Pixal3D: Free Pipeline

The most honest node in this pack: no inputs, no outputs, one job. It drops Pixal3D's cached pipeline and MoGe singletons out of VRAM. That's the whole thing, and it's a no-op if nothing's loaded - you can't break anything with it.

You might be tempted to skip it, but there are two genuinely good reasons to use it, and both come straight from how the pack behaves.

The fragmentation problem

Pixal3D is a ~14 GB resident pipeline, and it's a VRAM hog in a subtle way: the author has measured ten consecutive warm Image to Mesh runs succeeding and then the eleventh OOMing. The CUDA allocator gradually loses contiguous space across runs, so you get a crash that isn't your settings - it's just fragmentation. Queuing this node between batches resets the allocator and keeps the eleventh run from being the one that ruins your evening. If you're generating a lot of assets in one session, get in the habit: generate, free, generate.

Getting VRAM back from the hoarder

The pack's default is keep_warm: true on Image to Mesh, which deliberately parks the ~14 GB pipeline in VRAM so the next run is ~3 minutes instead of a 7–10 minute cold load. That's a great default until you're done with 3D for the day and want to run something else - then 14 GB of idle Pixal3D is just dead weight. You have two ways out: flip keep_warm to false on the generation node (it auto-frees at the end of that run), or leave the default alone and queue this node when you're switching tasks. The Free node is the non-destructive option: it doesn't change your workflow's settings, it just releases memory now.

Worth knowing: it's a release, not a reset. The next Pixal3D run pays the full cold-load again - weights into RAM, transfer to GPU, the whole 7–10 minutes on a fresh restart. That's the trade for getting the memory back, and it's usually worth it if you're not about to run more Pixal3D.

Installing

It ships with the pack, so the install is the pack's: ComfyUI Manager → search ComfyUI-Pixal3D → install → restart, or clone manually and run install.py with ComfyUI's Python:

cd $HOME\Documents\ComfyUI\custom_nodes
git clone https://github.com/dreamrec/ComfyUI-Pixal3D.git
cd ComfyUI-Pixal3D
& "$HOME\Documents\ComfyUI\.venv\Scripts\python.exe" install.py

The pack wants Windows 10/11 x64, Python 3.12 + PyTorch 2.8 + CUDA 12.8, an NVIDIA RTX 30/40/50 card (≥16 GB VRAM, 24 GB+ recommended), and ~50 GB of disk for the one-time ~26 GB model download. The wrapper is MIT, but the Pixal3D model underneath is academic / non-commercial and not licensed for EU use.

In the end this is a three-word answer dressed as a node: it frees VRAM. If you run Image to Mesh more than a couple of times per session, you'll be glad it exists - it's the closest thing the pack has to a "garbage collector" button, and it beats restarting ComfyUI to get your card back.