Skip to content

Reference Image Setup

Tired of a different character generated every scene? Reference images are the solution.

This can be done via Cloud API or a local ComfyUI setup. But the requirement is that the model and the workflow must be able to handle Loading Images.

This page only gets the local workflow running. For what Orb feeds it, and when, see Reference images.

Before you start

  • Finish ComfyUI Setup first.
  • This continues from that page's local Krea 2 section, so it wants the same hardware — 24GB+ VRAM.

Install the custom node

The workflow uses nodes ComfyUI does not ship with. Install them:

cd ComfyUI/custom_nodes
git clone https://github.com/lbouaraba/comfyui-krea2edit
# restart ComfyUI

There is nothing else to install — the node pack needs no extra Python packages.

Do this before you import the workflow into Orb. Orb checks every node against your server, if you get the error 'Krea2EditGroundedEncode', make sure you installed the pack to the correct folder and restarted ComfyUI.

Download the models

Three of these four files are the same ones the Krea 2 section already used. If you did that section, you only need the LoRA.

  1. LoRA — https://civitai.red/models/2761113/krea-2-identity-edit

    Put krea2_identity_edit_v1_2.safetensors in ComfyUI/models/loras/. This is the part that does the editing. Grab the biggest version.

  2. Checkpoint — https://civitai.red/models/2760803/dasiwa-krea2-or-turbo-or-raw?modelVersionId=3151280

    If you haven't, download the int8 version and put it in ComfyUI/models/checkpoints/.

  3. Text encoder — https://civitai.red/models/2731465/qwen3-vl-4b-abliterated-comfyui-krea-2-text-encoder-bf16-fp8?modelVersionId=3070870

    If you haven't, download the fp8 version and put it in ComfyUI/models/text_encoders/.

  4. VAE — https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/blob/main/split_files/vae/qwen_image_vae.safetensors

    If you haven't, download the full weights and put it in ComfyUI/models/vae/.

Restart ComfyUI, or refresh the UI.

Try it in ComfyUI

  1. Download KreaEdit_Default.png and drag it into ComfyUI. The embedded workflow loads automatically.
  2. If drag and drop doesn't work, go to ComfyUI -> File -> Open, then select the image.
  3. The Load Image node is empty. Upload any picture into it — a face works best for seeing what the LoRA does.
  4. The prompt is an instruction, not a description. Write what should change, like she is now wearing a red coat, and leave the rest unsaid.
  5. Click Run, wait, and an image shows up.
  6. Export/Save the output as a PNG. That file carries the whole workflow, which you import into Orb next.

You can also import the default PNG above straight into Orb without touching ComfyUI.

Import it into Orb

Follow Import a ComfyUI workflow, with two things to watch:

  • At the slot check, the positive prompt is the Krea2EditGroundedEncode node that has text in it, and the negative one is the empty one. Orb usually picks right — glance at the node numbers anyway.
  • Leave Width and Height on None. This workflow takes its size from the reference image, so the output follows the shape of the picture you send, not the style's resolution. (grounding_px is not a resolution and Orb never offers it.)

Then, on the style using this workflow, set the one row under Reference images to character reference. A row left Off keeps the filename the workflow shipped with, which doesn't exist on your server, and the render fails in this workflow. But cloud API is more flexible and can handle this case.

Next step

Reference images covers the sources you can pick per row, how the character reference image works, and what happens on reroll.