Multimodal

3D generation

A dedicated POST /v1/meshes endpoint — text-to-3d, image-to-3d, and multi-image-to-3d, across 4 labs (Tripo3D, Meshy, Hyper3D/Rodin, Hunyuan3D). Async job/poll, same shape as video generation: submit a request, get a job id back immediately, poll it until status is completed. Output is a textured mesh — GLB always, plus FBX/OBJ/USDZ/MTL and texture maps depending on the model — mirrored to our own storage so the URLs never depend on the upstream provider's own hosting.

Every model takes either prompt (text-to-3d) or image_url/ image_urls (image-to-3d / multi-image-to-3d) — never both, and a model rejects whichever one it doesn't support. Billed flat at creation time from your model + option choice (texture quality, rigging, animation, PBR materials, etc. — see the API reference for exactly which options apply to which model); polling is free.

Generate a 3D model from text

import requests, time

job = requests.post(
    "https://videorouter.sh/api/v1/meshes",
    headers={"Authorization": "Bearer llmr_sk_live_..."},
    json={
        "model": "tencent/hunyuan3d-v3.1-pro-text-to-3d/fal",
        "prompt": "a low-poly wooden chair",
    },
).json()
# -> {"id": "fal:...", "status": "queued", ...}

while True:
    status = requests.get(
        f"https://videorouter.sh/api/v1/meshes/{job['id']}",
        headers={"Authorization": "Bearer llmr_sk_live_..."},
    ).json()
    if status["status"] in ("completed", "failed"):
        break
    time.sleep(5)
# status["data"] -> [{"url": "...model.glb", "format": "glb", "kind": "mesh"}, ...]

const createRes = await fetch("https://videorouter.sh/api/v1/meshes", {
  method: "POST",
  headers: { Authorization: "Bearer llmr_sk_live_...", "Content-Type": "application/json" },
  body: JSON.stringify({
    model: "tencent/hunyuan3d-v3.1-pro-text-to-3d/fal",
    prompt: "a low-poly wooden chair",
  }),
});
const job = await createRes.json();
// -> {"id": "fal:...", "status": "queued", ...}

const statusRes = await fetch(`https://videorouter.sh/api/v1/meshes/${job.id}`, {
  headers: { Authorization: "Bearer llmr_sk_live_..." },
});
// poll on an interval until status.json().status is "completed"/"failed"

curl https://videorouter.sh/api/v1/meshes \
  -H "Authorization: Bearer llmr_sk_live_..." \
  -H "content-type: application/json" \
  -d '{
    "model": "tencent/hunyuan3d-v3.1-pro-text-to-3d/fal",
    "prompt": "a low-poly wooden chair"
  }'
# -> {"id": "fal:...", "status": "queued", ...}

curl https://videorouter.sh/api/v1/meshes/fal:... \
  -H "Authorization: Bearer llmr_sk_live_..."

A real end-to-end run of the example above takes roughly 3–4 minutes (Hunyuan3D; other models vary) and returns 5 output files: the mesh in GLB/OBJ/MTL form, a texture map, and a thumbnail preview PNG — all mirrored to our own storage, never a link back to the upstream provider.

Which models

Every row is served through fal.ai today — no cross-provider failover pool yet, one host per model. Browse the full catalog with live pricing and worked examples per model at the models catalog (filter by Model function → Text-to-3D / Image-to-3D, or Output Modalities → 3D).