is intended to avoid unnecessary drift from clear prompts. The paper also advises caution with high guidance scales and with BOG, because both can affect structural fidelity and text correctness. Licensing is Apache 2.0, with no special usage restrictions beyond the license terms (Chen et al., 14 Jul 2026).