CivArchive
    Krea 2 Regional Multi-LoRA — Multi-Character + Bounding-Box Layout Control By Fedor - v1.0
    NSFW
    Preview 135892100
    Preview 135892101
    Preview 135892102
    Preview 135892105

    # Krea 2 Regional Multi-LoRA — Multi-Character + Bounding-Box Layout Control

    Put multiple character LoRAs in a single Krea 2 image, each one locked to its own bounding box — no bleed, no merged faces, no averaging. And it's not just for LoRAs: draw and describe boxes for objects, props, backgrounds, and extra subjects too, exactly like Ideogram 4's bounding-box prompting.

    ## What it does

    Normal LoRA loading applies everywhere, so two character LoRAs smear into each other. This node injects each LoRA's effect only into the image tokens inside its box, at forward time — outside the box the effect is multiplied by zero. It's a hard spatial guarantee, not an attention-bias nudge the model can ignore.

    Pair it with an Ideogram 4-style prompt builder and every box does double duty:

    - Every box places its described content via Krea 2's Qwen3-VL text encoder (a table, a neon sign, a dog on the left — Krea 2 honors the placement).

    - LoRA boxes additionally lock in a specific trained identity on top of that placement.

    Sketch the whole scene as boxes, describe each one, and drop LoRAs into the boxes that need a precise face. Objects and characters, all placed by the same boxes.

    ## Features

    - Unlimited regions — 2 characters or 10, add a row per box

    - Region rows auto-sync to the boxes you draw (draw a box, a row appears)

    - Hard per-region LoRA masking (activation-delta injection)

    - Bounding-box layout control for non-LoRA elements too

    - fp8-safe — never touches quantized model weights

    - Runs at Krea 2's native CFG 1

    ## Requirements

    - ComfyUI with Krea 2 support (recent build)

    - Models: krea2_turbo_bf16 (UNet), qwen3vl_4b_bf16 (CLIP, type krea2), qwen_image_vae (VAE)

    - Custom node: ComfyUI-Krea2-Regional-MultiLoRA (this workflow's node)

    - ComfyUI-KJNodes (for the box-drawing prompt builder)

    - Character LoRAs trained against Krea 2 (e.g. via ai-toolkit)

    ## How to use

    1. Write your scene prompt in the box builder (setting/lighting/camera — not the characters).

    2. Draw one box per element, in order. Rows appear automatically in the LoRA node.

    3. Assign a LoRA to each character box; leave object/scenery boxes as plain descriptions.

    4. Sampler: euler / bong_tangent / 8–12 steps / CFG 1.

    5. Queue.

    ## Tips

    - Keep character boxes from overlapping to avoid bleed.

    - If seams show, raise seam_feather a touch (0.12–0.15).

    - Row order must match box order (row 1 = first box drawn).

    - Character LoRAs must be Krea 2-trained — FLUX/SDXL LoRAs load but look wrong.

    ## Node + workflow (GitHub)

    Full source, install instructions, and the example workflow:

    https://github.com/CliffNodes/Krea2-Multi-Character-Lora-Node-w-bounding-box

    License: MIT.

    Description

    FAQ

    Comments (9)

    basicslowbroJul 7, 2026
    CivitAI

    it kinda works, I have to crank my loras to like 2.35. sometimes it looks splotchy, especially the skin texture.

    interestingly only characters trained on AI toolkit seem to work, like Onetrainer recently updated for Krea training, but the node seems to not merely ignore OT loras, but completely alter the race and physical appearance of it

    yajukunJul 7, 2026

    Interesting. After reading your comment I read the desc on github and it does mention 'The reference trainer is ai-toolkit', maybe that was the only type of character loras it was tested with?

    ftrbzmogul355
    Author
    Jul 7, 2026· 1 reaction

    @yajukun Correct I only have AI toolkit loras and I have only tested with those. Shouldn't a a Lora behave like a Lora regardless of which program it was created though? Forgive my ignorance I've only ever used AI toolkit

    ftrbzmogul355
    Author
    Jul 7, 2026· 1 reaction

    @yajukun Here is an image of my workflow for ref https://ibb.co/ZRDj1mZH

    ftrbzmogul355
    Author
    Jul 7, 2026· 1 reaction

    @yajukun I also added a section for recommended settings to the github

    yajukunJul 8, 2026

    @ftrbzmogul355 Thanks, will give it a shot.

    nekotxtJul 7, 2026
    CivitAI

    It somewhat works: the basic image is there. But when I try to make adjustments (reposition the boxes) it mostly ignores my instructions.

    Sometimes it respects the prompt and location surprisingly well, most of times it does little or nothing.

    I still have to figure out the process.

    ftrbzmogul355
    Author
    Jul 7, 2026· 1 reaction

    I added some recommended settings to the git hub. Here is a screenshot of my workflow for ref https://ibb.co/ZRDj1mZH

    brnfd24434343dJul 9, 2026· 1 reaction
    CivitAI

    Is the "fedor bypass" just another alternative to the bypass 2vector and 3vector loras?

    Workflows
    Krea 2

    Details

    Downloads
    506
    Platform
    CivitAI
    Platform Status
    Available
    Created
    7/6/2026
    Updated
    7/26/2026
    Deleted
    -

    Files

    krea2RegionalMultiLoraMulti_v10.zip

    Mirrors

    krea2RegionalMultiLoraMulti_v10.json

    Mirrors