SuperAPI AI Directory
Partner with us中文AI newsFavorites
All toolsAI imagesBoogu-Image

Boogu-Image

Boogu-Image-0.1 is an Apache-2.0 open-source image generation and editing model family, including variants such as Base, Turbo, Edit, and Edit-Turbo, supporting text-to-image, fast generation, image editing, and Chinese and English text rendering. The project is positioned for research purposes, and the official statement explicitly states that it does not provide any paid API, subscription, or commercial service. Its training data scale is approximately one-tenth that of some existing open-source models, and under limited training compute it improves generation and editing performance by enhancing model understanding capability, data quality, and the training pipeline. The repository provides checkpoints and inference code, supports Python 3.10, CUDA 12.6, and PyTorch 2.7.1 environments, and has already added support for an NPU branch and vLLM-Omni inference service. The official team has also released the technical report arXiv:2607.13125 and the Boogu Arena preference evaluation leaderboard. The applicable resolutions are 1K, 1.5K, and 2K (Turbo only 1K), supporting multiple aspect ratios. The project has known limitations in areas such as world knowledge, image-to-image consistency, text rendering stability, human body structure in complex poses, and details of small faces and small limbs, and it has not open-sourced all training and system details, making it unsuitable for direct production deployment without additional safeguards.

Open website

About this tool

Boogu-Image-0.1 is an open-source unified image generation and editing model family released under the Apache-2.0 license, which the official team says achieves performance close to closed-source systems with a training data scale approximately one-tenth that of some existing open-source models. The project is released by the Boogu team and is positioned as a research project rather than a formal model release.

Purpose and Positioning

The project is aimed at open multimodal understanding and generation research, with the goal of advancing the open-source ecosystem in the direction of unified image generation and editing. The official team emphasizes that the excellent performance of closed-source multimodal understanding and generation systems such as Nano Banana Pro and GPT-Image-2 does not come from a single model, but rather from a highly unified combination of system capabilities; Boogu hopes to improve generation and editing results by systematically enhancing model understanding capability, data quality, and the training pipeline under training compute far lower than that of closed-source systems. This repository provides checkpoints and inference code for Boogu-Image-0.1.

Model Family and Main Capabilities

Boogu-Image-0.1 includes the following variants:

  • Boogu-Image-0.1-Base: a text-to-image base model that emphasizes diversity and controllability, suitable for fine-tuning and downstream development, and mainly aimed at ultra-high-density text rendering scenarios. It supports 1K, 1.5K, and 2K resolutions.
  • Boogu-Image-0.1-Turbo: a distilled variant with the same parameter count, usually requiring only 3~4 steps, focusing on high-quality generation and photographic realism while retaining bilingual text rendering and prompt-following capabilities. It supports 1K resolution.
  • Boogu-Image-0.1-Edit: an image-to-image image editing and transformation variant that supports up to 2K resolution, but results are more stable at 1K; currently it supports only one reference image. It supports 1K, 1.5K, and 2K.
  • Boogu-Image-0.1-Edit-Turbo: a four-step distilled fast inference variant of the editing model.

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, 16:9.

The officially listed capabilities include:

  • Photographic-grade generation: accurately understands photographic prompts, generates high-quality images with natural lighting, coherent composition, and faithful details, and maintains coherent subject, background, and spatial relationships in complex real-world scenes.
  • Text rendering: supports a large number of text-dense designs such as posters, stamps, documents, interfaces, brand guidelines, and handwriting boards, with readable structure, stable typography, and robust Chinese and English bilingual rendDye.
  • Stylized generation: covers miniature 3D scenes, Chinese gilded aesthetics, dazzling fantasy visuals, anime portraits, mythological character art, etc. It is not merely style transfer, but stable, aesthetically pleasing, and prompt-faithful creative generation.
  • Image editing: supports object insertion, replacement, and removal, attribute and material modification, background and scene replacement, as well as faithful style transfer, while maintaining the coherence of the source subject and composition.
  • Personalized poster design and product rendering: generates personalized poster layouts and clean product visualizations, maintaining consistent branding, refined typography, and product-grade lighting composition.
  • Precise text editing: supports fine-grained editing of text in images, allowing replacement, addition, or deletion of Chinese and English characters, with flexible adjustment of font, weight, color, and layout.

In the officially announced Boogu Arena preference evaluation, the Boogu-Image-0.1 family ranked highly among the evaluated open-source and closed-source systems. The official recommendation: when workloads are dominated by dense or ultra-dense text rendering, it is recommended to run Boogu-Image-0.1-Base at 2K output resolution to obtain the best layout fidelity and character accuracy; if photorealistic quality is sought, Turbo is usually a better default choice.

Usage and Installation

The officially provided test environment is Python 3.10, CUDA 12.6, PyTorch 2.7.1.

Before inference, the model weights need to be downloaded to the local models/ directory; the official recommendation is to use the Hugging Face CLI. After downloading, point inference to the local path via --model models/Boogu-Image-0.1-Base.

Regarding Flash Attention: the repository provides utils/get_flash_attn.py, which can automatically install a compatible flash-attn wheel for the current environment. The script first searches mjun0812/flash-attention-prebuild-wheels, then tries the two cxx11abi variant release wheels from the official Dao-AILab/flash-attention, and finally falls back to compiling from source via pip install flash-attn --no-build-isolation.

The repository provides a PyTorch native T2I inference method. Ascend NPU inference support requires switching to the npu branch and performing setup and according to the instructions in that branch's README.md.Usage.

For complete CLI options, device settings, offloading strategies, cache acceleration, Torch Compile, FP8, and batch inference details, see INFERENCE_GUIDE.md. Official note: --enable_torch_compile may occasionally produce all-black output on some GPUs/models; if this occurs, please disable this option first.

In addition, vLLM-Omni now supports Boogu-Image, and this model can be run through the vLLM-Omni inference server. ComfyUI-Boogu has also been released.

Applicable Scenarios

  • Text-to-image generation and image editing experiments for research purposes
  • Chinese-English bilingual text rendering, text-dense designs such as posters and brand guidelines
  • Generation tasks requiring fast inference (Turbo / Edit-Turbo)
  • Base model experiments requiring fine-tuning and downstream development (Base)
  • Integration and evaluation within ecosystems such as ComfyUI and vLLM-Omni

Limitations and Considerations

The official Responsible AI & Limitations statement notes that Boogu-Image-0.1 is intended for research purposes only and is not suitable for direct production deployment without additional safeguards. Known limitations include:

  • World knowledge gaps: There is a clear gap compared with strong closed-source systems when it comes to common sense, domain knowledge, real brands or people, famous landmarks, celebrities, products, and understanding of complex contexts; measuring this capability is extremely costly, and the true gap may be greater than what the measured scores reflect.
  • Image-to-image consistency and contextual scenarios: In editing tasks where the input subject, identity, layout, or details must be strictly preserved, image-to-image consistency is still not stable enough; in some contextual generation scenarios, it still lags behind Seedream 5.0 and Nano Banana Pro.
  • Text rendering stability: Long text, dense typography, small font sizes, and complex design layouts may still exhibit typos, missing characters, or layout drift; text rendering currently focuses on Chinese and English, and other languages have not been specifically optimized, so results may decline noticeably.
  • Complex pose human body structure: In multi-person interaction, occlusion, exaggerated actions, or unusual viewpoints, hands, limbs, and body structure may appear unnatural or inconsistent.
  • Small faces and small limbs: Due to the use of the open-source FLUX.1 VAE, reconstruction loss is relatively large, and details such as small faces, small limbs, eyes, and text may exhibit artifacts or instability.
  • Limited release scope: Due to resource, engineering complexity, and release boundary constraints, not all training and system details have been open-sourced; the current release aims to balance reproducibility, usability, and sustainable maintenance.

Downstream users are responsible for their own content moderation, verification, and compliance checks.

Official special statement: The Boogu team currently does not provide any paid API, subscription, or commercial service; any paid product or service offered under the name Boogu-Image or similar variants is unrelated to this project and unofficial. Please verify carefully and protect your personal privacy and financial security.

Related tools

Browse category