Textured 3D from a Single Image: Building a TRELLIS2 Pipeline

Generating 3D models from images used to mean one of two compromises: fast output with no textures (you get a grey shell you have to hand-texture), or high-quality results locked behind commercial tools. That changed with TRELLIS2 — Microsoft's open-source, 4-billion-parameter model (MIT license) that turns a single image into a textured mesh with PBR materials. This article is the pipeline we actually built and battle-tested last week: FLUX → TRELLIS2 → Blender → Sketchfab.

The Pipeline at a Glance

FLUX image generation (text prompt)
        │  "floating in empty space, no shadow" — critical
        ▼
TRELLIS2 in ComfyUI (Windows, RTX GPU)
        │  sparse structure → shape → texture (three-stage SLAT generation)
        ▼
UnWrapAndRasterizer  ← the make-or-break node (bakes the PBR texture)
        ▼
TRELLIS2ExportMesh (.glb)
        ▼
Blender (MCP-driven from the RPi)
        │  center it, relight it, decimate safely
        ▼
Sketchfab / Unreal (Draco-compressed GLB, <100MB)

Step 1 — Generate the Source Image Right

TRELLIS reconstructs what it sees. If your source image has a shadow under the subject, the model will happily bake that shadow into the mesh as a flat base plane — a pedestal you didn't ask for. The single most important prompt trick we found:

"The ship is FLOATING IN EMPTY SPACE with NO SHADOW, NO GROUND PLANE, no surface beneath it, isolated against a transparent background, clean bottom silhouette"

Append a variant of that to every FLUX prompt. It costs nothing and eliminates the most common artifact we hit.

Step 2 — TRELLIS2 in ComfyUI

Installation (Windows)

Do not install TRELLIS2 by hand with pip install + git clone. The dependency matrix (xformers, cumesh, o_voxel, nvdiffrast, flex_gemm) must match PyTorch exactly, and manual installs pull the wrong versions with spectacular failures ("DLL load failed while importing _C"). Use ComfyUI-Easy-Install:

1. Extract it to a fresh folder (e.g. C:\ComfyUI-Trellis)
2. Run the Trellis2 add-on (it auto-switches to Torch 2.8.0+cu128 and FlashAttention)
3. One gated model to fetch: facebook/dinov3-vitl16-pretrain-lvd1689m — accept the terms on HuggingFace, then huggingface-cli download it into ComfyUI/models/facebook/
4. Add --listen to the startup batch file so other machines on your network can drive it

The Node Chain

For textured output the basic ExportMesh is not enough — it produces geometry only. You must run the texture through UnWrapAndRasterizer:

LoadImage → Trellis2LoadModel → Trellis2MeshWithVoxelGenerator
         → Trellis2UnWrapAndRasterizer → Trellis2ExportMesh

UnWrapAndRasterizer unwraps the mesh and bakes the generated texture into a UV map. Miss it and you get a clean-looking mesh with zero color — the classic "outputs empty {}" trap.

Settings That Actually Matter

One empirical finding from a week of ship models: for clean geometric shapes (spaceships), the 512 pipeline with the sdpa backend produced fewer artifacts than 1024 + flash_attn. Detail isn't always the win — it amplifies defects too. For organic subjects, 1024 is the safer default.

Step 3 — Blender: Learn What Not to Do

TRELLIS outputs a non-manifold shell — a single layer of faces, not a closed volume. Every instinct you have from hard-surface modelling will betray you:

Safe operations only: vertex centering (shift geometry to origin — no topology change), viewport framing, lighting, and material toggles. TRELLIS also imports the model off-center under a world empty — clear the parent and shift vertices so the geometric center sits at the origin, otherwise it floats ~25cm off in each axis.

Swap the default Blender light for a 3-area-light rig (Key + Fill + Rim) — the difference in presentation is night and day, and it matters for previews, renders, and Sketchfab thumbnails.

One Blender 5.2 quirk: imported GLB materials default to Alpha Hashed blend mode and snap back if you try to set OPAQUE via the Python API. Fix it manually in the UI: Material Properties → Settings → Alpha Clip, clip threshold 0.01.

Step 4 — Shrinking It for Sketchfab (100MB limit)

A 1024-pipeline TRELLIS export with 4096px textures can be 200–350MB. The free Sketchfab tier caps at 100MB. Our working recipe:

1. Planar Decimate (Dissolve mode, 10° angle, UV delimiting on) — removes only coplanar faces, keeps textures intact, no holes. Collapse mode would tear the shell.
2. Merge by Distance (threshold 0.0001) — the dissolve leaves millions of orphaned vertices; merging them drops file size without touching geometry.
3. Export with Draco compression (level 6, position 14 / normal 10 / texcoord 12, JPEG textures at 80%).

Real result on a spaceship: 6.2M polys + 4096px textures (~350MB) → 1.4M polys + 1024px + Draco (~30MB) — a 10x reduction, well under the limit, and Draco decompresses client-side in the browser with no visible quality loss.

The Alternatives We Evaluated

Because why not ask "is there something better?" — we tested the field:

The takeaway: for textured output, TRELLIS2 is still the best open option we found. TripoSG is the one to keep an eye on for geometry-heavy work; Modly's value is convenience, not quality.

What's Next

• Unreal + Nanite: the 1536_cascade pipeline outputs are the ones to feed Nanite — film-quality geometry that still real-time renders
• Automate the loop: upload image → TRELLIS2 → auto-optimize → web-embed, so we never touch Blender for routine exports
• Embed on the web: Sketchfab's embed viewer is the easy path; self-hosted <model-viewer> GLB embeds are the lightweight alternative

The full workflow files, exact node JSON, and the Blender recipes live in our trellis2-3d-pipeline skill — if you're building on Windows with an NVIDIA GPU, that's a complete copy-paste starting point.