Trellis Image to 3D 🖼️➡️🎲

You hand this node a photo of a thing, and it hands you a 3D model you can spin, export as a .glb, and drop straight into Blender, Unity, or your game. No multi-view renders, no photogrammetry rig - one image in, mesh out. It's the payoff node of the IF_Trellis pack. The checkpoint loader is the toll booth; this is the ride.

Under the hood it's Microsoft Research's TRELLIS (arXiv 2412.01506), wrapped into ComfyUI by ImpactFrames-YT. TRELLIS's trick is its SLAT representation - a "structured latent" that can be decoded into whatever 3D format you want: a radiance field, 3D gaussians, or a mesh. The node runs the whole pipeline in two rectified-flow passes. First a sparse-structure sampler lays down a coarse voxel skeleton; then the SLAT sampler fills in geometry and appearance. What comes out is decoded to both a gaussian and a mesh, which are baked into the .glb you save.

That's why the sampling inputs come in pairs. The ss_guidance_strength / ss_sampling_steps pair (defaults 7.5 and 12) controls the structure pass; slat_guidance_strength / slat_sampling_steps (defaults 3 and 12) control the latent pass. The defaults are genuinely sane - before you touch them, try just changing seed to get variants. Only reach for more steps when geometry looks blobby or detail is mushy.

The inputs a beginner actually sets:

• mode - single for one image, multi if you feed several reference angles of the same object. Multi is where TRELLIS shines: back and side views fix the "mushroom on legs" failure where the model never knows what's behind the object. multimode (stochastic vs multidiffusion) only matters in multi mode.
• images (and optional masks) - the source picture. The pipeline preprocesses it, and that's where the rembg dependency earns its keep: it cuts the subject out. A clean subject on a simple background gives you a much better mesh than a cluttered scene.
• texture_mode - blank, fast, or opt. opt bakes a higher-quality texture with fewer visible seams, at the cost of time. Start on fast; switch to opt when the seam work bothers you.
• mesh_simplify - a simplification ratio, so the tooltip is backwards-ish: lower means more polygons, 1.0 means aggressive simplification. Keep 0.95 as a starting point; only drop it when you need dense geometry for a hero asset.
• save_glb is on by default and is the point. render_video gives you a rotating preview .mp4 (the fps input controls smoothness), save_gaussian drops the .ply, and save_texture / save_wireframe export extra maps. The wireframe map is oddly useful as a retopo reference.

The node is an output node, so it returns three things: model_file (a path to the .glb), video_path (path to the preview), and texture_image (the baked texture as a ComfyUI IMAGE). Files land in ComfyUI/output/<project_name>/.

Install is the real wall, and it's shared across the whole pack: you need the --recurse-submodules clone, a micromamba env, Visual Studio Build Tools on Windows, CUDA 12.4, and a bunch of compiled extensions (nvdiffrast, diffoctreerast, kaolin, spconv, SageAttention). The README says it plainly: not tested with the portable ComfyUI embedded Python. You'll also want an NVIDIA card with 8GB+ VRAM. If that all sounds like a weekend project - it is. The first run also quietly downloads the TRELLIS-image-large model (several GB) and a DINOv2 encoder, so don't panic when nothing happens for a while.

Where people get burned: a bad cutout. If rembg fumbles the background, the mesh inherits the mistake - fix the source image first. And if the whole install feels like too much, know that a lighter TRELLIS2 wrapper with prebuilt wheels and PBR support exists by now; this pack is the original, and the original has opinions about your environment.