| name | multiview-3d |
| description | Use when the user wants a high-quality 3D model from multiple views, a reference photo expanded into front/side/back views, Tripo multiview reconstruction, or a multi-angle image-to-3D pipeline. |
Multi-View 3D Pipeline Skill
You are an expert at orchestrating the multi-angle 3D generation pipeline. This skill uses trident-mcp for Tripo multiview generation/reconstruction, and can optionally use gemini-media-mcp for more controlled reference-image creation or iterative angle edits.
This is a composed workflow skill. trident-mcp itself remains independently useful in any MCP-capable client. Pair it with gemini-media-mcp when the workflow needs richer image ideation or manual control over each generated angle.
Prerequisites
Minimum:
- trident-mcp — Needs
multiview_to_3d. Prefer also having get_balance, task_status, download_model, image_to_multiview, and edit_multiview.
Optional:
Before starting, verify the available tools. If multiview_to_3d is missing, tell the user to configure trident-mcp. If only Gemini tools are missing, use Tripo's native image_to_multiview route when possible.
The Pipeline Workflow
Phase 1: Understand Intent
What does the user want in 3D? Understand:
- Subject — Character, object, vehicle, creature, etc.
- Style — Realistic, stylized, cartoon, game-ready?
- Reference — Do they have an existing image, or should we generate one?
Phase 2: Acquire Reference Image
Three paths:
"I have a reference photo":
- User provides an image path or URL
- Confirm it's suitable for 3D reference (clear subject, good lighting, no heavy occlusion)
- Proceed to Phase 3
"Generate an image first":
- Use
text_to_image from trident-mcp or generate_image from gemini-media-mcp
- Craft a prompt focused on a single object with clear form and clean background
- Example: "A detailed fantasy sword with ornate handle, clean white background, product photography style, front view"
- If using
text_to_image, poll the task and use the completed output.imageUrl as the source image; do not pass the image task ID as an image input.
- Show the result, offer to regenerate if needed
- Proceed to Phase 3
"I have multiple angles already":
- Skip directly to Phase 4 with the user's images
Phase 3: Generate Multi-Angle Views
Two approaches, offer the user a choice:
Approach A: Native Tripo Multiview (Recommended When Available)
Use image_to_multiview from trident-mcp to create ordered multi-angle views from a single reference image. If a generated view is wrong, use edit_multiview with per-view prompts entries for targeted corrections before reconstruction.
This is the shortest path when the source image is already strong and the user does not need manual control over every intermediate angle.
After image_to_multiview or edit_multiview succeeds, prefer passing that successful task ID to multiview_to_3d as taskId. Extract individual view URLs only if the tool version lacks task-id reuse.
Approach B: Iterative Gemini Views
Generate each angle separately using edit_image from gemini-media-mcp, producing more consistent results:
- Front view — Use the reference image as-is, or generate a clean front view
- Left side view —
edit_image with prompt: "Show this same [subject] from the left side, 90 degrees rotated, same style, same lighting, clean background"
- Back view —
edit_image with prompt: "Show this same [subject] from the back, 180 degrees rotated, same style, same lighting, clean background"
- Right side view —
edit_image with prompt: "Show this same [subject] from the right side, 270 degrees rotated, same style, same lighting, clean background"
Show each generated view to the user before proceeding. If a view looks wrong, regenerate it.
Approach C: Character Sheet (Single Generation)
Generate a 2x2 grid of views in one shot using generate_image:
Prompt template:
Character sheet of [subject description], 2x2 grid showing FRONT view (top-left), LEFT SIDE view (top-right), BACK view (bottom-left), RIGHT SIDE view (bottom-right). Clean white background, consistent lighting, same proportions in all views. Professional reference sheet style.
After generation, the individual views need to be cropped from the grid. If ImageMagick or similar is available, use convert -crop to split the 2x2 grid into 4 images. Otherwise, ask the user to crop manually.
Note: Iterative is recommended because character sheets sometimes have inconsistent proportions between views.
Phase 4: 3D Reconstruction
Feed the generated views to trident-mcp:
- Run
get_balance before credit-spending work when a live Tripo key is configured.
- Call
multiview_to_3d with either the successful multiview taskId, or 2-4 angle images (paths/URLs) in front/left/back/right order.
- Use
v3.1-20260211 for best quality, or P1-20260311 when the user needs low-poly/game-ready topology.
- Use options such as
quad, smartLowPoly, generateParts, faceLimit, textureAlignment, and enableImageAutofix when topology or texture alignment matters. For P1, omit smartLowPoly because P1 is already low-poly; use faceLimit instead.
- Poll
task_status until complete — this typically takes 1-3 minutes.
- Download the result with
download_model in its actual task output format.
If the user needs FBX, OBJ, STL, USDZ, or 3MF instead of the task's native output, run convert_format first and then download the conversion task.
Phase 5: Post-Processing
After the 3D model is ready, offer:
"Multi-view 3D model generated! Saved to: [path]. What would you like to do?"
- Retopologize — Clean lowpoly version for game use
- Convert format — Export to FBX, OBJ, STL, USDZ, 3MF
- Texture/refine/segment — Improve texture, refine quality, or split parts
- Rig/animate — Check riggability, rig, or retarget animation
- Stylize — Apply LEGO, voxel, Voronoi, or Minecraft style
- Send to Blender — Import via the preferred Blender connector (official first, community fallback; use the 3d-to-blender skill)
- Regenerate views — Try different angles for better geometry
- Done — Keep this model
Tips for Best Results
- Clean backgrounds: White or neutral backgrounds produce better 3D results
- Consistent lighting: All views should have similar lighting direction
- Single subject: Multi-view works best with a single, centered object
- Crisp details: Higher quality reference images produce better geometry
- 2 views minimum, 4 ideal: Front + side is minimum; front + left + back + right is ideal
