Scenario 3D Asset Workflows

Overview

Scenario runs text-to-3D, image-to-3D, and 3D-to-3D models behind the same MCP generation loop used for images. The most reliable pipeline generates a concept image first, then feeds it to an image-to-3D model; direct text-to-3D exists (txt23d) but image-to-3D has the larger catalog and more art direction control. Per-family contracts: scenario-meshy, scenario-rodin, scenario-sparc3d. Walkable scenes and Gaussian splats: scenario-3d-worlds. Retexturing a finished mesh or scene with PBR materials: scenario-patina-retexture. Connection and the core generation loop: see the scenario 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

Worked example: concept image to game-ready mesh

A realistic sequence for "make a 3D treasure chest prop". Before step 1, run the recommend call for every later step (step 2, plus any rig or retexture): it is the plan check, and an entry marked requires_plan_upgrade fails with a 403 naming the plan (ModelAccessRestrictedError) only after the steps before it were paid, so settle the upgrade or pick another member first.

1. Generate the concept: pick a text-to-image model via recommend with the user's own words as prompt, then model_schema_get and model_run with a prompt describing a single centered subject on a plain background. If the user has a reference, upload_asset it (plus upload_asset_complete when multipart) and pass that asset ID instead.
2. recommend with capability="img23d" and the user's own words as prompt. Live members include the Hunyuan 3D, Meshy, Tripo, and Trellis families, and Scenario's own LLM-based 3D generation, GPT-6 Astra 3D, and Claude Opus 5.5 3D (the cheaper of the two at authoring time; compare both with dry_run), which reconstruct one object from 1 to 8 photos or renders into an editable mesh with named parts and per-part PBR materials. Their builds suit props and hard-surface subjects, not characters, and need those images (a prompt alone returned 400 at authoring time). The price rises with buildEffort from standard to high (the default) to maximum. faceBudget caps the delivered triangle count at export (10k to 50k for a game-ready asset; the default keeps fine detail and is not that). To change the result, edit it (see Refining meshes).
3. model_schema_get on the chosen model. 3D schemas vary widely: single image vs multi-view arrays, polycount targets, PBR toggles, topology choices.
4. model_run with parameters={"image": "asset_xxx", ...} and wait=false (dry_run=true first to price a batch), then jobs_wait with job_ids=["<job_id>"] (re-call it with the returned pending_job_ids as job_ids if it times out). A downstream step (rig, retexture) has no payload to dry_run until its input mesh exists: quote it from recommend as an estimate, then re-price it with dry_run on the real asset before launching. Never dry_run it with a placeholder id: file ids resolve before plan access is checked, so the call returns 404 and proves neither price nor access.
5. asset_display with the output asset_id to preview, then asset_download with format: "glb" and curl -L -o chest.glb "<url>" for engine import. asset_download converts a mesh when format is glb, fbx, or obj; omitted, it returns the stored original, a Blender file for the LLM-based members. Ask for fbx for a Unity project without a glTF importer, and keep the GLB too: at authoring time the FBX embedded only two images, so it may not carry every PBR map.

Multi-view models accept several images of one subject from different angles; the count and the ordering vary per model, so take both from model_schema_get (the first image is usually the front view).

Texture and lighting controls

Several image-to-3D families split texture from geometry, and each dial is one model_schema_get away, so read them before promising a look. Authoring-time examples from the Tripo members:

• texture: false returns a bare mesh with no texture and is a cost_impact field, the cheap path when the user will texture in a DCC. textureQuality (fast, standard, detailed, extreme) also moves the price, and a texture version picker, left empty, keeps the provider's default.
• delight strips lighting and shadows baked into the reference image so the mesh lights correctly in the user's engine. It defaulted to on; turn it off only when the painted shading is the art style, as on a hand-painted prop. The export stays a lit material either way, so for no engine relighting at all the user sets an unlit shader on import.
• The pbr flag's description said PBR on, its default, ignores the texture parameters. A texture setting the user asked for is honored only with pbr: false there, so read that description on the chosen member and say which one won.
• Geometry and texture take separate seeds (seed, textureSeed): hold seed and the image fixed and vary textureSeed alone for texture variants on one shape.
• autoSize scales the output to real-world meters and defaulted to off: set it to true for a real-scale engine import, or the mesh keeps its native size. The asset's dimensions still read the unit-normalized mesh (longest side 1.0) because the meters sit in the GLB's node scale, so dimensions is not proof it failed.

Inspecting results

Hosts without MCP Apps get the app_url dashboard link instead of the asset_display viewer. The viewer's capture button calls capture_3d_view, an app-only tool: it uploads the current camera view as a new image asset and posts the asset_id back into the conversation. Use that capture as a reference image for follow-up generations or similarity search. Never call capture_3d_view yourself; it requires PNG canvas data only the viewer has.

Refining meshes

3D-to-3D utilities cover retexturing, remeshing, UV unwrapping, and part segmentation: recommend with capability="3d23d" plus the operation in the user's own words. Most take the source asset_id in a kind: "3d" file field, usually named model (also mesh, file3d, sceneFile); 400 Input model is required or Provide a reference image or a 3D model means the mesh went in under another name (an image field, or a URL), never that the tool wants something else.

The LLM-based members of step 2 also edit an existing mesh, prop or character: sceneFile takes its asset_id (any 3D model, returned in the same format), images stays empty, prompt states the change, and buildEffort fits the change, not the object (standard for a recolor, high or maximum to add or reshape parts). GPT-6 Astra 3D's description also claimed Mixamo-named rigs, ARKit face shapes, animation loops, UV re-unwraps, and modular kits, so it competes in those recommend calls. At authoring time recommend priced no edit, but the price followed buildEffort, not the mesh, so a dry_run on any mesh in the project quotes it before the first paid step. A recolor kept every part name and the face count; a reshape is uncapped (faceBudget is ignored) and took a 4,817-triangle cat to 788,954 triangles, so plan a remesh before an engine. A replaced material keeps its old name; only added parts get new ones. The edit is generated: compare what the prompt did not name against the source.

Splitting a finished mesh into parts is a contested lane, so it stays a recommend pick: at authoring time several vendors offered mesh segmentation with a granularity control, one combined the split with a PBR retexture, and image-to-parts members build the parts from the picture instead. A convincing mesh is not a game-ready one: part separation, joint placement, materials, and animation are each their own pass, and polygon caps differ by topology on the members that offer both (a quad cap sat well under the triangle cap on one), so read the slider's max off the schema instead of promising a count. A named provider feature is a valid search target, but an empty result or one member's body-plan enum does not prove a capability is absent platform-wide. For an unmet need such as a hinged prop, use recommend with that need and inspect the returned schemas; if none exposes it, report it as unverified in the inspected models, without inventing a call from the provider's own site.

A staged chain toward game-ready

One image-to-3D run bakes its texture onto UVs the generator chose, so a mesh headed for an engine usually goes further run by run, each output's asset_id feeding the next tool's kind: "3d" field: geometry first (a multi-view member when several angles of the subject exist, with texture: false where the schema offers it, since texture comes later, and pbr: false beside it on members whose pbr default ignores the texture settings), then a UV unwrap, then a retexture on the unwrapped mesh from a prompt, an image, or a style reference, then the rig last, on the topology that ships. Read each retexture schema for a setting that rebuilds UVs, which would discard the unwrap, and for the same pbr default: at authoring time the Tripo texturing member defaulted to PBR, which ignores its texture options. When a schema says nothing about UVs, prefer a member with an explicit keep-existing-UVs setting. Each step after the first is a 3d23d recommend pick (see Refining meshes) and its own billed run, so dry_run it on the real input before launching. Read input caps before chaining: one unwrapper at authoring time took meshes up to 30,000 faces, so set the generator's face limit under it where the schema has one (the Tripo multi-view member exposed faceLimit and a smartLowPoly topology switch), or remesh down first. Check each output with asset_display before feeding it on, or, where the host shows no viewer, read face count, UVs, and dimensions off asset_get and compare them step to step, and look at the record's thumbnail asset, which can lag the job by seconds (hasUVs only proves UVs exist, not that the layout is clean; an unwrapper's "albedo" PNG may be a preview, not the atlas): faces that render dark or inside out are flipped normals, which a later step inherits, and scale and pivot can shift between steps, even within one vendor's chain (autoSize above), so confirm both in a DCC before the rig. When a step returns sibling formats (a GLB beside an OBJ), feed the GLB onward unless the next schema asks for another or the mesh is quads, which stay in FBX (see Rigging and animation); when none is a GLB, feed the OBJ or FBX when the next schema lists it (a dry_run may price a call without checking the file format, so the real run is what confirms it), else convert with asset_download format. Fixing normals, origin, or naming by hand is DCC work, not an MCP call (the scenario-blender-expert skill drives Blender; it installs locally like the siblings above, never over MCP, so skill_get cannot fetch it).

Rigging and animation

Rigging is a separate 3d23d step on a finished mesh, not a generator flag: recommend with that capability plus the rigging need in the user's own words.

Body plan picks the model. Humanoid models take the mesh and little else (a front-facing hint, or an approximate height, depending on the model) and infer a biped skeleton. Non-biped work goes to a model exposing rigType, whose values cover quadruped, hexapod, octopod, avian, serpentine, and aquatic. A static prop has no body plan to rig: if recommend answers ask_user, part segmentation is the nearer step.

A character headed for a rig is built for it from the concept on. Generate the concept in a T- or A-pose (full body, front view, arms clear of the torso, plain background), since riggers infer joints from the pose. When the user wants quads, pick an image-to-3D member whose schema offers quad topology and set it, because the default is triangles: a quad switch with a lower face cap on the Tripo P2 and 3.1 members at authoring time, while scenario-rodin and scenario-meshy name their families' own controls. Quad output arrives as FBX, and glTF stores triangles only: converting it with asset_download format: "glb", or feeding it to a step that accepts GLB alone, triangulates it again. Keep the FBX, rig it with a member whose file field lists FBX, and check the face type of every output in a DCC before promising quads. A rig that comes back as GLB is triangulated: deliver the quad FBX beside it and say so.

Three schema details decide whether the output is usable:

• Formats. Rigging models accept GLB, and often OBJ, FBX, or STL. None exposes an output-format field, so the rig comes back as GLB or FBX and most descriptions do not say which: expect a DCC pass when the engine needs the other.
• Size ceiling. A max_size on the file input is the exception rather than the rule (one humanoid rigging model caps at 30 MB). Check the schema before assuming a large mesh needs decimating.
• Animation versus rig. Where a rigger exposes animation it retargets a preset clip, and its allowed_values are rig-type prefixed (quadruped:walk), so read them rather than guess. By default only the retarget file comes back: set includeRiggedModel to keep the plain rigged mesh too (the two come back as separate assets with identical metadata, so asset_get tells them apart: the retarget has properties.hasAnimations true and an animationFrameCount, the plain rig false and null, and their order is not guaranteed).

When only motion is wanted, motion-transfer video models animate a still character image with no skeleton at all: see scenario-video. Video-to-motion models that auto-rig an uploaded mesh are the one place an outputFormat enum picks the engine target.

Common mistakes

• Running model_run without model_schema_get: 3D model parameters differ far more between models than image models do.
• Passing a local file path as an image input: upload_asset first, then pass the returned asset ID.
• Hardcoding model IDs: catalogs rotate (a deprecated:<replacement_id> tag names the successor). Re-discover each session, recommend for the need or search for a name.
• Pasting raw asset URLs into chat instead of calling asset_display.
• Promising a named skeleton, an influence count, or bones for wings and extra limbs that the chosen member does not document: pick the closest body plan, then finish the rest in a DCC.
• Sending a biped to a rigType model: the enum has no biped value, because humanoids have their own rigging models.