FLUX Depth Preprocessor (Deprecated)

Its own display name says the quiet part: FLUX Depth Preprocessor (Deprecated). So lead with that - this node is deprecated, and if you're building a depth workflow today you want a dedicated depth pack, not this. It's documented here for anyone maintaining an old graph that still calls it.

What it did: take an IMAGE, run a depth-estimation model over it, and output a depth map IMAGE - a grayscale image where brightness encodes distance (white close, black far). That map then feeds a Flux depth ControlNet to constrain a new generation to the same spatial layout. It shipped alongside Nunchaku's early Flux ControlNet support as a convenience preprocessor.

How it works

Depth estimation turns a flat photo into geometry: one image in, a usable depth map out, no stereo rig or lidar needed. This node wrapped a depth model so you could generate that map inside the same graph as your Nunchaku Flux, rather than bouncing to a separate preprocessor pack. The depth map doesn't have to match your generation exactly - it encodes relationships (foreground vs background), which the ControlNet uses to place your new content in the same spatial structure.

Why it's deprecated: depth preprocessing is a solved, well-served corner of the ecosystem, and it makes more sense to lean on the packs that specialize in it than to keep a bespoke one in Nunchaku. Those packs carry the whole modern lineup - Depth Anything v2, Marigold, and friends - with far more options.

The inputs and outputs that matter

• image (IMAGE) - the source photo to extract depth from.
• model_path - the depth model to run. Empty until you have a compatible model downloaded.

Output is an IMAGE - the depth map, which you wire into a depth ControlNet apply node.

How to install it

It's bundled with the ComfyUI-nunchaku pack. ComfyUI Manager → search "ComfyUI-nunchaku" → install, or

cd ComfyUI/custom_nodes
git clone https://github.com/mit-han-lab/ComfyUI-nunchaku

then restart. It needs a depth model file present to populate model_path.

Common issues & troubleshooting

Use a real depth pack instead. For anything new, reach for comfyui_controlnet_aux (Fannovel16) or Kijai's ComfyUI-DepthAnythingV2. Depth Anything v2 Large is the community's default for the quality/speed balance; Marigold when you want maximum sharpness and don't mind the wait. All of them slot into a Flux depth ControlNet the same way.

Empty model_path. No depth model downloaded where this node looks. A dedicated depth pack handles model downloads more cleanly - another reason to migrate.

Depth map looks soft / ControlNet won't follow it. For ControlNet, you want sharp edges and clear local contrast more than metric accuracy. A crisper model (Depth Anything) tracks better than a gradient-heavy one - a lesson worth carrying to whatever pack you move to.

ControlNet strength. Once you have a map, weight 0.4–0.7 gives loose guidance that preserves creativity; 0.8–1.1 locks the layout tightly. Ending the ControlNet around 40–60% of the way through the steps keeps the composition while freeing detail.