Kandinsky 2.2 Serverless API

Kandinsky inherits best practicies from Dall-E 2 and Latent diffusion, while introducing some new ideas.

PlaygroundHistory
APIPricing
~2.00s
POST /v2/kandinsky2.2-txt2img · submit + poll
 1# pip install "segmind>=1.1.0"
 2# export SEGMIND_API_KEY="YOUR_API_KEY"
 3import segmind
 4
 5# Async (v2): submit to the queue and block until COMPLETED.
 6# run() returns the final result dict (600s deadline, 1.0s poll by default).
 7result = segmind.run(
 8    "kandinsky2.2-txt2img",
 9    prompt="masterpiece, best quality, portrait of an old man, 50mm, solo, natural skin texture, realistic eye and face details, dark, deep shadow, darkness, moonlight, award winning photo, extremely detailed, fine detail, highly detailed, extremely detailed eyes and face, piercing red eyes, detailed clothes, skinny, gothic, native american clothing, analog film, stock photograph,",
10    negative_prompt="lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, out of frame, extra fingers, mutated hands",
11    samples=1,
12    num_inference_steps=25,
13    img_width=512,
14    img_height=768,
15    prior_steps=25,
16    seed=9863172,
17    base64=False,
18)
19print(result["status"])                      # COMPLETED
20print(result.get("output"))                  # model output (e.g. media URL)
21print(result["metrics"]["inference_time"])   # server compute seconds
22
23# --- Or submit + poll manually (track request_id, control the cadence) ---
24from segmind import SegmindClient, InferenceFailed, InferenceTimeout
25
26client = SegmindClient()                      # reads SEGMIND_API_KEY
27payload = {
28    "prompt": "masterpiece, best quality, portrait of an old man, 50mm, solo, natural skin texture, realistic eye and face details, dark, deep shadow, darkness, moonlight, award winning photo, extremely detailed, fine detail, highly detailed, extremely detailed eyes and face, piercing red eyes, detailed clothes, skinny, gothic, native american clothing, analog film, stock photograph,",
29    "negative_prompt": "lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, out of frame, extra fingers, mutated hands",
30    "samples": 1,
31    "num_inference_steps": 25,
32    "img_width": 512,
33    "img_height": 768,
34    "prior_steps": 25,
35    "seed": 9863172,
36    "base64": False,
37}
38job = client.submit_async("kandinsky2.2-txt2img", **payload)
39print(job.request_id)                         # available immediately
40try:
41    result = job.wait(timeout=600, interval=1.0)
42except InferenceTimeout as e:
43    print("still running:", e.request_id)
44except InferenceFailed as e:
45    print("failed:", e.detail)

API Endpoint

POSThttps://api.segmind.com/v1/kandinsky2.2-txt2img

Parameters

promptrequired
string

Prompt to render

base64optional
boolean

Base64 encoding of the output image.

Default: false
img_heightoptional
integer

Image resolution.

Default: 768
Allowed values :
7681024
img_widthoptional
integer

Image resolution.

Default: 768
Allowed values :
7681024
negative_promptoptional
string

Prompts to exclude, eg. 'bad anatomy, bad hands, missing fingers'

num_inference_stepsoptional
integer

Number of denoising steps.

Default: 20Range: 20 - 100
prior_stepsoptional
integer

Number of denoising steps.

Default: 25Range: 1 - 100
samplesoptional
integer

Number of samples to generate.

Default: 1Range: 1 - 4
seedoptional
integer

Seed for image generation.

Default: -1

Response Type

Returns: Text/JSON

Asynchronous requests (v2)

Use Async for video, long-running (>~60s), or high-concurrency workloads; Sync is simplest for fast image & LLM calls. Async submits a request and you poll it to completion.

  1. 1
    POST /v2/kandinsky2.2-txt2img

    Submit — returns request_id, status_url, response_url

  2. 2
    GET /v2/requests/{id}/status

    Poll — until COMPLETED or FAILED

  3. 3
    GET /v2/requests/{id}

    Result — final response body

Status states

QUEUED— Accepted, waiting for a worker
PROCESSING— Running on a worker
COMPLETED— Done — result body is ready
FAILED— Errored (incl. content/RAI blocks)
  • A FAILED request is served as HTTP 422 — the body still carries the error detail.
  • An unknown or expired request_id returns HTTP 404.
  • Results are retained for 1 hour, then expire.
  • Content / RAI blocks surface as FAILED, not a separate state.
  • Track completion by polling the status endpoint.

Common Error Codes

The API returns standard HTTP status codes. Detailed error messages are provided in the response body.

400

Bad Request

Invalid parameters or request format

401

Unauthorized

Missing or invalid API key

403

Forbidden

Insufficient permissions

404

Not Found

Model or endpoint not found

406

Insufficient Credits

Not enough credits to process request

429

Rate Limited

Too many requests

500

Server Error

Internal server error

502

Bad Gateway

Service temporarily unavailable

504

Timeout

Request timed out