meshy / textTo3D

Text-to-3D is the demo that sells the whole category: type "a weathered stone golem" and out comes a 3D mesh. The KB is clear that in the open ecosystem, true text-to-3D barely exists - it's usually text-to-image followed by image-to-3D stitched into one graph. Meshy is one of the few places that runs the full pipeline natively, and this node wraps it: one prompt, one cloud job, textured mesh out, run through the Civitai Orchestration API on fal's infrastructure.

So this is the node to reach for when you want the demo experience - a prompt and a mesh, no two-stage wiring, no local 3D stack. It's also the node where the quality caveats land hardest, because text-to-3D inherits every weakness of both stages: the geometry is still triangle soup, and the texture is only as good as the intermediate image the model imagined.

Inputs and outputs

• prompt (required) - the only required input. Describe the object; texture words ("weathered," "brass," "hand-painted") carry a lot of weight because they feed the texture stage.
• mode - full default (vs. preview). Full is the real job; preview is presumably the cheap/fast check.
• enable_prompt_expansion - false default; let Meshy flesh out your prompt if it's sparse.
• target_polycount (30000), topology (triangle), symmetry_mode (auto), should_remesh (true), texture_prompt, enable_pbr (false), enable_rigging (false), enable_animation (false), seed - the same mesh pipeline knobs as the imageTo3D sibling.

Outputs: model and fbx_model (STRING), thumbnail (IMAGE), rigged_model/rigged_fbx_model, animated_model/animated_fbx_model, basic_animations, workflow_id, raw_json.

Install

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

Or ComfyUI Manager → "Civitai Comfy Nodes" → Install → restart. Authenticate; 3D jobs bill in Buzz per mesh.

Gotchas

Manage expectations hard on this one. The KB's top-voted community take on text-to-3D quality - "extremely low-quality 3d assets" from the 3D professionals' perspective - applies in full, and the professional's carve-out is the same one that matters here: fine for static props, 3D printing, and silhouette work; a retopology-and-re-UV job for anything animated or game-bound. enable_rigging and enable_animation look like free wins; in practice the generated rigs are where "entirely unusable for interactive applications" lives, so treat them as experiments, not deliverables. If the mesh comes out weak, the levers are texture_prompt, mode, and rewriting the prompt to be more concrete about shape - Meshy's text understanding is better than its topology, and the prompt is what you actually control.