GPT Image 2 Not Working? 7 Fixes for Failed Generations
By Peter Zhu, Founder ·
When GPT Image 2 fails, the error rarely tells you why. In practice the causes fall into seven buckets — and the most common one isn’t a bug at all, it’s the type of credits you’re spending. Match the symptom you’re seeing to the section below and jump straight to the fix.
July 22, 2026 · 6 min read
On this page
- 1. “Insufficient credits” — but your balance isn’t zero
- 2. It worked at 1K but fails at 4K
- 3. The generation is rejected the instant you hit Generate
- 4. It spins for minutes, then fails with a generic error
- 5. The text in the image comes out garbled
- 6. It ignores part of a long prompt or edit instruction
- 7. Your reference image is the problem
- FAQ
1. “Insufficient credits” — but your balance isn’t zero
Symptom: The generate button errors with a credits message even though you clearly have a balance.
GPT Image 2 is a premium model, and premium models only run on paid credits. Free credits — the 20 you get at signup and the 10 from each daily check-in — work with Shotari’s own Basic and Pro models, not with GPT Image 2, the Nano Banana family, Seedream, or any video model. So a balance made entirely of free credits reads as “insufficient” the moment you pick GPT Image 2, even though the number on screen isn’t zero.
Fix: Check what kind of credits you hold. To keep using free credits, switch the model to Shotari Basic or Shotari Pro. To run GPT Image 2, add a credit pack (from $29.90) or a subscription (from $20/mo) — both unlock every premium model, remove the watermark, and include commercial rights.
2. It worked at 1K but fails at 4K
Symptom: A standard generation goes through, then the same prompt at a higher resolution errors on credits.
GPT Image 2 costs 5 credits at standard (1K) resolution, but resolution multiplies that: 2K is roughly 1.5× and 4K is 2×, so a single 4K generation costs about 10 credits. If your paid balance sits between 5 and 10, standard output succeeds and 4K quietly fails — it’s a cost jump, not a bug.
Fix: Generate at 1K first to lock in the composition, then re-run your favorite at 2K or 4K once. You only pay the higher rate on the final upscaled pass, not on every draft — so drafting cheap and finishing high is the right workflow.
3. The generation is rejected the instant you hit Generate
Symptom: No spinner, no wait — the request bounces back immediately, often mentioning content policy.
Every prompt and every uploaded image is screened before GPT Image 2 ever runs. Requests naming real public figures, brands, logos, violence, or suggestive content get blocked, and the filter is deliberately cautious — so ordinary words occasionally trip it (“shoot a portrait” is a classic false positive).
Fix: Reword the flagged idea instead of resubmitting the same prompt: describe the style, not the celebrity (“a 1970s rock frontman in a leather jacket”); describe the mood, not the violence. If an uploaded photo was blocked, crop out logos, on-image text, or bystanders and try again.
4. It spins for minutes, then fails with a generic error
Symptom: The generation runs a long time and returns nothing, or “no image was returned.” Retrying sometimes works.
Upstream capacity gets tight after model launches and during US evening hours, and it surfaces as timeouts or empty results. Nothing is wrong with your prompt or account — and because credits are only charged for completed generations, a failed run leaves your balance untouched.
Fix: Wait a minute and retry once. If it keeps timing out, switch to a sibling for the session — GPT Image 1.5 is a faster OpenAI model with clean text, and Nano Banana 2 runs on entirely different infrastructure, so an outage on one rarely hits the other.
5. The text in the image comes out garbled
Symptom: Everything looks right except the lettering — misspelled words, dropped characters, or mangled small type.
Rendering readable text is one of GPT Image 2’s strengths, but it still has limits. Long paragraphs, tiny type, unusual fonts, and multiple text blocks in one image are where it breaks down — the more words you ask for, the more likely a character slips.
Fix: Put the exact words in quotes in your prompt (a poster that reads “Grand Opening”), keep it to a few short lines, and specify placement (“centered headline, small caption below”). If a word still comes out wrong, regenerate — text is one of the higher-variance parts of any image model, and a second pass usually lands it.
6. It ignores part of a long prompt or edit instruction
Symptom: The generation succeeds but skips half of what you asked, or an edit to a reference photo isn’t applied.
This is the failure people misread as “the model is broken.” GPT Image 2 has strong instruction-following, but a prompt stuffed with ten requirements — or contradictions like “minimalist poster with lots of ornate detail” — forces it to drop some. On edits, vague instructions (“make it better”) give it nothing concrete to change.
Fix: One change per generation. Get the composition right first, then fix lighting, then text — iterating beats one giant prompt. For edits, name the specific change (“replace the background with a plain studio grey, keep the subject untouched”). Our prompt-writing guide has a 5-part formula that prevents most of these misses.
7. Your reference image is the problem
Symptom: Text-only prompts work, but attaching a photo makes the generation fail or return something unrelated.
Oversized files, exotic formats, and corrupt images are the usual suspects. Screenshots of photos (rather than the photo file itself) and heavily compressed images also confuse the model into ignoring or misreading the reference.
Fix: Use JPG, PNG, or WebP under a few megabytes, and export a fresh copy instead of re-uploading a screenshot. GPT Image 2 accepts up to 3 reference images on Shotari — if a combination fails, test each image on its own to find the bad one.
Still stuck?
Isolate the problem by running the same prompt on a sibling model — if it works on GPT Image 1.5 or Nano Banana 2 but not GPT Image 2, it’s capacity, not you. Our prompt-writing guide fixes the ignored-instruction class of failures for good, the GPT Image 2 vs Nano Banana Pro comparison helps you pick the right model before you spend credits, and pricing shows exactly which credits unlock premium models.
FAQ
Why does GPT Image 2 say insufficient credits when I have a balance?
GPT Image 2 is a premium model that only runs on paid credits. Free credits (signup and daily check-in) work with Shotari's own Basic and Pro models. Add a credit pack or subscription to run GPT Image 2, or switch to a Shotari model to keep using free credits.
How many credits does GPT Image 2 cost?
5 credits at standard (1K) resolution. Higher resolutions multiply the cost — 2K is about 1.5× and 4K is 2×, so a single 4K generation costs roughly 10 credits. Draft at 1K, then upscale your favorite once.
Do I lose credits when GPT Image 2 fails or times out?
No. Credits are only charged for completed generations. If a run errors out, times out, or returns no image, your balance is untouched — so you can safely retry.
Why is the text in my GPT Image 2 image misspelled?
Text rendering is a strength but not perfect. Keep it short, put the exact words in quotes, and specify placement. Long paragraphs and tiny type are the most likely to break — regenerate once and it usually lands.
GPT Image 2 keeps failing — what should I switch to?
Try GPT Image 1.5 (a faster OpenAI model with clean text) or Nano Banana 2, which runs on different infrastructure. If your prompt works on a sibling model but not GPT Image 2, the issue is capacity, not your prompt.