BOOGU: Load Base Pipeline

BOOGULoadBasePipeline is the "serious output" loader in the ComfyUI-Boogu pack. It loads Boogu-Image-0.1-Base - the ~50-step, full-quality text-to-image model that the community's Boogu hype is mostly built on. If you want the best image a Boogu graph can give you and you're willing to wait, this is the loader you reach for.

Before the setup lecture: this pack is the legacy path for Boogu-Image. The model is supported natively in ComfyUI now (Comfy-Org PR #14523, repackaged files on HuggingFace at Comfy-Org/Boogu-Image), and the README literally says "for most users, do not install this custom node." Everything below is for people who specifically want the old pipeline-folder workflow - or who want to understand what the native nodes are abstracting away.

What it does

This loader finds the Boogu-Image-0.1-Base folder under ComfyUI/models/boogu/, constructs a HuggingFace BooguImagePipeline from it, and returns a single custom-typed pipeline output that plugs into BOOGUGenerate. The other two loaders are the same class with a different model folder swapped in - Base, Edit, and Turbo are just three variants of one loading mechanism.

Boogu-Image is a ~10B unified generation/editing model with a Qwen3VL-8B LLM text encoder and a Flux VAE, released June 2026 under Apache 2.0. The Base variant is the one that does full-quality text-to-image (and can also edit, since it's unified). It is not step-distilled like the Turbo variant, so it wants the classic ~50 inference steps for the image to converge.

The inputs that matter

There are only three required inputs, and two of them are one-time decisions:

• dtype - bf16 (default), fp16, or fp32. Keep bf16. It's the author's default, it's what the model was trained for, and fp32 just buys you double the VRAM for nothing.
• device - cuda (default, maps to cuda:0), cuda:0–cuda:7, or cpu. Pick your GPU. CPU technically works and is a patience exercise.
• enable_model_cpu_offload - default off. Flip this on if the model won't fit in VRAM; it shuttles components between RAM and VRAM and slows generation, so only use it when you must.

One optional input: force_reload. The loader caches loaded pipelines, so rebuilding the same graph reuses the loaded model instead of reloading it from disk. Set force_reload when you change dtype/device on a graph you've already run - it drops the cached copy so your change actually takes effect.

The output, pipeline (type BOOGU_PIPELINE), feeds directly into BOOGUGenerate's pipeline socket. That's the whole wiring: Load Base → Generate → PreviewImage.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/boogu-project/ComfyUI-Boogu

or search "Boogu" in ComfyUI Manager. Two extra things this pack needs that Manager won't fetch for you: the standalone boogu Python package from the Boogu-Image repo (pip install -e /path/to/Boogu-Image), and the model folder dropped at exactly ComfyUI/models/boogu/Boogu-Image-0.1-Base/. At ~10B diffusion weights plus the 8B encoder, be ready for a serious download and a serious VRAM footprint.

Common issues

• ModuleNotFoundError: No module named 'boogu' - you skipped the standalone package install. Install it, or switch to native ComfyUI support.
• "Resolved model path does not exist" - the folder isn't at ComfyUI/models/boogu/Boogu-Image-0.1-Base, or you expect extra_model_paths.yaml to work (it doesn't with this legacy loader; only native support honors it).
• VRAM errors - flip enable_model_cpu_offload on, or go the native route and grab an fp8/GGUF file, which the legacy loader can't run at all.

The base model is the quality pick, not the quick one. If you want speed, that's what BOOGULoadTurboPipeline is for - and that one, as they say, comes with its own trade-offs.