Boogu-Image shipped four variants in two weeks, then hotfixed two of them

The anti-scam notice contradicts the release log eleven entries below

The first thing the document does is warn you away from somebody else. It states that the team does not currently provide any paid API, subscription or commercial service, that anything offered under the name Boogu-Image or a variant such as `booguimage`, `Boogu Image` or `Boogu` is not affiliated with the project, and that readers should verify carefully before paying anyone. The same block adds that Boogu-Image-0.1 is a research project only and not an official model release. Eleven news entries further down contradict that last sentence: Base, Turbo and Edit all arrive on 2026-06-16, Edit-Turbo on 2026-06-30, each with a Hugging Face revision and a live demo behind it. So the document asks you to distrust commercial offers carrying its name while shipping dated public checkpoints under its own name. Both are only consistent if the caution is aimed at resellers rather than at the weights.

Two of the four variants needed artifact hotfixes inside a month

Checkpoint names here are not quality signals, because two of the four variants were republished within weeks. Turbo shipped on 2026-06-16 and a Turbo-hotfix followed on 2026-06-25, described as a minor patch release with new weights and no feature changes, fixing visual artifacts at different aspect ratios and background overfitting artifacts. Edit-Turbo shipped on 2026-06-30, and the 2026-07-08 entry calls its own update an Image-to-Image hotfix addressing severe image quality degradation plus poor performance on removal and other editing tasks. That entry is the published record admitting the first public editing checkpoint was weak on the two things it was released for. Both fixes live in dated Hugging Face revisions, `hotfix-20260625` for Turbo and `hotfix-1k-20260708` with `hotfix-1k5-20260708` for Edit-Turbo, and the repository has no GitHub releases to compare them against. Record which revision you loaded, because the family version number does not change.

2K is supported, 1K is recommended, and one reference image only

The resolution guidance contradicts itself on purpose. The Edit entry states the model supports resolutions up to 2K but that results are more stable at 1K, and the hotfix entry then recommends downloading the 1K checkpoint for more stable results, which is the same preference stated a second time. The capability and the advice disagree, and the demos reflect it: separate online demos exist for the 1k and 1.5k resolution variants of Edit-Turbo, while Base and Turbo have one demo each. The same Edit entry states that only one reference image is supported for now, with more planned. So the editing path is a single image tool at a resolution you are told to lower, and the two demos let you see both settings before committing to a checkpoint.

`requires-python` stops at 3.12 and triton is declared on Linux x86_64 only

The install surface is a `pyproject.toml`, and the text gives no command to run. `requires-python` is `">=3.10,<3.13"`, so 3.13 and everything newer are excluded outright, and torch is pinned from 2.7.1 to below 2.12. Two pins are tighter than the rest: `kernels>=0.14,<0.15` permits a single minor series, and `triton>=3.0` is declared only when the platform is Linux on x86_64, so it is skipped entirely on macOS and Windows. `webdataset` and `omegaconf` point at the batch data and configuration layers, and `cache-dit` at caching diffusion steps.

[project]
name = "boogu-image"
version = "0.1.0"
requires-python = ">=3.10,<3.13"
dependencies = [
    "torch>=2.7.1,<2.12",
    "diffusers[torch]>=0.35.2,<0.39",
    "transformers[torch]>=4.57.3,<6",
    "kernels>=0.14,<0.15",
    "triton>=3.0; platform_system == 'Linux' and platform_machine == 'x86_64'",
    "cache-dit>=1.3,<2",
    "webdataset>=1.0,<2",
    "omegaconf>=2.3,<3",
]

The package version tracks the model family rather than the code, sitting at 0.1.0 for Boogu-Image-0.1, so nothing in the manifest moves when a checkpoint is hotfixed.

Four `_fp8` scripts at the root, and no fp8 anywhere in the prose

Eight shell scripts sit at the repository root and four of them are `_fp8` twins: `test_base.sh` with `test_base_fp8.sh`, `test_turbo.sh` with `test_turbo_fp8.sh`, `test_ti2i.sh` with `test_ti2i_fp8.sh`, and `test_ti2i_turbo.sh` with `test_ti2i_turbo_fp8.sh`. That maps one to one onto the four variants across two precision modes, yet no news entry or guide line in the README refers to fp8 or quantization at all. The inference entry points are lopsided in the same way. Base and Turbo each have a plain script and a simple one, `inference.py` with `inference_simple.py`, and `inference_turbo.py` with `inference_turbo_simple.py`. The two editing variants have only the simple form, `inference_ti2i_simple.py` and `inference_ti2i_turbo_simple.py`, with no plain counterpart at the root. `quick_start.sh` and an `INFERENCE_GUIDE.md` are there too, and the text of neither appears in the README.

The wheel picks up `boogu*` only, so `utils/` never ships in it

The distributed package is narrower than the clone. The manifest sets `include = ["boogu*"]` under package discovery, so only the `boogu/` directory is packaged, while `utils/`, `demo_scripts/`, `assets/`, `input_image_examples/` and `batch_data_samples/` stay in a git checkout and are absent from an installed distribution. A `requirements/` directory sits next to the manifest as well, which means the dependency story exists twice. Two READMEs sit at the root, `README.md` and `README_CN.md`, so the notice at the top of this one has a Chinese counterpart. Combined with a package version fixed at 0.1.0 while checkpoints move through dated revisions, no version number in the repository identifies which weights you are running. The file tree is the only place that carries that information, and only down to the revision.

Four routes to the weights, and the newest one is on a branch

This repository offers `INFERENCE_GUIDE.md`, `quick_start.sh` and the root level inference scripts. ComfyUI support arrived twice on the same day, 2026-06-17, as `ComfyUI-Boogu` published both under the Comfy-Org account on Hugging Face and under `boogu-project` on GitHub. The vLLM-Omni inference server gained support on 2026-07-22 through an official recipe hosted in the vllm-project repository, so the fastest documented path runs somebody else's server instead of this code. The newest item, dated 2026-07-23, is NPU backend support on a separate `npu` branch, described as initial and inviting feedback. That same date is the repository's last commit, so the freshest work in the project lives on the branch that a clone of the default branch will not give you.

The Arena leaderboard is self-built and its 1K prompts are not out yet

Boogu Arena is the evaluation, and the same team that made the models built it. The document states they could not evaluate on LM Arena directly, so they use an LLM to generate diverse user personas, ask each persona to produce image generation prompts, and arrive at 1K+ test prompts. Those prompts are described as something that will be released publicly for community reproduction, so in the text as it stands they are not available, which means the leaderboard cannot be reproduced from what the repository holds. The leaderboard itself is embedded as an image, so no score appears in the text, and the document invites teams with questions about the results to contact the authors so they can work toward a more objective, fair and reproducible evaluation. The repository description goes further than the body does, claiming near closed source performance with an order of magnitude less data, while the body says only that the training data scale is roughly one order of magnitude smaller than some existing open source models, with no table, no numbers and no named comparison set.

Editorial conclusion

Read the notice at the top before anything else. It states that no paid API or subscription is offered and that anything sold under the Boogu name is unaffiliated, so a checkout page carrying this model's name is not evidence of an official release. For research use the checkpoints are the interesting part, but select revisions by name rather than by version, since both hotfixes live in dated Hugging Face revisions while the package version stayed at 0.1.0. Two things are worth checking before loading weights. Your Python has to sit inside the >=3.10,<3.13 window, because 3.13 is excluded and triton is declared only on Linux x86_64. And decide whether the evaluation claims matter to you, because the Arena prompts are generated by an LLM and are not published yet, so the leaderboard cannot be reproduced from anything in the repository today.

Frequently asked questions

Official sources

1. boogu-project/Boogu-Image on GitHub
2. Issues
3. License: Apache-2.0
4. Project website
5. README