EditUtils: Qwen Config Preparer lrzjason

This is the node that makes multi-reference editing work in EditUtils. QwenConfigPreparer_EditUtils takes one image and turns it into a config block describing how that image should be treated - whether it goes into the reference latent, into the Qwen2.5-VL text-encoder vision stream, which one is the "main" image, at what resolution, with what cropping. Then it appends that block to a configs list and passes the list on. Chain a few of these together, one per image, and EditTextEncode_EditUtils receives a list of however many references you want - the README's "unlimited images" claim, with the caveat that your VRAM will object before the node does.

Mechanically each config is a Python dict with a fixed set of keys, and every one maps to a real widget:

• to_ref - put this image into the reference latent (the appearance/identity stream that the sampler uses). Default on.
• ref_main_image - mark this as the image. Its latent becomes the output latent that seeds sampling, and it's the one that generates pad_info. Default on for the first node in the chain; only one image can be main, and the pack force-flips later ones off.
• ref_longest_edge (default 1024), ref_crop (pad/center/disabled), ref_upscale (lanczos/bicubic/area) - the resize-and-align treatment, same logic as LongestEdgeImageProcess_EditUtils. pad pads to the VAE unit with black, which is what keeps high-res refs (2xxx–3xxx) clean.
• to_vl - also feed this image to Qwen2.5-VL, the vision-language encoder that gives the model semantic understanding of what's in the picture. This is a separate stream from the VAE latent.
• vl_resize, vl_target_size (default 384), vl_crop, vl_upscale - how to shrink the image before it hits the VL encoder (VL is fed small; it reads meaning, not pixels).
• mask - optional, applies only to this image's reference latent (if its size doesn't match the image, it's dropped with a console warning).
• ref_resize_mode - longest_edge (longest dimension = ref_longest_edge) vs area (total pixels = ref_longest_edge²).
• rope_x_offset / rope_y_offset - ROPE position offsets, mostly a Krea2 regional-editing feature; only effective when ref_pos_match_target=False on the apply node.

Outputs: configs (the accumulated LIST, wire it onward or into the encode node) and config (the just-created block for this image, handy for debugging or feeding a single-image encoder).

Where people get burned: thinking to_ref and to_vl are redundant. They're two different streams doing two different jobs - the VAE latent drives low-level appearance (keep this dress, this pose), the VL stream drives semantic understanding (this is a person sitting on a bench). Turn to_vl off and the model still sees the reference as pixels but understands much less about it; turn to_ref off and it understands but has no pixel-level anchor. For a normal edit, both on is right.

Install

Part of lrzjason/ComfyUI-EditUtils. Install via ComfyUI Manager (search "ComfyUI-EditUtils") or cd ComfyUI/custom_nodes && git clone https://github.com/lrzjason/ComfyUI-EditUtils, then restart. No pip deps.

Troubleshooting

If a config silently ignores your mask, check the mask dimensions match the image - mismatches are dropped with a console print, not an error. If the wrong image is driving the output latent, you've got ref_main_image on more than one node; the pack forces only the first to be main, so reorder or untick. And the "unlimited images" pitch has a real ceiling: every reference costs VAE encode time and sampler memory, and the Qwen-Image-Edit family is 20B - start with two refs, add more only when the VRAM gods permit.