Task failures

Branch on it. The right response ranges from "retry immediately" to "never retry, and you have been charged".

Quick reference#

Failure code	Retry?	Credits refunded?
`SAFETY.INPUT.*`	Never	No
`SAFETY.OUTPUT.*`	No	Yes
`INPUT_PREPROCESSING.SAFETY.TEXT`	Never	Yes
`INPUT_PREPROCESSING.INTERNAL`	Yes, after a delay	Yes
`INTERNAL.BAD_OUTPUT.*`	Yes, after fixing the prompt	Yes
`ASSET.INVALID`	No	Yes
`THIRD_PARTY.UNAVAILABLE`	Yes, after a wait	Yes
`INTERNAL` or null	Yes, after a delay	Yes

SAFETY.* — content moderation

Codes beginning SAFETY. mean the task was rejected by content moderation, which runs on both inputs and outputs.

SAFETY.INPUT.* means an input was rejected; SAFETY.OUTPUT.* means the generated result was. The final component indicates the likely source — SAFETY.INPUT.TEXT points at prompt text. Where several inputs would be rejected, you get the first one tested.

Treat the third component as diagnostic only. Runway makes a best-effort attempt at accuracy, but it may not correspond exactly to your inputs — you can receive SAFETY.INPUT.TEXT on a task that passed no promptText at all. Do not surface these codes to end users.

Credits are not refunded for SAFETY.INPUT.* failures, unlike every other failure type. Do not retry them: you will be charged again for the same rejection.

Repeated moderated requests can lead to account suspension. If user-supplied prompts reach the API, moderate them upstream.

INTERNAL.BAD_OUTPUT.* — rejected for quality

These generations were rejected by Runway's internal systems for quality or system-error reasons. INTERNAL.BAD_OUTPUT.01 is by far the most common.

Usual causes:

Logos, watermarks or overlaid text in the input media
A prompt that explicitly asks for text to be generated
A prompt asking for a prompt to be written rather than giving one directly — for example "write a prompt for a sunset" instead of "a sunset"

Retrying is worthwhile if you correct the prompt or input first. An unchanged retry will usually fail the same way.

INPUT_PREPROCESSING.SAFETY.TEXT

Input prompt text was rejected for content moderation reasons. Do not retry.

INPUT_PREPROCESSING.INTERNAL

Something went wrong performing content moderation itself, rather than the content failing it. Retry, but add a delay first.

ASSET.INVALID

One of your inputs is unacceptable for this task type — typically wrong dimensions, wrong duration, or another media property outside the model's limits.

Do not retry: the same asset will fail again. Check the asset against the inputs reference, particularly the per-model aspect ratio windows and duration ceilings.

THIRD_PARTY.UNAVAILABLE

A third-party-provided model failed to return output, usually because of an upstream outage or load shedding. Affects models such as veo3.1, grok_imagine_1_5, happyhorse_1_0 and gemini_omni_flash.

Do not retry immediately — an immediate retry is unlikely to succeed. Wait, then retry. If your product can tolerate it, falling back to a Runway-native model such as gen4.5 is a good degradation path.

INTERNAL, or a null value

An internal problem processing the task. Retry with a delay.

Handling failures in code#

waitForTaskOutput throws TaskFailedError when a task fails, with the full retrieve payload on error.taskDetails:

import { TaskFailedError } from '@runwayml/sdk';
 
const NEVER_RETRY = ['ASSET.INVALID', 'INPUT_PREPROCESSING.SAFETY.TEXT'];
 
try {
  const task = await client.imageToVideo
    .create({ model: 'gen4.5', promptText: prompt, ratio: '1280:720', duration: 5 })
    .waitForTaskOutput();
 
  return task.output[0];
} catch (error) {
  if (!(error instanceof TaskFailedError)) throw error;
 
  const code = error.taskDetails?.failureCode ?? 'INTERNAL';
 
  if (code.startsWith('SAFETY.INPUT')) {
    // Not refunded, never retry. Show the user a generic message —
    // the code itself is diagnostic only and may not be accurate.
    return rejectWithGenericMessage();
  }
 
  if (NEVER_RETRY.includes(code)) return reportInvalidInput(code);
 
  return scheduleRetryWithDelay(code);
}

HTTP errors — failures at the request level
Inputs — the constraints behind ASSET.INVALID
Go-live checklist — testing failure paths before launch

Handle failures before they reach users

The go-live checklist walks through every failure path this page describes.

Read the checklist Create your account

Updated Aug 15, 2026

Task failures

Quick reference#

Handling failures in code#

Related#