CivArchive
    Elusarca's Krea 2 Detail Enhancer | Edit LoRA - v1.0
    Preview 135842119
    Preview 135842121
    Preview 135842145
    Preview 135842134
    Preview 135842136
    Preview 135842155
    Preview 135842142
    Preview 135842139
    Preview 135842123
    Preview 135842129
    Preview 135842131
    Preview 135842141

    Hi, recently Ostris, creator & maintainer of AI Toolkit, released a new lora training method and custom ComfyUI node for editing images with Krea2 despite model's text-to-image nature.

    Upon that I trained several detail enhancement LoRA's and sharing the best one.

    Please be aware, this LoRA and the edit model is highly experimental and it's not creating great results all the time. I'm merely sharing my experiment for inspiring other developers & community members.

    Some important notes:

    -Krea2 is not an edit model, so do not expect Flux.2 Klein/Qwen Image Edit precision and performance from it. It alters the image sometimes.
    -It might have faulty outputs with horizontal aspect ratio.
    -It changes the lighting and color slightly.

    Trigger word: enhance this image

    Prompt I'm using:

    Enhance this image to high resolution with rich fine details. Sharpen all textures and surfaces, add microdetails and natural grain. Increase clarity and definition across all elements while preserving the original composition, lighting, and atmosphere. Image can be illustration or real photo, keep the original input style.


    Workflow I'm using: https://huggingface.co/reverentelusarca/krea2-detail-enhancer-edit-lora/blob/main/workflow-comfyui-krea2-detail-enhancer-edit-lora.json

    Edit Node: https://github.com/ostris/ComfyUI-Krea2-Ostris-Edit

    Explanatory post from Ostris: https://x.com/ostrisai/status/2073428647273447480

    Description

    FAQ

    Comments (14)

    phazeiJul 6, 2026· 11 reactions
    CivitAI

    FYI to anyone who sets this up: The image must be 1MP, any larger and it won't work, it'll have the image in the top left corner, and you'll see some other image behind it being generated as well.

    Also, without a LoRA, the editing doesn't work, each edit must have a specific LoRA that is trained to do the exact edit being requested.

    reverentelusarca
    Author
    Jul 6, 2026

    Thank you for noticing this, I've always resize the image to 1K before sending it to the node, that's probably why I didn't came across.

    hirez80462Jul 12, 2026

    true dat. there should be a resizer before so we dont need to go hunting.

    Dull_bbbbJul 31, 2026

    @reverentelusarca is there a way you can push it to 2K for the next versions? This looks awesome and I'm wanting to try it out today.

    KisukeJul 6, 2026· 6 reactions
    CivitAI

    It didn't work. It totally changes the image and turns it into a mosaic.

    reverentelusarca
    Author
    Jul 6, 2026

    Then you should be doing something wrong. If mosaic is happening then lora/node not being applied properly. I took a snapshot for my workflow just now so you can compare with yours: https://ibb.co/8DtrmHFS

    reverentelusarca
    Author
    Jul 6, 2026

    and this is the mosaic artifact you are talking about. It happens if LoRA is not applied properly(in the snapshot I bypassed the LoRA Loader node): https://ibb.co/0yBBFtmH

    KisukeJul 6, 2026· 1 reaction

    @reverentelusarca I noticed the rgthree LoRA Loader doesn't work in my ComfyUI for some reason. So I used the standard LoRA Loader instead, and it worked. Thanks.

    VinotintaJul 18, 2026
    CivitAI

    does this do the same thing as this ? - https://civitai.com/models/2729908/krea-2-detail-slider

    reverentelusarca
    Author
    Jul 19, 2026

    Absolutely not. Concept-slider LoRAs are very easy to train and do not require a traditional curated dataset. For example, when training a weight-slider LoRA, instead of prompting the model with “extremely fat,” you teach it to represent that concept through numerical weight values, similar to weighted prompting such as (fat:1.5). So its basically doesn't teach anything new to the model.

    My LoRA, and the method I used to train it, includes many before-and-after image pairs—essentially the same structure used when training an edit-model LoRA.

    VinotintaJul 19, 2026

    @reverentelusarca meh, well i went with the other thing anyway, thanks for your response.

    flippa788Jul 20, 2026
    CivitAI

    Works great with the identity edit workflow. Good work.

    ProvenFlawlessJul 21, 2026
    CivitAI

    It gives some interesting results even when you don't use for it EDIT. Just use it as a lora on its own.

    reverentelusarca
    Author
    Jul 21, 2026

    oh wow, can you give me some ideas to test it out?