Krea2 Depth Map ⚡

Krea2 is the best open base model of its window, and its most popular control option is the depth Control LoRA - a LoRA that makes Krea2 follow a photo's geometry when you edit or re-frame it. Before that depth map can reach the model, something has to derive it from your source photo. Krea2 Depth Map is that something: image in, depth map out, via Depth Anything V2, made for feeding into Krea2Img2Img's control_image input (or control_mode=manual).

Why you'd reach for it

The one-node Krea2Img2Img path already has control_mode="auto_depth", which derives the depth map in-line with the same model - so for the single most common graph you don't need this node at all. This one is for when you want the map visible and inspectable first: build the graph by hand, preview the depth before it touches the sampler, reuse the same map across several generations, or feed the map into something else entirely. It's the explicit-node convention, mirroring how the pack exposes its Flux Klein depth detector. If auto_depth is the express lane, this is the manual gear lever.

The inputs that matter

• image - the source photo. Anything you'd want structure from.
• ckpt_name - Depth Anything V2 size: vits (smallest/fastest) → vitg (largest/slowest), default vitb. First use auto-downloads into models/depth_anything_v2/.
• resolution - processing resolution, default 512.

One output, IMAGE - the depth map, ready for control_image on Krea2Img2Img.

How it works

Depth Anything V2 (DINOv2 encoder + DPT decoder head) has been the ecosystem's default ControlNet depth preprocessor since 2024 precisely because it reads relative depth off real photos without complaint - near objects white, far objects dark - which is exactly the signal a depth Control LoRA was trained on. This port is consolidated from comfyui_controlnet_aux into the pack's single-file convention, and the weights self-download, so there's no second install step and no HuggingFace runtime in your workflow.

Installing it

It ships inside the ChrisColeTech pack with all the other Krea2 nodes:

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

Restart; it's under 🤖 CCTech/Krea2. Nothing extra to download by hand - the checkpoint pulls itself on first run.

Gotchas

Two things actually bite people. First, depth convention: if your result looks like the model thinks near is far (background popping forward), the preprocessor's convention is flipped relative to your LoRA's training - flip control_invert on the img2img node, don't blame the map. Second, and this is the one that surprises people coming from ControlNet: not every "Krea2 depth/canny LoRA" you find is a Control LoRA. The depth LoRA is the widened-projection kind that takes a control_image; other Krea2 LoRAs with canny in the name are ordinary in-context LoRAs that take an edit_reference instead. If you wire a depth map into the wrong mechanism you get nothing at all - Krea2Img2Img will tell you, because it raises when a Control LoRA is loaded with no usable image and ignores a control image with no Control LoRA, but only if you're using the node the map is meant for.