BOOGU: Generate

BOOGUGenerate is where Boogu-Image actually happens. The three loader nodes just drag a pipeline into your graph; this is the node that reads your instruction, feeds it to the model, and hands you an image. If you're coming from the community excitement around Boogu-Image - the June 2026 Apache-2.0 model that people immediately started pitting against Flux Klein and Qwen-Image-Edit - this is the node you'll spend your time in.

One honest caveat before anything else: this whole pack is legacy. Boogu-Image is now supported natively in ComfyUI (Comfy-Org's PR #14523, model files on HuggingFace at Comfy-Org/Boogu-Image), and the pack's own README tells you to use that instead. The legacy path below still works and still teaches you the model's knobs - just know the install story is more awkward than it needs to be.

How it works

The pipeline you feed in is a full HuggingFace BooguImagePipeline, and it's not your father's CLIP-based diffusion stack. Boogu-Image is a ~10B unified generation-and-editing model whose prompt encoder is a Qwen3VL-8B vision-language model, not CLIP. That matters: the VLM actually sees the input image, so your "instruction" can be a plain-sentence edit ("turn the background into a night street, keep the lighting") rather than a tag soup. The diffusion part is flow-matching style with a separate Flux VAE handling the latent space. You're telling an LLM what to do and letting it write the conditioning, which is why prompt weighting syntax and the 77-token wall are both dead on arrival here.

The inputs that matter

Most of this node's inputs are defaults you should leave alone. The ones you'll actually touch:

• instruction - the plain-language prompt, in any language. The default is a Chinese example ("a brown bear photographed in the forest"); replace it.
• width / height - generation size, 64–4096, snapped to 16. 1024² is a sane starting point.
• num_inference_steps - 50 for base/edit, and the turbo pipeline likes far fewer (more below).
• text_guidance_scale - the author's default of 4.0 is a good start. This is the flow-matching "CFG-ish" knob; crank it up for stricter adherence, and remember distilled models have guidance partly baked in, so don't chase 7+.
• seed - 0 to 2^64. Same seed + same inputs = same image.
• num_images_per_instruction - 1–8, batch size.
• negative_instruction - ships with a sensible default negative; on LLM-encoded models this field is more advisory than it used to be.
• input_image (optional) - an IMAGE tensor. This is what makes it an edit node: connect an image and your instruction becomes "edit this." It gets preprocessed before the VLM reads it, governed by the max_input_image_* and max_vlm_input_pil_* caps (those defaults are fine; the VLM-side cap keeps the 8B encoder from melting).
• instruction_override (optional) - a STRING you can wire from an external Qwen/LLM rewrite node for prompt enhancement before generation (deliberately not bundled in this pack).

The advanced block - system_prompt_follows_task_type, use_boosted_orthogonal_guidance and its bog_* range/mu/interval, plus turbo's use_dmd_student_inference and dmd_conditioning_sigma - is research-grade. Boosted orthogonal guidance (BOG) is a sampling tweak for prompt alignment; leave it off unless experimenting. DMD options only apply to the turbo pipeline.

The single output, image, is a standard IMAGE tensor - wire it to PreviewImage or SaveImage like any other.

Installing (the legacy way)

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

or search "Boogu" in ComfyUI Manager. Then the part people miss: this node needs the standalone boogu Python package and the model folders under ComfyUI/models/boogu/ (each Boogu-Image-0.1-* directory). Get the package from the Boogu-Image repo (pip install -e /path/to/Boogu-Image). It's heavy - ~10B diffusion plus the 8B encoder - so budget VRAM accordingly, and if you're on a tight card the native ComfyUI route with fp8/GGUF files is the better deal.

Common issues

• ModuleNotFoundError: No module named 'boogu' - the classic. You installed the node but not the standalone package. The README's answer: skip this pack and use native ComfyUI support.
• "Resolved model path does not exist" - the legacy loaders only look in ComfyUI/models/boogu/, and they won't honor extra_model_paths.yaml. Native support uses standard model folders instead.
• Turbo + an input image errors out - the turbo pipeline is strictly text-to-image; disconnect input_image.

If you hit any of those, honestly: the fix with the least friction is updating ComfyUI and using the native Boogu nodes. The custom node remains for people who specifically need the old pipeline-folder workflow.