Krea 2 Identity LoRA Tutorial: Perfect Face Consistency in ComfyUI
This guide walks you through using the Krea 2 Identity Edit model inside ComfyUI to edit images while preserving a person's face and identity. You'll learn how to change backgrounds, outfits, lighting, and entire scenes with nothing more than a text prompt.
What You'll Learn
Prerequisites
Before starting, ensure you have:
- ComfyUI installed and running - See our ComfyUI Installation Guide if you need help
- ComfyUI Manager set up for custom node installation
- A GPU with sufficient VRAM - The raw model requires at least 8GB VRAM; the turbo variant works on lower-end GPUs
- Custom nodes for Krea 2 installed in your ComfyUI custom nodes folder (specific node pack referenced in the transcript)
If your hardware isn't powerful enough or you want to speed up generations, consider using RunPod's cloud GPU service.
Special Offer - $5 Credit Included!
When you sign up for RunPod using our affiliate link, you'll receive a $5 credit that can be used to generate up to 9,000 images and 300 videos. This gives you plenty of resources to explore ComfyUI and AI image/video generation without any upfront cost!
Custom Node Installation
Make sure you install the Krea 2 custom nodes before loading the workflow. Use ComfyUI Manager to simplify this process - see our ComfyUI Installation Guide for detailed instructions on custom node management.
Step-by-Step Process
Step 1: Download the Krea 2 Identity Edit Model
- Visit the Hugging Face page for Krea 2 Identity Edit (link provided in the video description)
- Select version 1.1 - This is the latest release at the time of recording
- Choose your model file:
- 1.83 GB version - Best overall quality (recommended)
- 128 version or smaller variants - For GPUs with lower VRAM
- Download the model and place it in your ComfyUI models directory
Model Variants
The raw model is available in different sizes. The 1.83 GB version provides the best results, while smaller variants sacrifice some quality for lower VRAM usage. The turbo model is also available for faster generation with fewer steps.
Pro Tip
For most users, the 1.83 GB version is the best choice. Only use smaller variants if you're running into VRAM limitations on your GPU.
Step 2: Set Up the Workflow in ComfyUI
- Load the provided workflow in ComfyUI (link in video description)
- Disable any two-person editing sections if you're starting with a single-person example
- Load your input image - The transcript uses an image of a woman as the example
- Set image dimensions to 1,536 x 1,536 (height and width)
Workflow Structure
The workflow includes separate sections for single-person and two-person editing. Make sure you're using the correct section for your use case. Disabling the unused section prevents confusion.
Pro Tip
Start with a single-person example first to test your setup. Two-person editing is more complex and shows slightly less accurate facial consistency.
Model Selection
You'll have two model options:
- Krea 2 raw - The standard model, best for quality
- Turbo model - Faster generation, good for beginners or low VRAM setups
Step 3: Configure Settings and Prompt
- Write your prompt describing the desired edit (e.g., "portrait of a woman, depth of field, soft lighting")
- Set sampling steps to 50 for the raw model
- Set CFG scale to 2.5 - The transcript notes that CFG 3 makes the face look "too plastic"
- Add a negative prompt when using the raw model to clean up unwanted artifacts
CFG Scale Impact
The CFG scale dramatically affects facial quality. Testing showed:
- CFG 2.5: Natural, realistic results
- CFG 3.0: Creates a "plastic" look on the face
Pro Tip
Keep your prompts simple and descriptive. You don't need complex prompts to get impressive results with Krea 2 Identity Edit. Simple prompts like "woman standing on a cliff wearing a black skirt" work well.
Negative Prompt Importance
The negative prompt is required when using the raw model (not the digital CFG version). It helps eliminate unwanted artifacts and improves overall image quality.
Step 4: Generate a Single-Person Image
- Run the workflow in ComfyUI
- Review the result - The transcript reports approximately 95% accuracy in preserving facial identity
- Check for preserved details: face, hairstyle, clothing details, and patterns (e.g., flower patterns on shirts)
Excellent Identity Preservation
The Krea 2 model preserves facial identity remarkably well. In testing, even small details like flower patterns on clothing were maintained. The transcript notes this is "one of the best" open-source identity editing models tested.
Pro Tip
For the best results, use the Euler sampler with the turbo model. The transcript reports excellent performance with this combination.
Step 5: Edit a Two-Person Image
- Enable the two-person editing section of the workflow
- Provide two reference images - one for each person
- Write your prompt (e.g., "photo of this woman standing next to a blue colored pipe")
- Generate and review - Note that facial consistency may be slightly less accurate than single-image editing
Two-Person Limitations
Identity editing with two reference images is more challenging. The transcript notes this limitation exists across all tested identity editing models (Gwen image, Flux context, Fire image, and others). Results are still "very usable" but facial consistency drops slightly.
Pro Tip
For the best two-person results, use images with clear, front-facing views and consistent lighting. Complex poses or angles increase the difficulty for any model.
Step 6: Use the Turbo Model for Faster Generation
- Load the Criya MXFP8 Turbo model instead of the raw model
- Change sampling steps to 8 (versus 50 for the raw model)
- Switch the sampler to Euler
- Keep remaining settings the same as the raw model setup
- Optionally test the ERHD scheduler - The transcript shows good results with this combination
Speed Comparison
The turbo model is significantly faster:
- Raw model: ~5-7 minutes on an RTX 4090
- Turbo model: ~40-50 seconds on an RTX 4090
Pro Tip
The turbo model is ideal for:
- Beginners who don't want to tweak CFG values
- Lower VRAM GPUs
- Rapid prototyping and testing different prompts
Turbo Model Quality
Even with only 8 sampling steps, the turbo model produces surprisingly accurate results. Clothing details, poses, and facial features are all well-preserved. The transcript specifically notes the "convincing" face quality with the turbo model.
Best Practices and Troubleshooting
Optimal Settings Summary
| Setting | Raw Model | Turbo Model | |---------|-----------|-------------| | Sampling Steps | 50 | 8 | | CFG Scale | 2.5 | 2.5 | | Sampler | Default | Euler | | Negative Prompt | Required | Optional |
Common Issues
- Plastic-looking faces: Lower CFG from 3.0 to 2.5
- Artifacts with raw model: Add a negative prompt
- Poor identity with two people: Use clear, front-facing reference images
- VRAM errors: Try the smaller model variants or the turbo model
Related Guides
- ComfyUI Installation Guide - Complete installation process and custom node setup
- Running ComfyUI on RunPod - Cloud GPU setup for faster generations
- Deploy Your First RunPod - Basic account and pod deployment
Next Steps
Now that you've mastered Krea 2 Identity Edit:
- Experiment with different prompts: Try changing backgrounds, outfits, and environments
- Test the two-person editing: Push the model's capabilities with complex scenes
- Optimize your settings: Fine-tune CFG values, sampling steps, and schedulers for your specific use case
- Explore other identity editing models: The transcript mentions Gwen image, Flux context, and Fire image as alternatives