Vision

Understanding

task is describe, read, ask (with question), label (with labels) or analyse_video (with frames, oldest first). Images travel inline as base64 or a data: URI (PNG, JPEG, WebP or GIF, up to 20 MB). A whole clip is not read by the network yet: send frames, or the call answers 503 coming_online.

curl https://api.tenzro.xyz/vision/understand \
  -H "Authorization: Bearer $TENZRO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"task": "describe", "image": "'"$(base64 < street.jpg)"'"}'

Vision describes what an image shows; it never identifies people, and questions that ask who someone is are refused.

Datasets and consent

Generated images and video go into a Dataset with dataset. Inputs are kept only when you ask, with save: say whether people appear, and when they do, give the Fieldwork consent_ref they were collected under. The consent record is re-read at every use; a withdrawn participant's images are never read again. See Fieldwork to collect images and video with consent.

{
  "task": "read",
  "image": "<base64>",
  "save": { "dataset_name": "Receipts", "people": false }
}

# Images that show people are kept only with the Fieldwork consent they were collected under:
"save": { "dataset_id": "ds_…", "people": true, "consent_ref": "fw:<project>:<participant>:<version>" }

Image and video generation

Images come back inline with their SHA-256 and a signed manifest in the C2PA assertion model (created by a trained algorithm, the model and its licence, the machine, the node's receipt). Output that cannot be signed is never delivered. The OpenAI SDK works against the /v1 base.

from openai import OpenAI
import base64, os

client = OpenAI(base_url="https://api.tenzro.xyz/v1", api_key=os.environ["TENZRO_API_KEY"])
img = client.images.generate(model="flux2-klein-4b", prompt="A lighthouse at dusk, in ink", size="1024x1024")
open("lighthouse.png", "wb").write(base64.b64decode(img.data[0].b64_json))

Video renders as a job. A job whose model is not serving yet stays queued and starts by itself.

curl https://api.tenzro.xyz/vision/videos -H "Authorization: Bearer $TENZRO_API_KEY" -H "Content-Type: application/json" \
  -d '{"model": "wan2.2-ti2v-5b", "prompt": "Waves rolling onto a beach at dawn", "seconds": 4, "dataset": {"dataset_name": "Beaches"}}'
# 202 { "id": "vjob_…", "status": "queued", ... }
curl https://api.tenzro.xyz/vision/videos/vjob_… -H "Authorization: Bearer $TENZRO_API_KEY"
curl https://api.tenzro.xyz/vision/videos/vjob_…/content -H "Authorization: Bearer $TENZRO_API_KEY" --output beach.mp4

Runs over a Dataset

Describe, detect or label every accepted image of a Dataset, such as a Fieldwork project's. Each result is an item in a new Dataset with its own provenance: the source item, the model, the machine, the receipt and the consent it was collected under.

curl https://api.tenzro.xyz/vision/runs -H "Authorization: Bearer $TENZRO_API_KEY" -H "Content-Type: application/json" \
  -d '{"dataset_id": "ds_…", "task": "label", "labels": ["kitchen", "street", "office"]}'

SDK and MCP

import { TenzroLabs } from "@tenzrolabs/platform/server";

const tz = new TenzroLabs(); // reads TENZRO_LABS_API_KEY
const { text, provenance } = await tz.vision.describe(await readFile("street.jpg"));
const job = await tz.vision.videos.create({ prompt: "Waves rolling onto a beach at dawn", seconds: 4 });
const done = await tz.vision.videos.wait(job.id);

MCP tools: describe_image, read_image_text, ask_about_image, label_image, analyse_video_frames, detect_objects, generate_image, start_video_generation, get_video_generation, start_vision_run, get_vision_run, list_vision_models. See MCP server.

Models and prices

Prices are in Tenzro Labs credits, from the catalogue. A model without a set price is not offered until it has one.

Reference

• GET /vision/status no key
• GET /vision/models no key
• POST /vision/understand vision:analyse
• POST /vision/detect vision:analyse
• POST /vision/images vision:generate
• POST /vision/videos vision:generate
• GET /vision/videos vision:generate
• GET /vision/videos/{id} vision:generate
• GET /vision/videos/{id}/content vision:generate
• POST /vision/runs vision:batch
• GET /vision/runs vision:batch
• GET /vision/runs/{id} vision:batch
• POST /vision/runs/{id}/cancel vision:batch
• GET /vision/outputs/{id} vision:generate
• GET /vision/outputs/{id}/content vision:generate