Boogu-Image

Boogu-Image-0.1 is an open-weight image generation and editing model family from the Boogu Team, published on GitHub as boogu-project/Boogu-Image and on Hugging Face under the Boogu org. It has four 10B-parameter versions — Base, Turbo, Edit and Edit-Turbo — released in June 2026 under Apache 2.0, with a technical report on arXiv (2607.13125).

Base is the full text-to-image model (25–50 steps, 1K to 2K, best for dense text). Turbo is a distilled text-to-image version that needs only 4 steps at 1K. Edit changes an existing picture from one reference image and a text instruction. Edit-Turbo is the 4-step distilled editor, released on 30 June 2026.

18 credits per generated image and 60 credits per edit with a reference image — about $0.18–$0.21 and $0.60–$0.69 depending on which credit pack you buy. Packs are one-time purchases with no subscription. Boogu Image is not on the free plan; Qwen Image 3.0 is.

Yes — bilingual Chinese and English text rendering is its main focus. The team prioritised the 3,500 most common modern Chinese characters in training. It still warns that long text, small fonts and complex layouts can produce typos or missing characters, and that other languages are not optimised.

The Boogu Team describes the release as a research project and says it offers no paid API, subscription or commercial service, so paid sites using the name are not affiliated with it. Kenerate AI is not affiliated either: we run a hosted copy of the open Apache 2.0 model so you can use it without a GPU.

ComfyUI has an official Boogu-Image tutorial with templates for Turbo and Edit. You download the fp8 diffusion model, the Qwen3-VL-8B text encoder and the VAE from Comfy-Org/Boogu-Image on Hugging Face into the matching models folders, then load the workflow. Turbo needs 3–4 sampling steps.

The GitHub README lists settings from 12 GB up: on 12–16 GB cards you use the fp8 weights with CPU offload, 24–32 GB runs fp8 directly at 1K, and the unquantized Base model is listed for 40–80 GB. In the browser on Kenerate you need no GPU.

Both handle English and Chinese text. In the Boogu technical report the authors claim state-of-the-art open-source results on Qwen-Image-Bench, but those are author-reported. On Kenerate Qwen Image costs 6 credits and Qwen Image 3.0 costs 9 (free plan, up to 3 references), so start there and try Boogu Image (18 credits) when you want a second open model for bilingual lettering.

The technical report gives GenEval scores of 0.84–0.85, below HiDream-O1-Image (0.90) and GPT-Image-2 (0.89), and the authors say GenEval differs from human preference. These are the authors' own numbers; we found no independent benchmark yet, so judge it with your own prompts.