Fix Gemini Image Generation Errors: 400, 429, 503 and Retry Rules

Gemini image generation errors come from two places, and the fix depends on which one you're in. If the Gemini app on the web, Android or iPhone won't make or edit an image, the cause is almost always your account, a usage limit or a policy refusal, and sending the same request again rarely helps. If your code calls the Gemini API, the response carries an error code, and Google's rules say what to do with it: retry 429 and 5xx errors with exponential backoff, and never retry 400, 402 or 403, because those only clear after you change something.

App users need only the first section. Developers can skip to the error-code table. Codes, retry rules and model IDs below follow Google's API errors reference and troubleshooting guide as of October 5, 2026.

Why Gemini isn't letting you generate images: six app causes

Match what you see on screen to a row, then take that row's next step. The requirements come from Google's Gemini image help page and usage limits page.

The causes are the same on iPhone, Android and the web, so switching devices only helps with the last row. Numbers such as "Something went wrong (5)" or "(13)" have no published meaning. Nano Banana Pro error 5 shows how to tell whether the failure is on your side or Google's.

Two app rules catch Pro users. Redo with Pro needs a Google AI plan (Plus, Pro or Ultra), and once you hit the daily Nano Banana 2 quota, you can't redo images with Pro either. Paid plans raise the shared usage limit: Plus 2x, Pro 4x, and Ultra higher still, per Google's limits page. Google publishes no fixed number of images per day.

A policy refusal has no workaround, and you shouldn't look for one. Google's Generative AI Prohibited Use Policy lists "circumvention of abuse protections or safety filters" as a violation in its own right, and repeated violations can restrict your Google account.

Is Gemini image generation down right now? How to tell

Run one control test. Ask for something plain, like a bowl of fruit on a table, in a new chat with no uploaded files. If that fails too, try another device and, if you have one, another Google account. Failures everywhere point to Google's side, so wait and try later. If only your original request fails, the cause is in that request or that account, and the table above applies.

For the API, check the Google AI Studio status page. Vertex AI incidents appear on Google Cloud Service Health. A run of 500 or 503 errors on small, valid requests while the status page shows an incident means you should wait rather than change your code.

Gemini API error codes for image generation: fix, wait or retry

Read the error code before you touch the prompt. The Code column is the code string returned by the Interactions API, which Google's image guide now uses for its examples. generateContent reports the same situations as an HTTP status plus a status value in the error body, such as RESOURCE_EXHAUSTED.

A few rows need more detail:

• 402 stops everything on the billing account. When the Prepay balance reaches $0, Google's billing page says every API key in every project linked to that billing account stops working until you add credits. Retrying only adds noise to your logs.
• 404 often means a retired model ID. On the Gemini API, gemini-3-pro-image-preview and gemini-3.1-flash-image-preview were shut down on June 25, 2026, the Imagen 4 models on August 17, 2026, and gemini-2.5-flash-image on October 2, 2026, according to Google's deprecations page. Older tutorials still use these IDs.
• A leaked key stops working. Google blocks keys that have been published, with the message "Your API key was reported as leaked. Please use another API key." Create a new key in AI Studio and keep it out of public code.
• Failed 400 and 500 requests aren't billed, but they still count against your quota. A loop that resends a broken request uses up your limit without producing an image.
• Region matters. The Gemini API and AI Studio work only in Google's list of available regions, which doesn't include mainland China, Hong Kong or Russia. AI Studio also requires users to be 18 or older.

Blocked or empty results: IMAGE_SAFETY, IMAGE_OTHER and NO_IMAGE

Some failures arrive with HTTP 200 and no image. In generateContent, look at finishReason on the candidate and promptFeedback.blockReason on the response. In the Interactions API, the same cases come back as the error codes below.

For the blocked codes, Google's instruction is to "modify your input and retry." Resending the same request is pointless. Adjustable safety filters are already off by default for current Gemini models, so setting BLOCK_NONE or OFF doesn't lift IMAGE_SAFETY, IMAGE_PROHIBITED_CONTENT or IMAGE_OTHER. That follows from the defaults in Google's safety settings guide, and the Interactions API doesn't accept custom safety settings at all. Developers have reported IMAGE_OTHER on plainly harmless prompts in Google's AI developer forum. If yours is one of them, simplify the request and report the case there rather than rewording it to slip past a filter. Nano Banana Pro's safety filters explains which settings you can and can't change.

429 on image models: rate limit, daily quota, spend cap or limit 0

A 429 can mean four different things, and only the first one goes away with a short wait. Google's rate limits page applies to all of them.

1. Per-minute limit (rate_limit_exceeded). Image models have an images-per-minute limit on top of requests and tokens per minute. Limits apply per project, not per API key, so adding keys to the same project changes nothing. Your current numbers are shown in AI Studio.
2. Daily quota (quota_exceeded). Requests-per-day quotas reset at midnight Pacific time. Retrying before then only produces more 429s.
3. Spend-based limit. Depending on your billing history, Google also caps spend over a rolling 10-minute window: $10 on Tier 1, $50 on Tier 2 and $200 on Tier 3. Hitting the cap returns 429 RESOURCE_EXHAUSTED. At Nano Banana Pro's standard price of $0.134 per 1K or 2K image, $10 covers about 74 images ($10 ÷ $0.134 = 74.6), so a fast batch on a new Tier 1 project can reach the cap in minutes.
4. Limit 0. Image output has no free tier on any current Gemini image model, so an image request from a project without paid billing fails on quota. Developers report seeing free_tier_requests with a limit of 0. If your project is paid and still shows it, read paid tier getting free_tier_requests limit 0. For what image output costs, see whether the Gemini image API is free.

The full 429 walkthrough is in how to fix the Gemini image 429 rate limit. For Nano Banana Pro batch jobs that stop mid-run, see Nano Banana Pro RESOURCE_EXHAUSTED.

Retry rules from Google: back off on 429 and 5xx, stop on 400, 402, 403

Google's troubleshooting guide sets four rules for retries:

• Retry only transient errors: 429, 408 and 5xx.
• Never retry 400, 402 or 403. They mean a bad request, an empty Prepay balance or a key problem.
• Use exponential backoff, for example 1, 2, 4 and then 8 seconds, and add random jitter so clients don't all retry at the same moment.
• Set a maximum number of attempts.

Don't assume the Python SDK retries for you. Google's troubleshooting guide says it "automatically retries transient errors up to four times with an initial delay of approximately 1 second and a maximum delay of 60 seconds." That isn't the default in google-genai 2.28.0. In a test on October 5, 2026 against a local server that always returned 429, a default genai.Client() sent one request and raised the error at once. That test used client.models.generate_content; client.interactions.create in the same version behaved differently and retried a 429 on its own (four requests in about three seconds). Retries started only after they were switched on explicitly:

from google import genai
from google.genai import types

client = genai.Client(http_options=types.HttpOptions(retry_options=types.HttpRetryOptions()))

With that setting, the same 429 produced 5 requests (the first plus 4 retries), with delays starting around 1 second and capped at 60 seconds. The JavaScript, Go and Java SDKs weren't part of that test, so check their retry settings in your own version.

Pick one retry layer. Either enable HttpRetryOptions and let the SDK handle 429 and 5xx, or keep the SDK default and run your own loop. If you do both, one image can turn into 25 requests (5 outer attempts × 5 SDK requests each), and every one of them counts against the same per-project limit.

For direct REST calls, or SDK calls without HttpRetryOptions, this helper encodes Google's rules. It contains no API call and runs as written:

import random

def should_retry(status: int) -> bool:
    """Google's rule: retry 408, 429 and 5xx; never 400, 402 or 403."""
    if status == 501:  # unimplemented: the feature isn't supported, so retrying can't help
        return False
    return status in (408, 429) or 500 <= status < 600

def backoff_delays(max_retries: int = 5, base: float = 1.0, cap: float = 60.0):
    """Exponential backoff with full jitter: up to 1s, 2s, 4s, 8s, 16s."""
    for attempt in range(max_retries):
        yield random.uniform(0, min(cap, base * 2 ** attempt))

# Example: a 503 is retried, a 402 is not
print(should_retry(503), should_retry(402))  # True False

Call your request function, check should_retry on its status, and sleep for the next value from backoff_delays before trying again. A quota_exceeded 429 is the exception to the backoff rule: when the message says the daily quota is used up, stop until the reset instead of burning through the retry budget.

Request checklist before you resend a 400

Most 400s on image calls come from a few fields. Check these before changing your prompt:

• Model ID. Use gemini-3.1-flash-image (Nano Banana 2), gemini-3-pro-image (Nano Banana Pro) or gemini-3.1-flash-lite-image (Nano Banana 2 Lite). Preview IDs and gemini-2.5-flash-image are shut down.
• Image size. Use an uppercase K: "1K", "2K" or "4K". Google's image guide says lowercase values such as "1k" are rejected. "512" works only on Nano Banana 2, and Nano Banana 2 Lite outputs 1K only.
• Where the size goes. In generateContent, put it in generationConfig.imageConfig.imageSize and aspectRatio. In the Interactions API, it goes in response_format as image_size and aspect_ratio. Sending one interface's fields to the other can return parameter_unknown or invalid_request.
• Endpoint version. Call https://generativelanguage.googleapis.com/v1beta/models/MODEL_ID:generateContent. Google notes that Beta features exist only under /v1beta.
• Response modalities. The API returns an error if responseModalities doesn't match a combination the model supports. Copy the combination from Google's current example for your model.
• Other parameters. Candidate count must be 1–8. Google recommends leaving temperature, topP and topK at their defaults for Gemini 3.x models.
• Multi-turn editing. Send the previous model turn back exactly as you received it. Removing parts or rebuilding the image part drops its thought signature.

The Gemini image API guide has working requests for all three current models.

Which guide to open for each Gemini image error

• 429 on any Gemini image model: Gemini image 429 rate limit
• Nano Banana Pro RESOURCE_EXHAUSTED during a batch: resume Nano Banana Pro image jobs after a 429
• Paid project still shows limit 0: free_tier_requests limit 0 on a paid tier
• 503 "model overloaded" or "deadline expired": fix Gemini 3 Pro Image 503 errors, or the Gemini high demand error for the app message
• IMAGE_OTHER or blockReason OTHER: Nano Banana 2 blockReason OTHER
• HTTP 200 without an image: Nano Banana API returns no image
• Thought signature errors in multi-turn edits: Nano Banana 2 thought signature error
• Every Nano Banana Pro error in one place: Nano Banana Pro errors and troubleshooting hub