☁️BizyAir Depth Anything V2 - Relative

Depth Anything V2 has been the default depth preprocessor since 2024, and unusually for this fast-moving field, nothing has displaced it - community mentions have held steady at 8–19 a month for two and a half years running, which is about as settled as a tool gets around here. It's also a genuinely rare case of ByteDance giving something away: the flagship generation products (Seedream, Seedance) stay locked behind an API, but Depth Anything is one of the components ByteDance open-sourced outright, and it's now load-bearing infrastructure across the whole ecosystem. This node runs it on BizyAir's cloud instead of your GPU.

What it does

image is the only required input - hand it a photo, get back a grayscale depth map where distance from the camera is encoded as brightness. That map is what you feed into a depth ControlNet slot afterward (this pack's depth_image input on the Controlnet Union pipeline, or depth on BizyAir SetUnionControlNetType, are the natural next stops).

ckpt_name picks the model size: depth_anything_v2_vitg (giant), vitl (large, the default), vitb (base), and vits (small) - a straightforward accuracy-versus-speed ladder, largest to smallest. Large is the community's settled default and a safe choice unless you have a specific reason to trade accuracy for speed. resolution (64–1024, step 64, default 512) controls the preprocessing resolution - higher pulls out finer depth detail at the cost of more compute; 512 is a solid middle ground for most ControlNet use.

Output

A single IMAGE - the depth map itself, ready to feed straight into whichever depth-conditioned node comes next in your graph.

Installing it

ComfyUI Manager: search BizyAir, install, restart. Manual:

cd ComfyUI/custom_nodes && git clone https://github.com/siliconflow/BizyAir.git

Restart afterward. No model checkpoint to download regardless of which ckpt_name you pick - all four sizes live and run on BizyAir's infrastructure, which is the actual value of routing this through the cloud rather than ComfyUI's stock Depth Anything preprocessor: you skip the download entirely. A BizyAir account with an API key set is still required before any node in the pack executes.

Where people get burned

Expecting metric depth. This produces relative depth - distance ranked against other pixels in the same image, not real-world units. That's exactly what ControlNet wants, but if you were hoping for something you could measure with, it's the wrong tool; metric accuracy isn't the goal here.

Running it on video frames and expecting temporal stability. Depth Anything is an image model - feed it a video frame-by-frame and you'll likely see flicker between frames, because nothing enforces consistency across them. Video needs a video-native depth model, not this preprocessor run repeatedly.

Cranking resolution to max by default. 1024 isn't free - it costs more compute per call for detail gains that often don't matter once the depth map gets fed into a ControlNet at typical strength. Start at 512 and only go higher if you can see the difference matters for your specific shot.