DownloadAndLoadDepthCrafterModel

The DepthCrafter pack has exactly two nodes, and this is the one that does the unglamorous setup work: it downloads the model, drops it into your models folder, loads it into VRAM, and hands you a ready-to-run pipeline. It's the node that never appears in finished workflows - you wire it in, set it once, and feed its output to the DepthCrafter render node. If you're getting "missing model" errors or wondering why your first run takes forever, this is where the answer lives.

Why it downloads more than you expect

Here's the part that surprises people. DepthCrafter is Tencent's video-native depth estimator, and it isn't self-contained - it's built on top of Stable Video Diffusion's img2vid-xt pipeline. So this node pulls down two separate model repos into ComfyUI/models/depthcrafter/:

• tencent_DepthCrafter/ - the actual DepthCrafter UNet (a config.json plus one safetensors)
• stabilityai_stable-video-diffusion-img2vid-xt/ - the SVD machinery it leans on: image encoder, VAE, scheduler, and feature extractor

It grabs these via the Hugging Face Hub on first run, pinned to specific revisions, then caches them so every run after that is instant. All of it loads at fp16. That's several gigabytes of downloads the first time, so budget for it - and for the load itself, which is the slow part on a first run.

The only two inputs, and the one rule

enable_model_cpu_offload         (default: true)
enable_sequential_cpu_offload    (default: false)

That's the whole interface. These control how aggressively the pipeline offloads to CPU:

• enable_model_cpu_offload - shuttles model components to CPU when they're not actively needed. The author's framing is roughly +25% VRAM saved for a small speed hit. Default on, and a sensible default it is.
• enable_sequential_cpu_offload - moves everything to CPU in sequence, saving the most VRAM (about +37%) but noticeably slower.

The one rule: enable only one at a time. Tick both and you're fighting yourself over the same VRAM. Tick neither and the whole pipeline just sits on the GPU - fine if you've got the headroom, a bad idea on an 8GB card.

Output and what it wires into

The node outputs a single depthcrafter_model (type DEPTHCRAFTER_MODEL) - an opaque handle to the loaded pipeline, not something you can inspect. Plug it straight into the DepthCrafter node's depthcrafter_model input and you're done. Nothing else consumes it.

Installing the pack

Two ways, same result. Via ComfyUI Manager: search "DepthCrafter" and install the entry called "DepthCrafter Nodes". Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/akatz-ai/ComfyUI-DepthCrafter-Nodes.git

Then restart ComfyUI. Its actual dependencies are light for a video pipeline: torch, diffusers>=0.29.1, and accelerate>=0.17.0 - you almost certainly have the first two already, and Manager handles the rest. The author also ships a prebuilt Docker-style environment if you use their Environment Manager, but that's optional.

Where people get burned

• First run needs internet. If the auto-download fails, the README has a full manual-download route with the exact file list and folder layout - the subdirectories inside the SVD folder matter, don't flatten them.
• Both offload toggles on. Pick one.
• The model license is non-commercial academic only. Fine for personal experiments, read it before you put anything commercial on top.

Nothing here is clever, and that's the point. It's a download-and-prepare node that does its job, and its job is the prerequisite for the actual interesting part of the pack.