Fastest fix: open a brand-new chat and ask for a red apple on a white table, soft lighting, photo style. If that one image comes back, your real prompt is being blocked by the safety filter or is too contradictory — strip named people, brands, and conflicting style words (jump to Step 2). If even the apple fails, you’re out of quota or the image service is throttled — check the in-product wait message and retry in a few minutes (Step 3).
As of June 2026, ChatGPT image generation runs on gpt-image-2, the model behind ChatGPT Images 2.0 (launched April 21, 2026). The old DALL·E 2 and DALL·E 3 paths were retired on May 12, 2026, so any guide telling you to “switch back to DALL·E” is out of date. In a conversation you describe the image, the model parses your prompt, runs the image pipeline (paid tiers add a “thinking”/reasoning pass), then streams a PNG back. A failure can land at any step: the safety filter blocks the input, the output gets blocked after rendering, your quota is exhausted, or the image GPUs are temporarily throttled.
The five causes below are ordered by how often they actually bite. The first two cover roughly 70% of real cases.
Which bucket are you in?
| What you see | Most likely cause | Go to |
|---|---|---|
I can't generate that image, sorry or the image flashes then vanishes | Safety filter (input or output) | Cause 1 |
You've reached your image generation limit / a wait-time countdown | Quota / rate window | Cause 2 |
| Some details ignored, wrong style, or a generic refusal on a long prompt | Prompt conflicts / too long | Cause 3 |
Long spinner then image generation is temporarily unavailable | Capacity throttle | Cause 4 |
| Works in native chat but fails inside a Custom GPT | Stale Custom GPT config | Cause 5 |
Common causes
1. Safety filter blocks the prompt or the output
gpt-image-2 runs two filters: an input filter (while parsing your prompt) and an output filter (after the PNG renders, before it streams). Either one tripping fails the whole request. Commonly blocked content as of June 2026:
- Real named people, especially photorealistic public figures — celebrities can also formally opt out of being depicted
- Photorealistic depictions of any public figure who is a minor (hard block)
- Copyrighted characters (
Mickey Mouse drinking Coke,Pikachu playing piano) - Violence, weapons, gore, hateful imagery
- Sexual content, and anything sexualizing minors
- Editing a real person’s face in a way that reads as a deepfake
- Some brand logos and trademarks inside scenes
Symptom: it returns I can't generate that image, sorry, or the image flashes for a moment then disappears (that flash means it rendered and the output filter caught it).
How to verify: replace every named person, brand, and IP with a generic description and retry. Trump becomes a middle-aged male politician with a red cap. If the generic version works, the filter was the cause.
2. Quota or rate window exhausted
gpt-image-2 is available on every tier, but the throughput is capped per plan. OpenAI only publishes firm numbers for Pro and above; the Free/Go/Plus figures below are community-observed as of June 2026 and shift with demand:
| Tier | Approximate limit | Window |
|---|---|---|
| Free ($0) | ~2–3 images | Rolling 24h |
| Go ($8) | ~20–30 images | Rolling (not officially stated) |
| Plus ($20) | ~50 images | Rolling 3h |
| Pro ($200) | Effectively unlimited, faster | Abuse guardrails only |
| Business / Enterprise | Virtually unlimited | Per contract fair use |
Note these are rolling windows, not midnight resets: a Plus user who burns 50 images in 20 minutes waits for the oldest slots to age out of the 3-hour window, not until the next day.
Error message: You've reached your image generation limit, often with a countdown.
How to verify: the error explicitly mentions a limit or shows a wait time. ChatGPT does not expose a clean “X of Y images left” counter, so the in-product message is the source of truth.
3. Prompt contradicts itself or is too long
The pipeline has an internal prompt budget (stable well below ~4000 characters). Common failure patterns:
- Asking for
cartoon styleandphotorealisticin the same prompt - Stacking conflicting colors:
blue background, overall warm tones, green subject - Trying to specify 20+ discrete details — the model silently drops some
- Mixing a non-English prompt with technical art terms that resolve more reliably in English
gpt-image-2 is far better at following long prompts than DALL·E 3 was (and hits ~99% character-level text accuracy across Latin and CJK scripts), but genuine self-contradiction still produces a wrong or refused result.
How to verify: try the minimal probe (a red apple on a white table). If that works, your complex prompt is the problem, not your account.
4. Capacity throttle or regional limit
When OpenAI’s image GPUs are saturated — typically US daytime or new-model launch days — some accounts and regions get throttled. OpenAI also states it may lower limits during peak hours. Symptom: a long spinner, then image generation is temporarily unavailable.
How to verify: wait 10 minutes, retry with the minimal probe. If it works, it was a transient queue. Check status.openai.com for an active image-generation incident before assuming it’s your account.
5. Stale Custom GPT or wrong model
Some older third-party Custom GPTs still reference the retired DALL·E endpoint or a removed model. If image generation fails inside a Custom GPT but works in native chat, the Custom GPT’s config is stale.
How to verify: leave the Custom GPT, return to native ChatGPT, and run the same prompt.
Shortest path to fix
Run these in order — each step is a 30-second check.
Step 1: Probe with a minimal prompt
Open a new chat (the + icon) and paste:
A red apple on a white table, soft lighting, photo style.- Image returned → your original prompt has a safety or length issue → Step 2
- Still fails → quota or capacity → Step 3
Step 2: Strip sensitive or conflicting elements
Walk this checklist on your original prompt:
- Remove all named entities: celebrities, politicians, brands, IP characters → generic description
- Drop conflicting style words: keep exactly one (
watercolorORphotoreal) - Remove competing colors: specify only a subject color and a background color
- Shorten to under 100 words: one subject, one action, one style, one lighting cue
Template:
[subject], [action or state], [style], [lighting], [one or two extra details]Working minimal example:
A young woman reading by a window, watercolor illustration,
warm afternoon light, soft pastels.Step 3: Check quota and wait
Read the in-product message — it tells you whether you’re rate-limited and shows any wait time. If you’re capped:
- Free → wait for the rolling 24h window to free up, or upgrade to Go/Plus
- Plus → the 3-hour rolling window returns slots gradually; if you batched a burst, partial capacity is usually back within 1–3 hours
- Go → expect a tighter rolling cap than Plus; space out requests
- Team / Business → ask your workspace admin to check shared usage
There is no settings page that shows an exact remaining-image count, so don’t hunt for one — trust the wait message.
Step 4: Switch entry point / model
- Leave any Custom GPT and return to native ChatGPT
- In the model picker, choose GPT-5.5 (image-capable by default; on paid tiers this enables “thinking”/reasoning before the image)
- If you were on a lightweight/mini variant that lacks image tools, switch to the standard GPT-5.5
Step 5: Try a different device / network
If nothing above works, open the same account on your phone over cellular and re-send the same prompt. Success there means it’s your computer’s network (a proxy, VPN, or corporate firewall blocking the image stream). Failure there means the backend is genuinely refusing.
Step 6: Use an alternative tool
When ChatGPT keeps refusing a legitimate image:
- Google Gemini (Nano Banana Pro / Gemini 3 image) — consistently among the top-ranked models on the text-to-image arena leaderboard
- Adobe Firefly — free tier, commercially safe training data
- Microsoft Copilot / Designer (the former Bing Image Creator) — now offers Microsoft’s in-house MAI-Image-1 alongside the older DALL·E 3 engine
- Midjourney — subscription, strongest for stylized art
How to confirm it’s fixed
- The minimal probe returns a clean PNG in a fresh chat.
- Your real prompt — with named entities and conflicting modifiers removed — returns an image without the flash-and-vanish behavior.
- No
temporarily unavailableorreached your limitmessage appears on retry.
If 1 succeeds but 2 still fails, the remaining blocker is content in your prompt, not your account — keep stripping named people and IP.
Prevention
- Save prompts that worked once; the same prompt usually still works the next day.
- Don’t stack 6+ style modifiers — pick at most three (e.g. style + lighting + camera angle).
- Use the official model picker, not a stale Custom GPT.
- For commercial work, avoid real names, brands, and copyrighted characters — both to clear the safety filter and to avoid downstream IP risk.
- Reuse prompt templates and tweak parameters instead of re-rolling, so you don’t burn your rolling-window quota.
FAQ
Why does my image flash on screen and then disappear? It rendered successfully, then the output safety filter blocked it before streaming. Something in the result — usually a recognizable real person, logo, or restricted content — tripped the post-generation check. Regenerate with named entities replaced by generic descriptions.
ChatGPT used to let me pick DALL·E — where did it go?
DALL·E 2 and DALL·E 3 were retired on May 12, 2026. All image generation now routes through gpt-image-2 (ChatGPT Images 2.0). There’s no separate DALL·E option to select anymore.
How many images can I generate per day on Plus? As of June 2026, Plus is community-observed at roughly 50 images per rolling 3-hour window — not a fixed daily number. Slots reopen as your oldest requests age past the 3-hour mark, so spacing requests out gets you more total images than batching them.
My prompt has no banned content but still gets refused. Why?
Two common reasons: the prompt self-contradicts (e.g. cartoon plus photorealistic), or a generic noun is matching a trademark or real person the filter recognizes. Run the minimal probe first; if it works, simplify your real prompt one element at a time.
Is the image service down, or is it just me? Check status.openai.com. If there’s no active incident and the minimal probe works on a second device, the issue is your prompt or your local network, not OpenAI.
Does image generation work on the Free tier?
Yes. gpt-image-2 is available on every tier including Free, but Free is capped at roughly 2–3 images per rolling 24 hours and does not get the paid “thinking”/reasoning pass.
Related
- AI image distorted faces
- AI image style inconsistent
- ChatGPT beginner guide
- ChatGPT prompt improvement
- ChatGPT model selection guide
Tags: #ChatGPT #Debug #Troubleshooting