Krea 2 Depth ControlNet Released: Depth-Conditioned Generation With ComfyUI Support

Community developer Tanmay Patil has released the Krea 2 Depth ControlNet, a ControlNet-LoRA that adds depth-conditioned generation to Krea 2. The single 862MB LoRA file lets you control the 3D structure of generated images while preserving full creative freedom over content and style.

The model is available on Hugging Face: Patil/Krea-2-depth-controlnet.

How It Works

Given any input image and a text prompt, the ControlNet extracts a depth map using Depth-Anything-V2, then generates a new image that preserves the same 3D structure and composition. The depth consistency (Pearson correlation between input depth and the depth of the generated output) reaches 0.98 with no prompt and 0.99 with a prompt, meaning the geometric structure is faithfully carried over.

The ControlNet is trained on Krea 2 Raw and works with both the Raw (28-52 steps, CFG 3.5) and Turbo (8 steps, no CFG) base models interchangeably. A LoRA scale dial lets you adjust the strength of depth adherence — from strict structural reproduction at 1.0 to more creative freedom at lower values.

Usage

The repository provides a simple inference script:

# Turbo base — fast, recommended (8 steps, no CFG)
python inference.py photo.jpg -p "a futuristic spaceship interior, cinematic lighting" \
    --lora depth-control-lora.safetensors

# Raw base — maximum quality (28-52 steps, CFG 3.5)
python inference.py photo.jpg -p "..." --lora depth-control-lora.safetensors --base raw

# No prompt: depth map is the only signal
python inference.py photo.jpg --lora depth-control-lora.safetensors --save-strip

# Weaker structure adherence
python inference.py photo.jpg -p "..." --lora depth-control-lora.safetensors --lora-scale 0.6

The --save-strip flag outputs a side-by-side comparison of the input image, extracted depth map, and generated output for easy evaluation.

ComfyUI Integration

The ControlNet is fully compatible with ComfyUI via the comfyui-krea2-controlnet custom node by facok. This node provides a straightforward interface for loading the depth ControlNet LoRA and applying it within the Krea 2 workflow.

To use it in ComfyUI:

1. Download depth-control-lora.safetensors from the Hugging Face repository
2. Install the comfyui-krea2-controlnet custom node
3. Load the LoRA through the custom node's interface and connect it to your Krea 2 workflow
4. Provide an input image as the depth reference

File Details

The base model stays frozen — only the LoRA weights are loaded, making it efficient to switch between different ControlNet modes.

License

The model is released under the Krea 2 Community License, consistent with the broader Krea 2 ecosystem.