DiffSynth: Boogu Image Loader

Boogu-Image is one of those drops that arrived with no warning and immediately got compared to the model of the month. First impressions from people playing with the Turbo demo were warm - "quite fast… understands public characters and renders them pretty decently… kind of same league or slightly superior to klein" - alongside the inevitable "it looks just like AI generated art". The thing that actually stood out was text: a four-pane storyboard prompt with quoted speech bubbles came back legible, which is not something you can say about most 2026 releases. Community GGUF quants showed up within days, and from there it went quiet, which is the normal arc for a model that isn't from a big-name lab.

There are three checkpoints and they are not interchangeable: Boogu-Image-0.1-Base, -Turbo, and -Edit. This Loader is the same node for all three - the difference is which files you declare and which Inference node you pair it with.

What it loads

Required: model_configs (merged list from DiffSynth: Merge ModelConfigs), torch_dtype (bfloat16 / float16 / float32), device (cuda / cpu).

Optional:

• processor_config - a single ModelConfig pointing at the MLLM/processor directory. Boogu is a transformer plus a vision-language front end, so this is not decoration: the shipped template wires a ModelConfig with pattern mllm/ into it.
• vram_limit - from a VRAM Limit node.

Output: pipe → DiffSynth: Boogu Image Inference.

The template's ModelConfig block, which is the cleanest thing to copy: for Boogu/Boogu-Image-0.1-Base (or -Turbo, or -Edit) declare transformer/*.safetensors, mllm/*.safetensors, vae/*.safetensors, merge those three, and hand the mllm/ directory on its own to processor_config. The globs matter - these repos keep their weights in subfolders, and a pattern that doesn't match anything gives you a download that "succeeds" and a loader that can't find a model.

Install

# ComfyUI Manager: search "DiffSynth-ComfyUI" → Install → restart

pip install git+https://github.com/modelscope/DiffSynth-Studio.git
cd ComfyUI/custom_nodes
git clone https://github.com/modelscope/DiffSynth-ComfyUI.git
cd .. && python main.py

Files download the first time you run the graph, from ModelScope by default. Boogu's repos live on HuggingFace too, so if the default source is crawling, set download_source to huggingface on each ModelConfig node.

Where people get burned

Mixing variants is the classic mistake: the Turbo weights with a 20-step inference, or Base weights with a 4-step one, both produce sludge that looks like a broken install. Turbo is a 4-step model - the template runs exactly 4 - and Base/Edit want the full 20-50. Also keep the pair together: Edit conditioning comes from the Edit checkpoint, so if the edit path seems to do nothing, check which repo you actually loaded. And as with every node in this pack, if the DiffSynth category is empty in your node menu, DiffSynth-Studio isn't installed; the pack imports it at startup and registers nothing when that import fails.