Meshy: Text to Model

No image, no reference - just a text prompt and Meshy's cloud turns it into a textured 3D model. Meshy: Text to Model is the purest "I want a thing" node in this family, and it's the right starting point when you're brainstorming props, creatures, or set dressing and don't want to hunt for reference images first.

The honesty rules still apply: it's a paid cloud API node (Comfy Org account + credits, ~$0.80 per text-to-3D run), and text-to-3D quality is famously more hit-or-miss than image-to-3D. You're asking the model to invent both the shape and the look from words alone, so expect drafts, not finals. That's fine - that's what the refine node is for.

How it works

You type a prompt, optionally pick a style, and the node uploads the request to Meshy's text-to-3D API via Comfy's cloud. Meshy generates the mesh (plus a texture pass by default), Comfy polls until it's done, and the result downloads as GLB and FBX into your output folder with a meshy_task_id for the next step. Everything is server-side; nothing runs on your GPU.

The inputs that matter

• prompt - up to 600 characters. This is where the craft lives. Vague prompts ("a sword") give you the most generic sword ever generated; specific ones ("an ornate ceremonial sword with a golden dragon-wrapped hilt and a blue gem in the pommel") give the model something to work with. Material words ("steel, leather, brass") help the texture pass a lot.
• style - realistic or sculpture. Realistic for game assets and props; sculpture if you want a stylized, carved look.
• should_remesh - true (default) cleans the mesh and exposes topology (triangle/quad) and target_polycount (100–300K). false returns the raw mesh.
• symmetry_mode - auto/on/off; on for anything symmetric, and prompts for characters and vehicles usually are.
• pose_mode - A-pose/T-pose for characters you'll rig.
• seed - a re-run control; results stay non-deterministic (the tooltip says as much).

Outputs: model_file (STRING), meshy_task_id (MESHY_TASK_ID, feeding refine/rig/texture), GLB, FBX.

Common issues

The biggest disappointment curve is quality: text-to-3D lags image-to-3D in every service, Meshy included. If the draft is close but not right, refine it rather than re-rolling the prompt from scratch - you'll keep the parts that worked. Second, 600 characters is a cap, not a target; short, dense, material-rich prompts beat walls of adjectives. And remember each run costs credits - generate a couple of drafts, pick the best, don't treat it like a free sampler.