Load TripoSR Model

TripoSR is a real model, not a Mixlab invention - Stability AI and Tripo AI released it as a fast feed-forward image-to-3D network, and Mixlab wraps it so you can load it as a node instead of running their standalone demo. Load TripoSR Model is the first half of that pipeline: it loads the weights and hands you a TRIPOSR_MODEL object that a downstream Mixlab node (elsewhere in the pack's 3D category) turns into an actual mesh from a single input image.

What TripoSR is actually good for is worth being honest about, because the initial hype ran a little ahead of the results. When it launched, the community reaction split fast: some got clean, usable meshes in a couple of seconds; others - running the exact same official demo - got "terrifyingly bad" results with distorted geometry and an ugly backside on nearly everything they tried. The consensus that settled out over time is closer to "genuinely fast and genuinely rough" - great for a low-res base shape you can pose or use as a background/reference asset, not something you feed straight to a 3D printer or a game engine without cleanup. One practical workflow that came out of that: generate a rough shape with TripoSR, then run a depth model like Marigold over it to add detail before printing, rather than treating TripoSR's raw output as final. Expect roughly 6GB of VRAM for a single-image run - light by generative-AI standards, which is part of why it's a fun node to have around even with rough output.

The node itself has exactly one setting: chunk_size, an integer defaulting to 8192, with a range from 0 to 10000. This maps to TripoSR's internal marching-cubes chunking during mesh extraction - it's a memory/quality knob, not a creative one. Lower values use less VRAM per step but can leave more seams or holes in tight spots; higher values are smoother at the cost of more memory and a slower extraction pass. If you're on a lower-VRAM card and the node runs out of memory, dropping chunk_size is the first thing to try before giving up on it entirely.

Getting the model files in place is the one step people skip. Per the README, you need two separate downloads, not one: TripoSR's own model.ckpt from Stability AI's Hugging Face repo, placed at models/triposr, and - easy to miss - facebook/dino-vitb16, placed at models/triposr/facebook/dino-vitb16. That second one is TripoSR's image encoder (a DINO Vision Transformer), and the node will fail to load without it even though it's a completely separate download from the main checkpoint. If Load TripoSR Model errors out on a fresh install, check for that DINO folder first - it's the most common thing people forget.

Installing the pack itself is the usual path: ComfyUI Manager, search "comfyui-mixlab-nodes," install, restart - or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/shadowcz007/comfyui-mixlab-nodes.git
cd comfyui-mixlab-nodes
install.bat

(or pip3 install -r requirements.txt in a venv). Restart ComfyUI, then place both TripoSR downloads before you try to run the node - installing the pack alone gets you the node in the menu, not a working model.