🚗 汽车深度图

ControlNet depth is great until you realize you don't want all the depth - you want the depth of just the thing you care about, and everything else zeroed out. A depth map of a car sitting in a busy street is useless for a workflow that only wants to preserve the car's geometry; what you need is the car's depth with the background removed.

The display name tells you exactly what this node was built for: 🚗 汽车深度图, "car depth map." 🚗 Text Segmented Depth runs three models in sequence to produce a depth map that contains only your target: Depth-Anything-V2 estimates the depth, Grounding DINO finds your object by text, and BiRefNet produces the fine mask - and the outputs include both the masked depth and the pieces you can reuse.

The pipeline

1. Depth-Anything-V2-Small (~95 MB) produces a full depth estimate of the scene.
2. Grounding DINO uses your 文字提示 to locate the target. The default text is literally car, and the node's whole default setup is tuned for automotive shots.
3. BiRefNet (~1 GB) mattes out the target precisely, and the depth map is masked to that region - everything else goes to zero.

There's a shortcut built in: wire the optional 外部深度图 (external depth image) and it skips the internal Depth-Anything step entirely, using your depth map as the source. Useful when you already have a depth map from elsewhere and just want it masked.

The inputs

• 文字提示 - target text (default car). Leave empty and BiRefNet auto-grabs the salient foreground, same trick as the pack's Text Segmenter.
• 识别严格度 (default 0.35), 分割分辨率 (512–2048, default 1024), 羽化半径 (default 2), 反转遮罩 - the same detection and mask controls as Text Segmenter; if you've used that node, none of these need explaining twice.
• 外部深度图 - optional, skips the internal depth model.

The outputs

• 遮罩深度图 - the masked depth map, your main product.
• 深度图 - the full unmasked depth estimate (Depth-Anything output).
• 遮罩 - the segmentation mask on its own, so you can reuse it in other branches.
• 预览 - a visual check.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/LouisLU1997/ComfyUI-louis-use Louis_use
pip install -r requirements.txt

Or "Louis Use" in ComfyUI Manager, then restart. First run downloads three models - Depth-Anything-V2-Small (~95 MB), Grounding DINO Tiny (~340 MB), and BiRefNet (~1 GB) - roughly 1.4 GB total, cached on HuggingFace and reused on later runs. Plus timm via requirements.txt for BiRefNet. It's the heaviest node in the pack alongside Text Segmenter; give it disk and VRAM.

The honest take

Three models in one node is impressive, but it's worth asking whether you need all three. If you have any external depth map - which is common, since you may already run Depth Anything as a ControlNet preprocessor elsewhere in the workflow - the 外部深度图 shortcut lets you skip a model and a download's worth of work. The value-add of this node is really "depth + mask, aligned and in one place," and that's exactly what you want for a car-depth ControlNet pass or an object-focused depth composite. For anything else, the single-purpose pieces are cheaper.