ComfyUI-KleinRefGrid

A single-node ComfyUI custom package that turns up to 4 reference images into a 2×2 stitched grid and injects it as reference_latents into the conditioning stream for the Flux Klein 9B model.

The node ships with a built-in image gallery UI — no external LoadImage nodes, no noodles. Click Add Images, drop up to 4 pictures onto the node, and they are uploaded, thumbnailed in a 2×2 layout, persisted with the workflow, and stitched + encoded at run time.

Features

• Flux Klein RefGrid node under category conditioning/flux_klein.
• In-node gallery: Add Images button, 2×2 thumbnail layout, per-image × remove button, Clear button.
• Gallery state is serialised with the workflow — save, reload, share.
• Every input image is normalised to a fixed 1000×1000 tile so the stitched grid is always exactly 2000×2000 pixels (4 MP), regardless of the input resolutions.
• strength float input that scales the reference latent before injection.
• Outputs the modified CONDITIONING plus the composed IMAGE.
• Live preview of the stitched grid shown on the node after each run (the same mechanism PreviewImage / KSampler uses).

Installation

Via ComfyUI Manager

Search for ComfyUI-KleinRefGrid and install.

Manual

cd ComfyUI/custom_nodes
git clone https://github.com/xb1n0ry/ComfyUI-KleinRefGrid

Restart ComfyUI. No extra Python dependencies beyond what ComfyUI already has.

Usage

1. Add the Flux Klein RefGrid node to your graph (conditioning/flux_klein category).
2. Click Add Images on the node and pick 1–4 reference images.
3. Wire:
   • conditioning ← your text / positive conditioning.
   • vae ← the Flux Klein VAE.
   • strength → 1.0 by default; lower/raise to weaken/strengthen the reference signal.
4. Pipe the conditioning output into your Flux Klein sampler stack. The grid_image output is optional and handy for previewing / saving the stitched reference.

How it works

• Each uploaded image is letterbox-fit onto a 1000×1000 tile.
• Missing slots (<4 images) are filled with black.
• Tiles are stitched into a 2000×2000 grid, VAE-encoded, scaled by strength, and appended to the conditioning's reference_latents list.

Requirements

• ComfyUI with Flux support.
• The Flux Klein 9B checkpoint and its matching VAE.

License

MIT — see LICENSE.