The open-source community has received a new tool for high-precision image editing: the unofficial Krea 2 Identity Edit (v1.1) LoRA file for the Krea 2 Raw model (12.9B MMDiT) has been released. This technology allows for changing scene details based on text instructions while preserving character identity and the overall composition of the frame.
What Happened
The developer has introduced LoRA version 1.1, optimized for the Krea 2 Raw model with the MMDiT architecture (12.9B parameters). The tool supports local edits, such as changing object colors, replacing clothing, and moving characters into new scenes. Integration into ComfyUI requires the use of a specialized node set, ComfyUI-Krea2Edit, which applies a dual conditioning mechanism via VAE tokens and the Qwen3-VL text encoder.
Context
The Multi-Modal Diffusion Transformer (MMDiT) architecture provides deeper integration between text commands and visual features. The use of dual conditioning (VAE tokens + Qwen3-VL) allows for a high level of control, which is critical for tasks requiring content manipulation without destroying the original image structure or altering facial features.
Why It Matters for the Industry
The emergence of high-quality LoRAs for identity control in multimodal models significantly simplifies the instruction-based editing process in the open-source segment. This brings the capabilities of local solutions closer to proprietary cloud services like Krea.ai and creates a foundation for developing new AI editors and agentic workflows focused on deep semantic editing.
Why It Matters for Users
ComfyUI users can now perform professional character and object editing in a local environment without resorting to expensive APIs and cloud computing. This allows for precise content personalization (e.g., changing clothes or surroundings) while minimizing the risk of "breaking" the face or composition, which was previously characteristic of simple generation methods.
What Is Not Yet Known / Limitations
There are legal risks associated with using such technologies to create deepfakes and potential violations of personality rights.
Sources
Author
Look at AI, Editorial Team