Rodin Gen-2

Choosing an input image

Image-to-3D is only as good as the photo you feed it. The model reconstructs what it can see, so the input should show one object, clearly lit, against a plain background. Busy backgrounds or several objects in frame give the model competing signals and muddy the result.

A small glazed ceramic owl figurine with a smooth rounded body, stylized folded wings and large round eyes, soft matte pastel glaze, single object centered and isolated, plain studio background, even soft lighting, product photography

The owl above was reconstructed from a single studio shot. A clean subject with even lighting gives the model an unambiguous read on its shape and materials, which is what separates a usable mesh from a lumpy approximation.

import { createClient } from '@runware/sdk'

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

const [result] = await client.run({
  model: 'hyper3d:rodin@gen-2',
  inputs: {
    images: [
      'https://example.com/owl.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": "hyper3d:rodin@gen-2",
            "inputs": {
                "images": [
                    "https://example.com/owl.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": "hyper3d:rodin@gen-2",
      "inputs": {
        "images": [
          "https://example.com/owl.jpg"
        ]
      }
    }
  ]'

runware run hyper3d:rodin@gen-2 inputs.images.0=https://example.com/owl.jpg

{
  "taskType": "3dInference",
  "taskUUID": "f47ac10b-58cc-4372-a567-0e02b2c3d479",
  "model": "hyper3d:rodin@gen-2",
  "inputs": {
    "images": [
      "https://example.com/owl.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 accepts a URL, base64 string, data URI, or a UUID from a previous generation or the Image Upload API. The request above passes one image.

Turning 2D art into 3D

The input doesn't have to be a photograph. Image-to-3D reconstructs from any clear single image, including flat illustrations and concept art, which makes it a quick way to turn existing 2D assets into 3D. A front-facing character drawing comes back as a model you can pose and light.

A flat vector illustration of a kawaii cartoon cactus character with a round smiling face and little stub arms, planted in a small terracotta pot, simple bold shapes and clean outlines, soft pastel colors, front facing, centered, plain background, sticker style

The model reads the drawing's shapes and gives the flat character real volume and a back the illustration never showed. Simple, clearly outlined art reconstructs more cleanly than busy or heavily shaded drawings.

Using multiple views

A single image is enough, but the model accepts up to five images of the same object, and more angles give it more to reconstruct from. The images should be a turnaround of one subject, and the first image seeds the materials, so lead with the angle that best shows the object's color and surface.

Only the front was generated directly. The other three angles were derived from it with an image-editing model, a quick way to build a multi-view set from a single shot.

The single-image model reconstructs the face it was shown and infers the rest. The four-image model has direct evidence for the sides and back, so the geometry it has to guess shrinks.

[
  {
    "taskType": "3dInference",
    "taskUUID": "9b1deb4d-3b7d-4bad-9bdd-2b0d7b3dcb6d",
    "model": "hyper3d:rodin@gen-2",
    "inputs": {
      "images": [
        "https://example.com/front.jpg",
        "https://example.com/side.jpg",
        "https://example.com/back.jpg",
        "https://example.com/top.jpg"
      ]
    }
  }
]

Preserving transparency

If your input is a cutout on a transparent background, set settings.useOriginalAlpha to true. The model then uses that alpha channel as the object's silhouette instead of detecting edges itself, which keeps masked catalog images and pre-cut assets tight. The setting requires image input, so it has no effect in text-to-3D.

"settings": {
  "useOriginalAlpha": true
}

Prompting for form

Text-to-3D rewards a different kind of prompt than an image model. You're describing an object, not a photograph, so the details that matter are its form and materials. The camera and lighting language that drives an image model has no role here, because the output is a model you light and frame yourself.

A workable prompt names the object, its overall shape, its style, and what it's made of:

The same model handles a soft, stylized creature and a hard-surface mechanical prop without any change in approach. Both prompts lead with the object, then layer on shape and material.

Being specific pays off

The model fills in whatever you leave out. A bare prompt like "a chair" produces a generic chair invented from scratch. A specific prompt constrains the result to the chair you actually want.

Both came from text alone. The detailed prompt locks in the silhouette and materials, while the bare one leaves every choice to the model. The more the asset matters, the more the prompt should say.