Digital artists and AI developers have long struggled with a recurring nightmare known as identity collapse. The scenario is always the same: you generate a perfect character or upload a high-resolution portrait, but the moment you attempt to change the background or swap an outfit, the person's face subtly shifts. A jawline softens, the eyes change shape, or the distinct ethnic markers that made the subject unique simply vanish. This instability has made professional-grade AI image editing a game of chance, where maintaining a consistent persona across different scenes requires tedious manual masking and hours of inpainting.

The Architecture of Precision and Efficiency

Krea 2 Identity Edit emerges as a community-driven solution to this instability, built upon the foundation of the Krea 2 Raw model. At its core, the system utilizes a single-stream Multi-modal Diffusion Transformer (MMDiT) boasting 12.9 billion parameters. Unlike traditional architectures that process text and images in separate silos, the MMDiT structure allows the model to handle both modalities simultaneously within the same neural network, leading to a more cohesive understanding of how a textual prompt should modify a visual subject. The latest v1.1 iteration focuses heavily on improving facial similarity and overall image fidelity. By enhancing the locality of the editing process, the model ensures that camera angles, body poses, and untouched environmental elements remain static while only the targeted areas change.

One of the most significant breakthroughs in v1.1 is its ability to handle multi-subject compositions. Previous models often suffered from identity bleeding, where the features of two different people in one image would merge or swap during the editing process. Krea 2 Identity Edit effectively isolates individual identities, allowing for precise object removal, replacement, and complex clothing changes without compromising the distinct look of each person. To ensure this power is not limited to those with enterprise-grade GPUs, the developers implemented Singular Value Decomposition (SVD). This mathematical technique decomposes large matrices to reduce model size without sacrificing significant quality. The result is a suite of low-VRAM variants: the r128 version requires only 0.91GB and the r64 version requires a mere 0.46GB, both delivering results nearly identical to the full v1.1 model.

To implement this workflow, users must utilize the ComfyUI-Krea2Edit node pack. This is necessary because the model employs a dual-conditioning method that combines Variational Autoencoder (VAE) tokens with Qwen3-VL encoding, a sophisticated vision-language model that provides the spatial awareness required for identity preservation. The environment can be configured using the following command:

bash

ComfyUI-Krea2Edit 노드 팩 설치

git clone https://github.com/lbouaraba/comfyui-krea2edit

Mastering the Latent Space and Grounding

Moving from installation to execution requires a strict adherence to input logic, as the model is highly sensitive to the order of operations. The system supports two primary input editing functions: a scene image and a person image. For the model to function correctly, the first image, designated as `source_latent`, must always be the background scene. The second image, `source_latent_b`, must be the person. Reversing this order leads to a catastrophic drop in output quality, as the model fails to correctly map the identity onto the environment.

Operational success depends on choosing the right model variant and configuration based on the task. For most additive or transformative tasks—such as changing colors, altering styles, or repositioning a subject—the Turbo model is the optimal choice. In these scenarios, developers should set the step count between 8 and 12 and maintain a Classifier-Free Guidance (CFG) value of 1.0. However, when the task involves significant subtraction, such as removing a large object from the frame, the Raw model is required. This more intensive process demands 20 steps and a CFG of 3.0 to ensure the void is filled naturally without leaving artifacts.

The most critical variable for fine-tuning the output is the `grounding_px` value. This parameter acts as the fulcrum between edit strength and identity preservation. A lower `grounding_px` value forces the model to follow the text prompt more aggressively, which can lead to more uniform changes across the scene but risks altering the subject's face. Conversely, a higher value locks the identity more firmly. The training range for v1.1 spans from 384 to 768, with 768 serving as the default. If a user encounters a duplication glitch where the screen appears split or the subject is mirrored, the solution is to lower the `grounding_px` value.

Resolution management is the final piece of the puzzle. The aspect ratio of the source image and the output image must be identical to prevent stretching or warping. Furthermore, the total resolution should be kept under 2 megapixels. When editing images containing two people, the most effective strategy to prevent identity mixing is to generate the initial image between 1 and 1.5 megapixels and then apply an upscaling pass afterward. Practical prompt examples for this workflow include:

python

프롬프트 예시

"create a photo of this person at a night market"

"replace the woman with a big orangutan"

"create a photo of this man next to the tractor"

Beyond these native settings, Krea 2 Identity Edit offers a level of flexibility that closed-source AI editors cannot match. It supports Low-Rank Adaptation (LoRA) stacking, allowing users to layer their own custom-trained character or style LoRAs on top of the base model. This means a creator can not only preserve a real person's identity but also force that identity into a specific artistic style or a fictional universe with surgical precision.

The rapid evolution of these open-source tools is effectively dismantling the monopoly that proprietary services held over high-fidelity identity editing.