EditUtils: Boogu Config Preparer lrzjason

The Boogu branch of EditUtils has a one-shot encode node (BooguEditTextEncode) that caps you at three reference images. BooguConfigPreparer_EditUtils is the node you reach for when three isn't enough - or when you want per-image control that the one-shot node doesn't expose. Boogu-Image, the 2026-06 model family this pack wraps, does multi-reference editing well, and this is how you actually point it at a pile of references.

Here's the mechanism, because it's the clever part. Each ConfigPreparer takes one image and all of the processing choices for that image, bundles them into a dict, and appends that dict to a configs list. The configs input lets you chain Preparers together: the first one has no list, so it starts one; the second one takes the first's configs output and appends its own config; and so on, for as many images as you want. The finished list goes into EditTextEncode_EditUtils.configs, which walks it and encodes every entry. The README is explicit that the multi-config path is unlimited - the "3 images max" limit only exists in the simple one-shot node.

For each image you're setting three groups of things. The reference group decides whether the image becomes a reference latent at all (to_ref), whether it's the main image whose latent becomes the sampler's starting point (ref_main_image), and how it's resized (ref_longest_edge, ref_resize_mode, ref_crop, ref_upscale). The vision group does the same for the Boogu vision tower (to_vl, vl_target_size, vl_crop, vl_upscale) - Boogu, unlike Flux 2 Klein, uses a vision tower on top of the reference latents, which is why these knobs exist at all. And then there are rope_x_offset / rope_y_offset, which shift where the reference sits on the canvas by nudging its RoPE position - genuinely useful for regional editing, but only if you've got BooguEditApply_EditUtils patching the model, since that's the node that reads the offsets back out of the conditioning.

Two settings that matter more than the others. ref_main_image is the one that decides which reference becomes the seed latent - set it on exactly one image or Boogu defaults to the first. And ref_crop: "pad" is the default for a reason: it letterboxes the image to a VAE-friendly multiple of 8 instead of cropping content off, and it's the mode that pairs with CropWithPadInfo_EditUtils on the way out.

Watch the mask rule: any mask you attach must be the same height and width as the image, or the node silently drops it with a console warning. That's not a crash, it's a "your inpainting is about to ignore you" signal. Install is the pack standard - Manager, search "EditUtils", or clone lrzjason/ComfyUI-EditUtils into custom_nodes and restart. The Boogu model, Qwen3-VL encoder and VAE are yours to supply.