Scenario Meshy 3D

Overview

Meshy on Scenario is a toolchain rather than one model: image-to-3D generators (Meshy 7 Image to 3D, Meshy 7 Multi Image to 3D, Meshy T2 Smart Topology), Meshy 6 Text-to-3D, and GLB-in, GLB-out utilities (Retexture, Remesh, UV Unwrap, Rigging, Animation). Work runs as a pipeline: generate a mesh, refine it, then rig or animate, each stage its own model_run whose model parameter takes the 3D asset id the previous stage returned. Discover members with search and treat model_schema_get as the contract: members disagree on defaults as basic as enablePbr (true on Image to 3D, false on Multi Image and Retexture at authoring time).

Connection and the core loop: see the scenario skill in this repo; model-agnostic 3D work (viewer, capture, engine import): the scenario-3d skill. If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.

Quick reference

Members and their traps (names from the live schema, caps at authoring time):

Member	Core input	Watch for
7 Image to 3D	image (1-4)	ultraMode takes a single image only; raw mesh by default
7 Multi Image to 3D	image (1-4, first = front view)	remeshes in-run; savePreRemeshedModel keeps the raw GLB
T2 Smart Topology	image (1-4)	animation-ready topology; targetPolycount caps at 15,000
6 Text-to-3D	prompt (600 chars)	shouldRemesh off by default
Retexture	model + one style input	geometry untouched; keep inputs near 30K polys; enableOriginalUv
Remesh	model	targetPolycount 100 to 300,000, resizeHeight, originAt
UV Unwrap	model	rejects meshes above 44,000 faces; Remesh down first
Rigging	model, heightMeters	humanoid skeleton and skin weights
Animation	model, actionId	auto-rigs, then applies the clip

One silent rule runs through every texture control: an image beats text. On the generators textureImage overrides texturePrompt (the text is ignored, not blended), and both need shouldTexture on (Text-to-3D has no such toggle: it always textures). Retexture is stricter: exactly one of textStylePrompt, imageStyle, or multiviewImage (1 to 4 views, first is the front). textureResolution runs 2k, 4k, or 8k; at 8k the PBR maps come back at 4K with no emission map (Meshy 7 never produces one), and 8k pairs only with triangle topology when remeshing. poseMode takes lowercase "a-pose" or "t-pose": set one on any character headed for rigging.

Generate raw, refine on purpose

Keep shouldRemesh off on Meshy 7 Image to 3D (the schema recommends it): take the raw mesh at generation, then control polycount with Remesh, which owns targetPolycount, topology, resizeHeight in meters (0 keeps scale), and origin placement. On the generator, targetPolycount does nothing while shouldRemesh is off. ultraMode buys finer geometry for extra cost but accepts one input image; when hidden sides matter more than micro detail, spend the budget on up to four views instead. Costs spread widely (at authoring time the single-image Meshy 7 p50 ran three times the multi-image one), so dry_run before batches.

Rig or animate, not both

Rigging and Animation are alternatives, not stages. Animation auto-rigs internally, then applies the clip named by actionId, a number from Meshy's animation library (ids listed at docs.meshy.ai/en/api/animation-library). Run Rigging alone when you want a rigged GLB for your own clips; go straight to Animation for a moving character, optionally setting postProcessOperation to change_fps (postProcessFps 24, 25, 30, or 60) or extract_armature. Both scale the rig from heightMeters, so pass the character's real height.

Worked example: photo to animated character

1. search with target="models", query="meshy", public=true. Single-image route: model_meshy-7-img23d (a live hit at authoring time: re-discover each session).
2. model_schema_get with that id, then upload_asset the character photo (see the scenario skill).
3. model_run with dry_run=true and parameters={"image": ["asset_photo"], "poseMode": "t-pose", "texturePrompt": "worn leather armor, muted palette", "enablePbr": true} for the estimate; then re-run with wait=false and jobs_wait with the job id, re-called with pending_job_ids on timeout, never a second model_run.
4. asset_display the GLB; check silhouette and texture before spending further.
5. search for the animation member, model_schema_get, then model_run with parameters={"model": "<GLB asset id from step 4>", "heightMeters": 1.8, "actionId": <id from the library>}.
6. jobs_wait, asset_display, then asset_download for engine import.

Common mistakes

• texturePrompt next to textureImage expecting a blend: the image wins, the text is ignored.
• Two style inputs on Retexture: it takes exactly one of prompt, style image, or multi-view images.
• ultraMode: true with several images: Ultra requires a single input image.
• UV Unwrap on a dense mesh: above 44,000 faces the job is rejected; Remesh down first.
• Rigging before Animation: Animation rigs by itself, so the chain wastes a job.
• targetPolycount on Image to 3D or Text-to-3D with shouldRemesh off: it does not apply.
• A bare string where the schema says array: one photo goes as ["asset_x"].