Krea2 ControlNet img2img ⚡

Plain Krea2 img2img changes a photo's pixels but leaves you at the mercy of the prompt for where things go. This node is the version that lets you steer the structure - depth, edges, pose - the way you'd use a ControlNet on an SDXL-era model. Except Krea2 doesn't have a ControlNet architecture; it has Control LoRAs, and the wiring is different enough that a dedicated node is worth having.

Here's the mental model. Krea2's Control LoRAs come in two flavors, and the pack's Krea2 Control LoRA Loader auto-detects which you've got: the widened-projection kind (trained to accept image tokens plus control tokens through the DiT's input projection) and the ordinary in-context kind. What nothing in the LoRA file tells you is what type of control signal it wants - depth, canny, something else. That's what this node's control_mode is for.

What it does

Everything Krea2 img2img does - prompt encoding, VAE-encoding your source photo, building the init latent - plus attaching a control image to the loaded Control LoRA. The control image gets resolved three ways:

• auto_depth (default) - derives a depth map from your images automatically using Depth Anything V2. Weights auto-download into models/depth_anything_v2/ on first use. Right for the depth Control LoRA (e.g. the depth-control-lora that circulates from the Krea-2-depth-controlnet release).
• auto_canny - plain cv2.Canny edge detection, no model, no download. Right for canny checkpoints.
• manual - no derivation; connect control_image yourself. Use it for pose/lineart/normal LoRAs the two auto modes don't cover - those preprocessors come from the separate ComfyUI-ControlNet-Nodes package, and you wire their output in here.
• none - skip control entirely for this call, even with a Control LoRA loaded. Handy for A/B-testing without rewiring the graph.

Connecting control_image explicitly always overrides auto-derivation. Two small companion settings: control_channel_mode (grayscale for depth, rgb for canny/pose) and control_normalize (per-image minmax for depth, none for edges).

The rest of the knobs

strength (default 0.6) is your img2img dial - how much of the init image to discard. width/height are output size; with images connected they resize them. Outputs are model, positive, negative, latent, denoise, which feed a stock KSampler directly. Minimal graph: Krea2 Model Loader → Krea2 Control LoRA Loader → this node → KSampler → VAE Decode.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/ChrisColeTech/ComfyUI-GGUF-Loader
pip install --upgrade gguf

Restart, drop your Krea2 model, Qwen3-VL-4B encoder, VAE and Control LoRA into the usual model folders. Node is under 🤖 CCTech/Krea2.

Gotchas

The guard rails are worth knowing. A Control LoRA loaded with no usable control image (and control_mode not none) raises immediately - it refuses to silently sample a half-configured model. The reverse is fine: control_image connected with no Control LoRA just gets ignored with a warning, so you can leave a preprocessor chain wired while you toggle the LoRA loader. Both are deliberate, and both save you from the "why is my output ignoring the depth map" class of debugging.