Meshy-6

What makes a good reference

In image-to-3D the model can only rebuild what the photo shows, so the input sets the ceiling on the result. Give it one object, evenly lit, on a plain background. A cluttered scene or a second item in frame splits the model's attention, and the mesh comes back soft.

A glossy ceramic teapot in deep teal with a curved spout, a rounded lid with a small knob and a polished bamboo handle, centered product photograph on a plain light gray seamless studio background, three-quarter angle, soft even studio lighting, the entire teapot fully visible with margin around it, sharp focus, photorealistic, no text

This teapot came from a single catalog-style shot. Flat light and an empty backdrop let the model read the curves and the glaze without inventing anything, and that read is the difference between a crisp mesh and a melted one.

import { createClient } from '@runware/sdk'

const client = await createClient({ apiKey: process.env.RUNWARE_API_KEY })
await client.connect()

const [result] = await client.run({
  model: 'meshy:meshy@6',
  inputs: {
    images: [
      'https://example.com/teapot.jpg'
    ]
  }
})

import asyncio
import os

from runware import Runware


async def main():
    async with Runware(api_key=os.environ["RUNWARE_API_KEY"]) as client:
        results = await client.run({
            "model": "meshy:meshy@6",
            "inputs": {
                "images": [
                    "https://example.com/teapot.jpg"
                ]
            }
        })


asyncio.run(main())

curl https://api.runware.ai/v1 \
  -H "Authorization: Bearer $RUNWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[
    {
      "taskType": "3dInference",
      "taskUUID": "f47ac10b-58cc-4372-a567-0e02b2c3d479",
      "model": "meshy:meshy@6",
      "inputs": {
        "images": [
          "https://example.com/teapot.jpg"
        ]
      }
    }
  ]'

runware run meshy:meshy@6 inputs.images.0=https://example.com/teapot.jpg

{
  "taskType": "3dInference",
  "taskUUID": "f47ac10b-58cc-4372-a567-0e02b2c3d479",
  "model": "meshy:meshy@6",
  "inputs": {
    "images": [
      "https://example.com/teapot.jpg"
    ]
  }
}

{
  "data": [
    {
      "taskType": "3dInference",
      "taskUUID": "f47ac10b-58cc-4372-a567-0e02b2c3d479",
      "outputs": {
        "files": [
          {
            "uuid": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
            "url": "https://im.runware.ai/image/os/a14d18/ws/2/ii/a1b2c3d4-e5f6-7890-abcd-ef1234567890.glb"
          }
        ]
      }
    }
  ]
}

inputs.images takes a URL, a base64 string, a data URI, or a UUID from an earlier task or the Media Storage API. The finished GLB arrives at outputs.files[].url.

Turning enhancement off

Before it reconstructs, Meshy runs a cleanup pass on your photo. On a casual phone shot that usually helps. On a render or a stylized asset you've already polished, the pass can repaint detail you meant to keep. Set settings.imageEnhancement to false and the model works from your image untouched.

A hand-painted tabletop miniature of an elderly wizard in flowing blue robes holding a gnarled wooden staff topped with a glowing crystal, a long white beard and a pointed hat, standing on a circular rocky base, centered product photograph on a plain light gray seamless studio background, three-quarter angle, soft even studio lighting, the entire miniature fully visible with margin around it, sharp focus, photorealistic, no text

"settings": {
  "imageEnhancement": false
}

The setting does nothing in text-to-3D, where there's no photo to clean up.

Adding more angles

A single photo is enough, but Meshy accepts up to four images of the same object, and every extra angle is one less surface the model has to invent. Shoot them as a turnaround, the same subject from different sides.

A stylized collectible astronaut character, front view, white and orange space suit with a large round tinted helmet visor, a chunky life-support backpack, ribbed joints at the elbows and knees, sturdy boots, standing upright in a neutral relaxed pose, plain white background, even studio lighting, full body centered with margin around it, clean game-ready collectible style

Show the exact same astronaut character from a direct side profile view. Keep the identical suit design, colors, proportions, and lighting. Plain white background, full body centered with margin around it.

Show the exact same astronaut character from a direct rear view, revealing the life-support backpack. Keep the identical suit design, colors, proportions, and lighting. Plain white background, full body centered with margin around it.

For this set I generated the front directly, then made the side and back from it with an image editor, a handy shortcut when no real turnaround exists.

From the front alone, the model has to imagine the life-support pack it never saw, and the guess shows. Hand it the sides and back too, and the rear hardware is built from evidence.

[
  {
    "taskType": "3dInference",
    "taskUUID": "9b1deb4d-3b7d-4bad-9bdd-2b0d7b3dcb6d",
    "model": "meshy:meshy@6",
    "inputs": {
      "images": [
        "https://example.com/front.jpg",
        "https://example.com/side.jpg",
        "https://example.com/back.jpg"
      ]
    }
  }
]

Describe the object

A text prompt for 3D reads differently from one for an image. There's no camera to place and no light to set, because you frame and light the mesh yourself afterward. What earns a place in the prompt is the object itself: its shape and what it's made of.

The recipe holds across very different objects. A soft creature, a hard-surface vehicle, and a potted plant all come back clean from prompts built the same way, each leading with the object before its shape and materials.

The more you say, the less it guesses

Whatever you leave out, the model settles on its own. Ask for "a backpack" and you get a capable but plain one. Spell out the fabric and the hardware, and you get the pack you had in mind.

Both are text-only. The longer prompt nails down the silhouette and the loadout. The short one hands every call to the model. Spend words in proportion to how much the asset matters, a positivePrompt has room for up to 600 characters.

[
  {
    "taskType": "3dInference",
    "taskUUID": "c4278a91-6c18-494d-a0d9-a4359dcbd783",
    "model": "meshy:meshy@6",
    "positivePrompt": "A rugged tactical hiking backpack in dark olive ripstop fabric with molle webbing, several zippered compartments, a rolled sleeping bag lashed to the top, side water-bottle pockets and padded shoulder straps"
  }
]