TRELLIS.2

TRELLIS is Microsoft's 3D generation line, and TRELLIS.2 is the newest generation: take a single image, get back a textured, physically-plausible 3D mesh. Locally, that means downloading a multi-gigabyte model, satisfying a pile of Python dependencies, and waiting minutes on a card with serious VRAM - which is why most ComfyUI users never get past the idea. This node runs TRELLIS.2 on Runware's cloud and hands you back a finished .glb file saved to your output folder. If you've ever wanted "turn this photo of an object into a 3D model I can open," this is the lowest-friction path there is.

The image-to-mesh input is required - this node is the image-conditioned path (unlike Hunyuan 3D in the same pack, which also does pure text-to-3D). You give it an image, it gives you geometry, materials, and UVs in one shot.

How it works

It sends a 3dInference request to Runware, which returns a GLB. The node downloads it (named runware_<random>.glb) into your ComfyUI output folder and outputs a STRING file_path. There's a lot of hidden machinery under TRELLIS.2 - it generates sparse structure, then refines shape and texture in separate stages - and the widget list here is basically the exposed knobs for that pipeline. Most of them you will never touch, which is fine; the defaults are sensible.

The inputs that actually matter

• image (required) - the reference. One clear, well-lit shot of the object you want. The cleaner the input, the better the mesh; busy backgrounds make it guess.
• meshFile - the sneaky good one. It takes a UUID or URL of an existing 3D model and re-textures it using your reference image, skipping geometry generation. If you already have a mesh and want new textures from a photo, this is a genuinely useful second feature hiding in the widget list.
• settings.geometryOnly - off by default; turns it on to skip texturing and get raw geometry. Good for blockout or if you'll re-texture later.
• settings.resolution - 512 / 1024 / 1536 voxel grid for generation. Higher = finer geometry and slower. 1024 is a reasonable default; drop to 512 for quick iterations.
• settings.decimationTarget - target polygon count (5000–1,000,000, default 500,000). Lower it for a lighter model - a game engine will thank you - at the cost of geometric detail.
• settings.dracoCompression.enabled - compresses the output GLB geometry with Draco (plus level, quantizationPosition, etc.). Flip this on if file size matters (web delivery, sharing); leave off if you'll keep editing the mesh.
• settings.remesh - on by default; produces cleaner topology. The remeshBand / remeshProject sub-knobs are for tuning that pass. Generally: leave remesh on unless you're chasing every last polygon.
• settings.textureSize - 1024–4096 texture map resolution (default 2048). 4096 if the texture detail matters; 1024 for quick versions.
• settings.textureFormat - WEBP (default) or PNG. WEBP is smaller; PNG if you need lossless.
• settings.steps and the guidanceStrength / guidanceIntervalStart / rescaleT trio for shapeSlat, sparseStructure, and texSlat - the three sub-stages each have their own denoise steps and guidance. This is deep-end tuning; most people should leave these alone. steps at 12 each is the default and fine.
• seed and numberResults (1–4) - take a couple of seeded variations; 3D quality varies a lot between runs and picking the best is the normal workflow.

The output is file_path - a path string to the saved .glb.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

Or ComfyUI Manager → search Runware → install → restart. API key via Settings, RUNWARE_API_KEY, or runware auth login. No local models - that's the point.

Honest gotchas

3D generation is the slowest thing in this pack; a cloud TRELLIS.2 run is not a quick job, and multi-stage pipelines mean the "Rapid"-style speedups don't apply. Budget time and expect the run info to show a real cost. Second, the output is a file path, not an image or video tensor - you'll find the .glb in your output folder and open it in any mesh viewer or Blender. And treat the image input seriously: a cluttered or low-contrast reference will give you a lumpy result, and no amount of settings tuning fixes a bad input.