LoRAtech - ComfyUI Hub

AI Influencer Master Class - Krea 2 Inpaint (perfect background) (Patreon)

Published:

2026-08-16 22:10:58

Edited:

2026-08-18 04:46:49

Imported:

2026-10

Downloads

Download Krea 2 Inpaint v3 - by LoRAtech (2).json

Content

Swaps the person in a reference photo for your character while keeping clothing, pose, and background intact. Works best with high quality reference photos where the person isn't too far away, results vary with different sizes, dimensions, and pixel counts since masking behavior shifts with the source image.

How to use it

1. Load your character LoRA Place your character LoRA into the LoraLoaderModelOnly node. Strength defaults to 1.0, lower it if the LoRA is overpowering the result.

2. Update the "Describe Person" prompt Start with your LoRA keyword, then a basic description like hair color and eye color. Example: "YourKeyword, a woman with natural blue eyes and rose gold hair. The edges of the woman blend in seamlessly with the background." Keep the last sentence, it helps the inpaint blend cleanly.

3. CR Upscale Image (reference photo size) Upscales your reference photo before inpainting. Default rescale factor is 1.5. If your reference is already large and high quality, lower the rescale factor or set it to 1.0. If it's small, raise it so there's enough pixel detail to work with.

4. Set your output resolution The Resize Images By Longer Edge node controls the final output size. Default is 1280. Higher values give a larger output but cost more VRAM and generation time.

5. Sampler denoise (inpaint pass) Controls how much the masked area gets regenerated. Default is 0.40. Raise it and the masked area follows your LoRA and prompt more strongly but keeps less of the original reference. Lower it and more of the original image survives but your character comes through weaker. Experiment in the 0.4 to 0.8 range.

6. Eye detailer denoise By default, only one detailer pass runs, using the eye bbox model, denoise 0.15. There's a second detailer node wired to a face bbox model (denoise 0.10), but it's disabled (muted) in this workflow out of the box. Raise or lower the active detailer's denoise for more or less detail. Stay in the 0.10 to 0.35 range, going higher starts changing the face instead of enhancing it. If you want an added face pass, turn that second node's mode back to active.

7. Describe Mask prompt Tells SAM what to mask, and the masked area is what gets regenerated. Default covers the entire body including exposed skin and hair. Make it more specific (like "the face and hair of the woman") to regenerate less of the image, or broader to regenerate more. Anything not covered by the mask stays untouched from the original reference.

Installing the required models

Every model below is a one-time download. Once it's in the right folder ComfyUI will pick it up automatically, no further setup needed. Your ComfyUI models folder lives at ComfyUI/models/ inside your ComfyUI directory, and each model goes into its own subfolder as noted below.

SAM model (sam_vit_b_01ec64.pth) — download from Hugging Face, place in ComfyUI/models/sams/

SAM 3.1 masking checkpoint (sam3.1_multiplex_fp16.safetensors) — download from Hugging Face, place in ComfyUI/models/checkpoints/

Face bbox detector (face_yolov8m.pt) — download from Hugging Face, place in ComfyUI/models/ultralytics/bbox/

Eye bbox detector (Eyeful_v2-Paired.pt) — download from Hugging Face, place in ComfyUI/models/ultralytics/bbox/

Text encoder (qwen3vl_4b_fp8_scaled.safetensors) — download from Hugging Face, place in ComfyUI/models/text_encoders/

VAE (qwen_image_vae.safetensors) — download from Hugging Face, place in ComfyUI/models/vae/

Diffusion model (krea2_turbo_fp8_scaled.safetensors) — download from Hugging Face, place in ComfyUI/models/diffusion_models/

Upscale model (RealESRGAN_x4plus.safetensors) — download from Hugging Face, place in ComfyUI/models/upscale_models/

Your character LoRA — whatever you trained or downloaded, place in ComfyUI/models/loras/

This one needs two extra models compared to the other two workflows: the SAM 3.1 checkpoint for masking, and the RealESRGAN upscale model.

Install commands (Windows)

Open Command Prompt or PowerShell and cd into your local ComfyUI directory first (the folder that contains the models folder), then run:

If models\ultralytics\bbox doesn't exist yet, create it first with mkdir models\ultralytics\bbox.

Install commands (RunPod / JupyterLab)

Your RunPod ComfyUI install always lives at /workspace/ComfyUI, so you can run these directly from any directory in the JupyterLab terminal:

Workflow by LoRAtech