comfy / trellis2 / shapeGen

When you're hunting the right silhouette, you don't want to pay for a full textured render every time. TRELLIS.2's shapeGen recipe is the geometry-focused pass of Microsoft's image-to-3D model, hosted on Civitai's cloud - the cheap, fast way to iterate a shape before you spend on texture.

What it is

comfy / trellis2 / shapeGen from the civitai/civitai-comfy-nodes pack. Same plumbing as every Civitai recipe node: IMAGE in, cloud job, mesh URL + thumbnail out. The discriminator marks this as the shape-generation operation, and that's where the value is - a shape run is cheaper and quicker than a full imageTo3D, so it's the right tool when you're still deciding what the object should be.

One honest difference from the Hunyuan3D side of the pack: the trellis shapeGen keeps the texture knobs around. The schema still exposes texture_steps, texture_cfg, and should_texture (default true), so this isn't a bare-geometry-only flag the way the Hunyuan shapeGen is - it's the shape-oriented pass, and you can dial texturing down or off to make it cheaper. Treat it as "shape first, texture optional" rather than "shape only".

The inputs

image_url is required - an IMAGE socket.

• model_version - pixal3D or trellis2. Try both on a test image once; the two backends genuinely differ in look.
• structure_steps, shape_steps, upsample_steps, texture_steps - the four pipeline stages, each with its own step count. For a shape-focused run you can lean on shape_steps and drop the rest.
• should_remesh (true), remesh_resolution (768), target_face_count (700k) - the cleanup controls. TRELLIS.2 remeshes by default, which is a big part of why its output looks cleaner than the raw generators.
• target_resolution (1536) - overall generation resolution; 1024 is noticeably cheaper if you're iterating.
• should_texture - set to false if you truly only want geometry; that skips the paint stages entirely.

The outputs

model (GLB URL), fbx_model, thumbnail (IMAGE), the mostly-empty rigged/animated slots, plus workflow_id and raw_json. For iterating, wire thumbnail → Preview Image and read the cost from raw_json - that's the whole loop.

Installing it

In ComfyUI Manager, search Civitai Comfy Nodes (publisher civitai), Install, restart. Or:

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

No model downloads - just requests and python-socketio.

Gotchas

• Buzz, metered per run. Shape runs are the cheap tier, but they're not free. If a silhouette is still being decided, that's fine; if you're brute-forcing hundreds of variations, check the cost column before you do.
• Credentials required. Civitai Auth node, CIVITAI_API_TOKEN, or sidebar OAuth. No token, no geometry.
• model is a hosted URL string, not a downloaded file.
• Geometry quality is better, not solved. The remesh pass helps, but you're still getting AI topology. Static props and prints: fine. Anything animated: retopo by hand, and be honest that hand-modeling may have been faster.