trellis2-gguf

name: trellis2-gguf description: "Trellis 2 GGUF ΓÇö local image-to-3D pipeline in ComfyUI on 6GB+ VRAM. Generates full 3D meshes with textures + UV unwrap from a single image. Use for characters, props, hard-surface assets, and architectural exteriors. NOT for room interiors, text-to-3D, or scene composition." tags: [3d, ai, trellis2, gguf, comfyui, image-to-3d, blender, local-ai, low-vram] skill_group: "3d-graphics" skill_category: "3D Graphics & Rendering"

Trellis 2 GGUF ΓÇö Local Image-to-3D

Open-source image-to-3D pipeline running fully locally, free, in ComfyUI. Generates textured 3D meshes with UV unwrap from a single input image. GGUF quantization allows 6GB VRAM GPUs (RTX 3050/4060 class) to run a 4B-parameter model ΓÇö faster than the full uncompressed version.

Research note: Full video analysis in research/videos/2026-04-15_trellis2-gguf-6gb-vram-local-3d.md

When to Use

• Converting a reference image (photo, concept art, sprite) into a 3D mesh with textures
• Generating hero props, characters, or architectural exteriors for game scenes
• Rapidly prototyping 3D assets for import into Blender or game engines
• Replacing or supplementing paid cloud services like Meshy for image-to-3D tasks
• Running 3D generation locally on 6ΓÇô16GB VRAM consumer GPUs

When NOT to Use

• Interior rooms, floor plans, or scene layouts ΓåÆ use SketchUp or manual modeling
• Text/prompt to 3D ΓåÆ use Meshy, Tripo3D, or 3DAIStudio (cloud)
• Rigging or animation ΓåÆ downstream Blender/Maya step, not this tool
• Production-ready topology ΓåÆ always retopo output in ZBrush, 3DA Studio, or Blender

VRAM & Quantization Guide

Key fact: GGUF Q4 uses ~half the VRAM of the full model and runs ~2 minutes faster.

Prerequisites

- Nvidia GPU with 6+ GB VRAM (CUDA required)
- Windows 10 or 11
- ComfyUI Easy Install (installed first)
- Trellis 2 default installation (prerequisite for the GGUF fork)

Installation

Step 1 ΓÇö Install GGUF Fork

1. Download install_trellis2_gguf.bat from pixel-artistry.com/trellis2gguf
2. Drop into your ComfyUI add-ons folder
3. Double-click ΓÇö installs the ArrowX GGUF ComfyUI fork + all dependencies
4. Wait for completion (internet speed dependent, ~5ΓÇô10 min)

Step 2 ΓÇö Download GGUF Model Weights

1. Download download_trellis2_gguf.bat from the same link
2. Drop into the add-ons folder and run
3. Auto-downloads all quantized model files, encoders, decoders, configs
4. Grab a coffee ΓÇö model download takes time

Step 3 ΓÇö Load ComfyUI Workflow

1. Launch ComfyUI ΓÇö ignore any red errors at startup (model path resolution, non-fatal)
2. Load the GGUF workflow from pixel-artistry.com/trellis2gguf
3. Do NOT reuse old Trellis 2 workflows ΓÇö GGUF uses different nodes

ComfyUI Workflow Structure

The GGUF workflow replaces the standard Trellis 2 Model Loader with the GGUF Load Model node:

[Image Input] → [GGUF Load Model (Q4/Q5/Q6/Q8)]
     Γåô
[Mesh Generation] → [Mesh Refinement]
     Γåô
[Texturing (4K)] → [UV Unwrap]
     Γåô
[Export: GLB / OBJ]

Key Node: GGUF Load Model

• Dropdown for quantization level: Q4 / Q5 / Q6 / Q8 / safetensor
• For 6 GB VRAM: select Q4_K_M
• Models must be pre-downloaded (Step 2)

Output Quality by Asset Type

Post-Processing Pipeline

Trellis 2 output is a draft mesh ΓÇö production use requires downstream steps:

Trellis 2 Output (GLB)
    Γåô
Retopologize in:
  - 3DA Studio (cloud retopo)
  - ZBrush (manual or ZRemesher)
  - Blender (Remesh modifier or manual)
    Γåô
Rig in Blender / Maya / Mixamo
    Γåô
Export to game engine

Why retopo is needed: Trellis 2 outputs dense, unstructured topology. Fine for static background assets or further sculpting. Needs clean edge loops for rigging.

Low VRAM Troubleshooting (6 GB cards)

If you hit OOM errors during generation:

1. Lower resolution: Change generation nodes from 1024 ΓåÆ 512
2. Reduce token count: Fewer tokens = less VRAM during inference
3. Lower texture resolution: In the texturing node (slightly less detail, mesh unchanged)
4. Close everything: Chrome, Discord, games ΓÇö every MB counts on 6 GB
5. Add startup flag: --disable-pinned-memory to ComfyUI startup args (frees held GPU memory)
6. Combine above: All three options stack ΓÇö use together if needed

Multi-View Generation

For better accuracy when you have multiple reference angles:

• Same GGUF loader node
• Feed multiple reference images into the workflow
• Multi-view workflow available at pixel-artistry.com/trellis2gguf
• Reduces shape ambiguity on complex assets

Integration with Other 3D Skills

vs. Cloud 3D Tools

Common Mistakes

Resources

• Installer + Workflows: pixel-artistry.com/trellis2gguf
• ComfyUI GGUF Fork: ArrowX fork (linked from installer page)
• Trellis 2 base repo: Visual Bruno / Microsoft (GitHub)
• Retopo tools: 3DA Studio (3daistudio.com), ZBrush, Blender Remesh
• Vault analysis: research/videos/2026-04-15_trellis2-gguf-6gb-vram-local-3d.md