Boogu-Image-0.0-Base-exp-pt

<p align="center"> <img src="assets/boogu-logo-title.svg" alt="Boogu-Image-0.1" width="420" /> </p> <h3 align="center">Boosting Open Agentic Multimodal Generation via Understanding under a Minimal Budget</h3> <div align="center"> <img src="assets/boogu-infinity-teaser.png" alt="Boogu-Image-0.1 Teaser" width="100%" /> <!-- ============== Badges ============== --> <!-- [](https://arxiv.org/abs/{{ paper_id }}) -->

Welcome to the official repository for Boogu-Image-0.1 !

English | 中文

</div>

---

⚠️ Attention: Regarding the -pt (Prompt Tuning) Version

Please note that the Prompt Tuning version (repositories with the -pt suffix) is purely an experimental release. We are sharing it solely as a resource for community research, study, and discussion; it is NOT intended to be the main release.

The model weights in the -pt repositories were trained on a small-scale, limited dataset. Consequently, its performance and generation quality are expected to differ from the main version.

Note: Models with the Prompt Tuning (-pt) suffix only support image generation or single-image editing at up to 1K resolution.

⚠️ Important Notice

The Boogu team does NOT currently provide any paid API, subscription, or commercial service for Boogu-Image. Any paid product or service offered under the name "Boogu-Image" — or any similar / variant name such as booguimage, Boogu Image, Boogu, etc. — is NOT affiliated with this project and is unofficial. Please verify carefully before making any payment, and stay vigilant to protect your personal privacy and financial safety.

Boogu-Image-0.1 is a research project only, and not an official model release.

📖 Introduction

Boogu-Image-0.1 is a competitive Apache-2.0 open-source unified image generation and editing model family, including Base, Turbo, Edit, and Edit-Turbo, and other variants that provide stable, practical capabilities for high-quality text-to-image generation, fast generation, image editing, and Chinese-English text rendering. Closed-source multimodal understanding and generation systems like Nano Banana Pro and GPT-Image-2 achieve remarkable performance not because of a single model, but through a highly unified suite of system capabilities. However, under training compute that is extremely limited compared with closed-source systems, we find that systematically improving a model's understanding ability, data quality, and training pipeline can still significantly improve image generation and editing performance. Specifically, compared with some existing open-source models, our training data scale is roughly one order of magnitude smaller. We hope our empirical study and open-source release will help advance the open-source ecosystem for multimodal generation and understanding.

This repository provides checkpoints and inference code for Boogu-Image-0.1.

📣 News

• 2026-07-23 🔥 NPU support is now available! Check out the npu branch for initial NPU backend support and instructions. We welcome feedback and bug reports!
• 2026-07-22 🔥 vLLM-Omni now supports Boogu-Image! You can now run Boogu-Image models with the vLLM-Omni inference server. See the official recipe for details and setup: vllm-project/vllm-omni Boogu-Image recipe.
• 2026-07-16 🔥 🚀 🚀 The Boogu-Image-0.1 Technical Report is released! We have released our technical report. If you find our series of models or the report helpful to your research or applications, please consider citing our paper. Thank you very much for your support! You can read the paper here: arXiv:2607.13125.
• 2026-07-08 🔥 Boogu-Image-0.1-Edit-Turbo (Image-to-Image hotfix) is released! This update addresses several issues in the previous version, including severe image quality degradation and poor performance on removal and other editing tasks. The new checkpoints are released on Hugging Face in the revisions hotfix-1k-20260708 and hotfix-1k5-20260708. We recommend downloading the 1K checkpoint for more stable results. Try the online demo for 1K res and online demo for 1.5K res.
• 2026-06-30 🔥 Boogu-Image-0.1-Edit-Turbo (Image-to-Image) is released! Four-step distilled variant of the editing model for fast inference. Try the online demo for 1k res and online demo for 1.5k res.
• 2026-06-25 🔥 Boogu-Image-0.1-Turbo-hotfix (Text-to-Image) is now online! The new checkpoint is released on Huggingface in the revision hotfix-20260625. This is a minor patch release. We fixed visual artifacts appear in different aspect ratio, background overfitting artifacts, and other artifacts. Model weights are updated, no feature changes.
• 2026-06-17 🔥 ComfyUI-Boogu powered by ComfyUI is released! Thank you, ComfyUI!
• 2026-06-17 🔥 ComfyUI-Boogu is released!
• 2026-06-16 🔥 Boogu-Image-0.1-Base (Text-to-Image) is released! The core text-to-image foundation model. Try the online demo.
• 2026-06-16 🎨 Boogu-Image-0.1-Edit (Image-to-Image) is released! Image editing and transformation capabilities now available. The model supports resolutions up to 2K, but the results are more stable at 1K. Try the online demo. Only support 1 reference image for now. Will try our best to support more reference images. Stay tuned! Boogu-Image-0.1-Edit on single-image editing is strong. More failure cases are welcome.
• 2026-06-16 🚀 Boogu-Image-0.1-Turbo is released! Four-step distilled variant for fast inference and photorealistic generation. Try the online demo.

<!-- - **[{{ 2026-06-DD }}]** 📄 **Technical report is released!** Read our findings on [arXiv](https://arxiv.org/abs/{{ paper_id }}). -->

🌐 Join the Community!

Boogu-Image is built to grow with its users. Share what you create, report issues, exchange ideas, and help shape what comes next.

<p align="center"> <strong>WeChat Group</strong> </p> <p align="center"> <img src="assets/Boogu-Image-Discussion-Group.png" alt="Boogu-Image WeChat Group QR Code" width="180"> </p>

🏆 Boogu Arena

Since we could not evaluate on LM Arena directly, we built Boogu Arena, an LM Arena-style preference evaluation. We use an LLM to generate diverse user personas, then ask each persona to produce image generation prompts, resulting in 1K+ test prompts that we will release publicly for community reproduction. The ELO leaderboard below spans leading closed- and open-source systems. We welcome teams with questions about the results to contact us so that we can work toward a more objective, fair, and reproducible evaluation.

<p align="center"> <img src="assets/boogu_elo_arena.svg" alt="Boogu Arena ELO Leaderboard" width="100%" /> </p>

✨ Highlights

• 📸 Beautiful and Precise Photography — Accurately understands photography prompts and generates high-quality images with natural lighting, coherent composition, and faithful details, preserving coherent subject, background, and spatial relationships even in complex real-world scenes
• 📝 Diverse and Stable Text Rendering — Supports a wide range of text-heavy designs — posters, stamps, documents, interfaces, brand guides, and handwritten boards — with readable structure, stable typography, and robust bilingual (Chinese/English) rendering across diverse layouts
• 🎨 Diverse and Beautiful Stylization — Handles stylized generation across miniature 3D scenes, Chinese-inspired gilded aesthetics, shining fantasy visuals, anime portraits, and mythic character art — not just style transfer, but stable, attractive, and prompt-aware creative generation
• 🖌️ Versatile Image Editing — Handles a wide spectrum of editing tasks, including object insertion, replacement and removal, attribute and material modification, background and scene replacement, and faithful style transfer across artistic looks, while keeping the source subject and composition coherent
• 🪧 Personalized Poster Design & Product Rendering — Generates personalized poster layouts and clean product visualizations with consistent branding, refined typography, and product-grade lighting and composition
• ✍️ Precise Text Editing — Enables fine-grained, in-image text editing — replacing, adding, or removing characters in both Chinese and English — and flexibly adapts fonts, weights, colors, and layouts to match different design intents
• 📊 Competitive General Performance — Demonstrates competitive performance across many scenarios and benchmarks, with the Boogu-Image-0.1 family ranking among the very top of evaluated open- and closed-source systems in Boogu Arena

📖 For the full set of practical lessons and an honest account of current limitations, see Responsible AI & Limitations below.

🔬 Scenario-wise Comparison

Beyond overall arena rankings, we break performance down by scenario across leading open-source peers. Ratings reflect our internal evaluation of typical prompts in each category.

• 📸 Photography with reliable text rendering — Boogu-Image-0.1-Turbo delivers realistic photography, while also offering solid performance on both simple and dense text rendering.
• 📝 Strong dense text rendering — Boogu-Image-0.1-Base shows competitive results on dense, layout-heavy text scenarios such as posters, documents, brand guides, and complex bilingual designs.
• 💡 Recommendation — When your workload is dominated by dense / ultra-dense text rendering needs, we recommend running Boogu-Image-0.1-Base at 2K output resolution for the best layout fidelity and character accuracy.

📥 Model Zoo

• Boogu-Image-0.1-Base: Foundation model with strong diversity and controllability — ideal for fine-tuning and downstream development. Mainly intended for ultra-dense text rendering; for photorealism, Turbo is usually the better default. Supported resolutions: 1K, 1.5K, 2K.
• Boogu-Image-0.1-Edit: Image editing and transformation variant. Supported resolutions: 1K, 1.5K, 2K.
• Boogu-Image-0.1-Turbo: Distilled variant with the same parameter count, typically requiring only 3~4 steps. Focuses on high-quality generation and photorealism while preserving bilingual text rendering and prompt adherence. Supported resolution: 1K.
• Boogu-Image-0.1-Edit-Turbo: Distilled variant with the same parameter count.

Unless otherwise specified, all model variants support the following aspect ratios: 1:1, 2:3, 3:2, 3:4, 4:3, 1:2, 2:1, 9:16, and 16:9.

🛠️ Installation

Tested environment: Python 3.10 · CUDA 12.6 · PyTorch 2.7.1

# Use a brand new conda environment
conda create -y -n boogu python=3.10
conda activate boogu
# Instal necessary dependencies
# PyTorch up to 2.11.0 with CUDA up to 12.8 is supported
# Check `requirements/<torch>_<cuda>.txt`
pip install -r requirements/torch2.7-cu126.txt
pip install -e .
python utils/get_flash_attn.py

or

bash quick_start.sh
conda activate boogu

Download Checkpoints

Download the model weights into a local models/ directory before running inference. We recommend using the official Hugging Face CLI:

pip install -U "huggingface_hub[cli]"

# Download to ./models/<model-name>
huggingface-cli download Boogu/Boogu-Image-0.1-Base --local-dir models/Boogu-Image-0.1-Base
huggingface-cli download Boogu/Boogu-Image-0.1-Turbo --local-dir models/Boogu-Image-0.1-Turbo
huggingface-cli download Boogu/Boogu-Image-0.1-Edit --local-dir models/Boogu-Image-0.1-Edit
huggingface-cli download Boogu/Boogu-Image-0.1-Edit-Turbo --local-dir models/Boogu-Image-0.1-Edit-Turbo

Example layout after download:

models/
└── Boogu-Image-0.1-Base/
    ├── model_index.json
    ├── mllm
    ├── processor
    ├── scheduler
    ├── transformer
    └── vae

Then point inference to the local path via --model models/Boogu-Image-0.1-Base.

Flash Attention

This repository provides utils/get_flash_attn.py to automatically install a compatible flash-attn wheel for your environment.

Requirements:

• Python and PyTorch with CUDA already installed
• Linux x86_64

# Auto: detect environment, download a prebuilt wheel, fallback to source build
python utils/get_flash_attn.py

# Force source compilation
python utils/get_flash_attn.py --build

The script first searches mjun0812/flash-attention-prebuild-wheels, then tries official Dao-AILab/flash-attention release wheels with both cxx11abi variants, and finally falls back to source compilation via pip install flash-attn --no-build-isolation.

🚀 Quick Start

PyTorch Native T2I Inference

# !/bin/bash

export HF_MODULES_CACHE="$(pwd)/.hf_modules_cache"

###########################################################
# Device settings (Important!)
export device="cuda:0"
export rewriter_device=$device
enable_inner_devices_manager=False
###########################################################

# Common settings
num_inference_steps=50
seed=0
text_guidance_scale=4.0
image_guidance_scale=1.0
pretrained_pipeline_name_or_path="Boogu/Boogu-Image-0.0-Base-exp-pt"
###############VLM input image size (Fixed)#################
max_vlm_input_pil_pixels=147456
max_vlm_input_pil_side_length=768
###########################################################


#####################1K resolution#########################
# Support 1K resolution only
resolution=1K
width=1024
height=1024
max_input_image_pixels=1048576
max_input_image_side_length=2048
##########################################################

experiment_name=demo_t2i_pt
use_prompt_tuning=True

instruction="A street photography shot of an elderly scavenger with a deeply weathered and wrinkled face in the center of the frame. The scene includes a trash can and a traffic light in the background. Shot on Leica camera, high photographic texture, classic street photography aesthetic, cinematic lighting, photorealistic."

case_name=t2i_oldman_en_pt_${resolution}_steps${num_inference_steps}_tg${text_guidance_scale}_ig${image_guidance_scale}_seed${seed}

mkdir -p outputs/${experiment_name}

python inference.py \
    --seed $seed \
    --pretrained_pipeline_name_or_path $pretrained_pipeline_name_or_path \
    --num_inference_steps $num_inference_steps \
    --use_prompt_tuning $use_prompt_tuning \
    --height $height \
    --width $width \
    --max_input_image_pixels $max_input_image_pixels \
    --max_input_image_side_length $max_input_image_side_length \
    --max_vlm_input_pil_pixels $max_vlm_input_pil_pixels \
    --max_vlm_input_pil_side_length $max_vlm_input_pil_side_length \
    --text_guidance_scale $text_guidance_scale \
    --image_guidance_scale $image_guidance_scale \
    --instruction "$instruction" \
    --output_image_path outputs/${experiment_name}/${case_name}.png \
    --num_images_per_instruction 1 \
    --enable_inner_devices_manager "$enable_inner_devices_manager" \
    --device "$device" \
    --rewriter_device "$rewriter_device"

echo "${instruction}" > outputs/${experiment_name}/${case_name}_prompt_text.txt

Hardware Notes

📖 For full CLI options, device setup, offload strategies, caching acceleration, Torch Compile, FP8, and batch inference details, see INFERENCE_GUIDE.md. Torch Compile note: --enable_torch_compile can occasionally produce all-black outputs on some GPUs/models. If that happens, disable it first.

⚠️ Responsible AI & Limitations

Responsible AI

Boogu-Image-0.1 is released for research purposes and is not intended for production deployment without additional safeguards. We took responsible-AI considerations into account during data curation, training, and evaluation; however the model may still produce outputs that are inaccurate, biased, or otherwise inappropriate.

Known Limitations

🌍 World Knowledge Gap

• For tasks requiring rich common sense, domain knowledge, real brands or people, famous landmarks, celebrities, products, or complex contextual understanding, Boogu still has a clear gap from strong closed-source systems
• This capability is extraordinarily expensive to measure; even Arena-style evaluation struggles to assess it fully, so existing benchmarks barely quantify this dimension and the real gap is likely larger than measured scores suggest

🖼️ Image-to-Image Consistency & In-Context Scenarios

• For editing tasks requiring strict preservation of the input subject, identity, layout, or fine details, Boogu's image-to-image consistency is still not stable enough
• Because our image-to-image capability focuses more on photography and text-generation applications, Boogu still trails Seedream 5.0 and Nano Banana Pro in some in-context generation scenarios

📝 Text Rendering Stability

• Boogu can handle many Chinese and English text scenarios, but long text, dense typography, small fonts, and complex design layouts can still produce typos, missing characters, or layout drift
• Text rendering is currently focused on Chinese and English; other languages are not specifically optimized and may degrade noticeably

🦴 Body Structure in Complex Poses

• In multi-person interaction, occlusion, exaggerated motion, or unusual viewpoints, hands, limbs, and body structure may still become unnatural or inconsistent

👤 Small Faces & Small Limbs

• Because we use the open-source FLUX.1 VAE, reconstruction loss is relatively large, so details such as small faces, small limbs, eyes, and text may still show artifacts or instability

📦 Limited Release Scope

• Due to resource constraints, engineering complexity, and release boundaries, we are not able to open-source every training and system detail
• The current open-source release aims to balance reproducibility, usability, and sustainable maintenance while providing a reliable starting point for community research and improvement

Downstream users are responsible for applying content moderation, validation, and compliance checks appropriate to their use case.

🙏 Acknowledgements

Closed-source systems such as GPT-Image, Nano Banana, and the Seedream series helped us understand the frontier capabilities and practical boundaries of unified understanding-and-generation systems. We thank the Qwen-Image, Z-Image, OmniGen2, FLUX, Lumina-Image-2.0 and broader open-source communities for the foundations they provide, and DeepSeek for strong open-source understanding models that support open-source unified multimodal systems.

📚 Cite this work

@misc{chen2026booguimage01,
      title={Boogu-Image-0.1: Boosting Open Agentic Multimodal Generation via Understanding under a Minimal Budget}, 
      author={Guoxuan Chen and Chufeng Xiao and Haoran Yang and Siyue Xie and Binxiao Huang and Ming Zhang and Cheuk Him Chau and Xinyu Fu and Yingzhao Lian and Tom S. Y. Li and Jintao Lin and Bowen Dong and Zian Qian and Yuhao Liu and Yuxuan Hu and Weikang Shi and Bin Zou and Bowen Zheng and Haoxuan Che and Chang Chen and Yuyang He and Heyang Sun and Tianyu Huang and Chong Hou Choi and Cheng Gong and Han Shi and Haoli Bai and Xihui Liu and Hongsheng Li and Qifeng Chen and Chao Huang and Rui Liu and Chenyang Lei},
      year={2026},
      eprint={2607.13125},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.13125}, 
}

📄 License

This project is released under the Apache-2.0 License.