InstantMeshRun

This is the node that does the party trick. You feed it one image of an object, and it hands you back a 3D mesh - geometry, UVs, and a texture map - plus a rotating preview video. It's TencentARC's InstantMesh, the sparse-view Large Reconstruction Model, wrapped into a single ComfyUI node, and it's genuinely one of the more satisfying things to watch run on this ecosystem. It's also from an archived pack, so read the setup notes below before you clear your weekend.

What it actually does

The pipeline is three stages, and understanding them explains every input on the node:

1. Background removal. Unless you turn it off, the input image goes through rembg (the u2net default) to cut the object out and resize the foreground. Clean background in, clean silhouette out.
2. Multiview generation. The trimmed image goes through a zero123plus diffusion pipeline (loaded from sudo-ai/zero123plus-v1.2) which imagines the object from several angles at once - a 3×2 grid of six views. This is the slow, VRAM-hungry part: it's 75 diffusion steps by default.
3. Reconstruction. Those views plus fixed camera poses go into the model the Loader gave you. It builds a triplane representation, then extracts a mesh via FlexiCubes, exporting an OBJ with an MTL material and a 1024×1024 texture PNG if you asked for the texmap.

The images output is the multiview grid itself (handy as a preview or for a sanity check); mesh_path and video_path are strings pointing at the saved .obj and .mp4 on disk. Files land in an output folder inside the pack's own directory, ComfyUI/custom_nodes/ComfyUI-InstantMesh/, not in ComfyUI's normal output folder - the single most common "where did it go" moment.

The inputs that matter

model and config come straight from InstantMeshLoader - wire both. After that, the ones you'll actually touch:

• images - an IMAGE input; feed it a LoadImage or anything that produces a clean image.
• diffusion_steps - default 75. Lower it toward 30-40 for quick drafts; quality genuinely suffers, but it's a fine tradeoff while iterating.
• view - 6 or 4 (the range is locked to those two). 4 uses a subset of the multiview poses and is faster but more fragile.
• export_texmap - keep on true. This is the difference between a colored mesh and a properly textured one you can take into Blender. Off gives you per-vertex colors only.
• save_video - the rotating turntable render. Great for showing off, but it renders 120 frames at 512×512 and can add a couple of minutes; leave it off until a result is worth it.
• remove_bg - on by default and usually what you want. If your input already has a clean white background, you can flip it off and shave some time.

Install - the part everyone underestimates

This pack is a two-layer dependency stack: it requires MrForExample's ComfyUI-3D-Pack to be installed and working first, plus Visual Studio Build Tools on Windows. Both repos ship install_windows_portable_win_py311_cu121.bat installers that you run from ComfyUI/custom_nodes/ while ComfyUI is stopped. Community reports put this whole stack at the top of the "hardest nodes to install" list, and one user on the release thread said following the instructions "completely effed" their install. Back up your ComfyUI folder before you try, seriously.

On the plus side: the release thread has someone confirming it runs on an 8 GB GPU, and the models auto-download from Hugging Face on first use - the loader grabs the .ckpt, and the run node pulls the zero123plus weights and a custom white-background UNet from TencentARC/InstantMesh. No manual model hunting, but budget a few GB of disk.

Where people actually hit walls

• Install failure is the #1 issue, and it's almost always the 3D-Pack half, not this node. If the Loader errors with missing imports, the 3D-Pack build didn't finish.
• VRAM. The diffusion stage plus the reconstruction model in fp32 can exceed 8 GB on fiddly inputs. Drop diffusion_steps, use view 4, or pick instant-mesh-base in the Loader.
• Messy inputs. rembg is good, not magic - flyaway hair and semi-transparency still trip it up, and what zero123plus imagines becomes what you reconstruct. Feed it a solid, well-lit object.
• The output mesh is not game-ready. It's a great starting point; for 3D printing or real use you'll still clean it up in Blender.

And the honest disclaimer: this repo is archived because ComfyUI-3D-Pack now supports InstantMesh natively. If you haven't committed to this pack yet, take the modern path. If you're already here, it works - just treat install day like a small project.