🔧 Dependency Installer

Here's the thing nobody tells you about the ArchAi3D Qwen pack: it's enormous, and its requirements.txt lists everything - SAM3's segmentation stack, Metric3D's deep-learning toolkit (opencv, timm, scipy, matplotlib), colour-science, google-genai, gdown. Install all of it and you're pulling in a small conda environment just to run one or two nodes. ArchAi3D_Dependency_Installer is the pack's answer: a node that installs dependencies in groups, so you only get what the nodes you actually use need.

How it works

You pick an action from a dropdown and run the graph:

• Check System - inspects your environment and reports what's installed, missing, and what GPU/CUDA/PyTorch it detected. Run this first; it tells you what you're in for.
• Install Core - PyYAML, numpy, Pillow, requests, tqdm. The pack's basics.
• Install SAM3 - the segmentation stack (sam3, einops, ftfy, pycocotools, scikit-image, scikit-learn, pandas, iopath).
• Install Metric3D - huggingface_hub, opencv-python, matplotlib, timm, scipy, addict, yapf.
• Install Color Tools - colour-science (for the BT.709 color correction node).
• Install API Tools - google-genai and gdown (Gemini + Google Drive nodes).
• Install All - everything above.
• Upgrade PyTorch (Blackwell) - a special action for RTX 50-series cards, upgrading to the PyTorch 2.8+ / CUDA 12.8+ combo that Blackwell needs.

There's an include_optional toggle (default True) for optional extras. Output is a status string summarizing what happened.

The reason a node like this exists: the author builds for RunPod-style serverless deployments where you want to auto-install deps on a cold start. It's also the node to run if ComfyUI Manager's dependency install failed partway and left you with a half-working pack.

The real use

You almost never need Install All. Map it to what you run: using SAM3_Segment? Install SAM3. Using Metric3D_Depth? Install Metric3D. Using only the Qwen prompt-builders? Core is enough. Every optional group you skip is a version of opencv or scipy you aren't fighting with later. ComfyUI's "one node pack, one dependency hell" reputation (the KB's ecosystem essay is blunt about it) is exactly what this node is dodging.

Caveats

Two. First, it shells out to pip from inside ComfyUI - that's normal for installer nodes, but it means you're letting a custom node run pip as your user. The pack is from a single indie author with a dual-license model, and while nothing here is suspicious, the habit of "know what your installer installs" is a good one; the ecosystem has had real supply-chain incidents. Second, Upgrade PyTorch (Blackwell) is a big deal - it changes the torch install under your ComfyUI and can break other packs. Only run it if you actually have a 50-series card and know your environment can take the change.

Install it (irony-free)

The node itself is in the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/amir84ferdos/ComfyUI-ArchAi3d-Qwen.git

No requirements needed for this node - it's pure stdlib. Restart ComfyUI, find it under ArchAi3d/Utils, and run Check System before anything else.