BOOGU: Load Edit Pipeline

The thing that made Boogu-Image a headline in June 2026 wasn't its text-to-image. It was that this ~10B Apache-2.0 model could edit - feed it an image and a plain sentence and it reliably changes the scene, adds characters, even does view synthesis. The BOOGULoadEditPipeline loader is the legacy ComfyUI door to that capability, loading the Boogu-Image-0.1-Edit model so BOOGUGenerate can use it as an image editor.

Early community comparisons put Boogu's edit model up against Flux Klein and Qwen-Image-Edit and found it holding scene context and character identity surprisingly well on complex instruction edits - the "new open king edit model" takes were flying. That's the capability you're unlocking here. If your workflow is "I have an image, I want it changed" - restyle a render, swap a background, add a person doing a specific action - this loader is the one you want.

What it does

Same mechanism as the Base loader, different model folder. It finds ComfyUI/models/boogu/Boogu-Image-0.1-Edit/, builds a BooguImagePipeline from it, and hands you a pipeline output to wire into BOOGUGenerate. The clever bit is how editing works: because Boogu's text encoder is a Qwen3VL vision-language model, the pipeline feeds your input image and your instruction to the LLM together. The VLM actually sees what's in the frame, so you can write "turn this into a night scene but keep the street sign" and it has a chance of actually doing it. Editing here is genuinely instruction-driven, not "img2img with a low denoise."

In BOOGUGenerate, connect your image to the optional input_image socket, set width/height to the image's aspect, and keep the steps up around the author's 50-ish default. The node automatically aligns output resolution to the input on the edit path, and the VLM side-limits (max_vlm_input_pil_*) control how the image is resized before the 8B encoder reads it - the defaults are sensible, leave them.

The inputs

Just like its siblings: dtype (keep bf16), device (cuda unless you must), enable_model_cpu_offload (on if it won't fit VRAM), and optional force_reload to drop the cached pipeline when you change dtype/device. That's the entire input surface. The one output, pipeline, goes to BOOGUGenerate - this time with an image also connected to its input_image.

The Edit and Base models are both loaded by the same underlying class, and the loader caches each variant separately. You can keep an Edit pipeline and a Base pipeline in one graph; they don't fight over the same cache entry.

Installing

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

or install via ComfyUI Manager (search "Boogu"). You also need the standalone boogu package from the Boogu-Image repo (pip install -e /path/to/Boogu-Image) and the model folder at exactly ComfyUI/models/boogu/Boogu-Image-0.1-Edit/. The download is big - ~10B weights plus the 8B Qwen3VL encoder.

Common issues

• ModuleNotFoundError: No module named 'boogu' - the node needs the standalone package; Manager can't install it for you.
• "Resolved model path does not exist" - the folder isn't in ComfyUI/models/boogu/, and this legacy loader ignores extra_model_paths.yaml.
• Edits look like loose img2img, not real edits - make sure you've connected input_image, and double-check the instruction actually describes an edit to the image rather than a brand-new scene.

And the standing caveat: this whole pack is the legacy path. Native ComfyUI support (Comfy-Org/Boogu-Image files, TextEncodeBooguEdit for edits) is what the README points most users to. Use this loader if you specifically want the old pipeline-folder workflow.