Troubleshooting
Every entry here is something that actually happened, with the cause that was actually found — not a guess at what might go wrong.
Your settings and keys
My keys and settings disappeared after restarting the app
If you were on the 1.0.0 desktop app, that was the app. It served itself on a different local port every launch, and the browser engine inside keeps saved data per port — so every restart opened a new, empty store. Nothing had failed to save; each launch looked somewhere different.
1.0.1 fixes it. Keys you typed into 1.0.0 are stranded under the old ports and are not recovered, so enter them once more. Then press Back up to a file in Settings.
“Could not read your manifest” or “…your settings and keys”
The app found stored data it could not read. It has not overwritten it: the original is kept under a .rescue copy, and saving is paused until you act. Press Try again first. If you have a backup file, Restore it.
This exists because older versions, on an unreadable read, fell back to defaults and then saved those defaults over your data a moment later.
Settings say “NOT saved”
Every save is read straight back to check it stuck. If it did not — storage full, a private window, something blocking site data — you see this instead of a green “Saved and checked”. Press Back up to a file before closing.
“Your prepayment credits are depleted” — but there is money in the account
This is the one that costs people an evening. Linking your project to Google Cloud billing is what causes it.
In order:
- A new project is on the free tier. It works.
- You link it to a Cloud billing account — to use a voucher, to use trial credit, or because something told you to.
- Google now marks the project paid tier.
- Paid tier bills against a separate balance called Prepay – AI Studio.
- Ordinary Google Cloud credit does not fund that balance. Neither does free trial credit, and nor do most vouchers.
- The balance is therefore zero, and every request is refused with a
429.
Two fixes, and you have to pick one:
| You want | Do this |
|---|---|
| The free tier back | Unlink the project from its billing account at console.cloud.google.com/billing |
| To actually pay | Add a Prepay – AI Studio balance at ai.studio/projects |
What does not work is being linked to billing with only Cloud credit.
Confirmed on a real account: a project with €262 of Cloud credit refused every call on every model. Unlinking restored it.
Which advice you get depends on the pool. A key in the free pool is told how to get the free tier back. A key in the paid pool is told “Paid tier is on, but the AI Studio prepay balance is empty” — and not to unlink, because the free tier has no picture allowance at all.
A paid key “works” but no picture is ever made
The key check says “The key works. Whether it has picture money, this cannot tell you.” That is deliberate. Text has a free allowance and pictures do not, so a key with an empty prepay balance passes a text check and still refuses every picture. The only real proof costs one picture, and the app never spends that behind your back — run one row when you want to know.
“Your project has been denied access. Please contact support.” (403)
A project-level block, not a billing state. More credit will not fix it, and neither will unlinking. Verified not to be a browser problem: the same key fails identically through a server-side proxy. Make a key in a fresh project, or take it up with Google.
A key lists 50 models and then refuses everything
Expected, and why the key check makes a real call. Listing models succeeds on a key that can do nothing else.
Model names that a chatbot suggested do not exist
On a live account, gemini-1.5-flash, gemini-1.5-pro and gemini-2.5-flash return 404 — This model is no longer available to new users. Press Load models and use what your key actually lists.
Cloudflare
Cloudflare did not work in the 1.0.0 desktop app
The part of the desktop app that forwards requests to Cloudflare crashed on the first one. It only worked when running from source. Fixed in 1.0.1, checked on a real install.
“Could not reach Cloudflare”
Cloudflare’s API sends no CORS headers, so a browser page cannot call it. The app forwards the request in both the dev server and the desktop app. If you see this, you have opened the built HTML file directly — use npm run dev or the desktop app.
Every Cloudflare image fails at once
If a request includes seed, flux-1-schnell rejects the whole request. Cloudflare’s own model page lists seed. It does not work. The app does not send it. Confirmed on a live account, 2 September 2026.
403 on a Cloudflare token
The token needs Account → Workers AI → Read.
OVHcloud
“OVHcloud is still busy after three tries”
It allows two pictures a minute without a key. The app already paces itself to that — one every 31 seconds, however many run at once — so this only appears when it stayed busy anyway. Wait a minute and run again.
My OVHcloud pictures are all square, and the seed does nothing
Correct. It accepts only a prompt and a negative prompt. Every picture is 1024×1024, and the same prompt twice gives two different pictures.
NVIDIA and other text providers
NVIDIA says “Failed to fetch”
NVIDIA answers a browser’s permission check without the header that allows it, so the browser refuses — not your key, not the address. The app forwards NVIDIA requests the same way it does Cloudflare. Checked 4 September 2026.
“Only show models that are free” does nothing for NVIDIA or Google
They do not publish a price per model, so there is nothing true to filter on. The app says so rather than guessing from model names. OpenRouter does publish prices, and there the filter is exact.
Pollinations
“Missing Turnstile token”
Pollinations stopped serving anonymous requests. A free token from auth.pollinations.ai fixes it.
Vision and lettering
“That model cannot look at pictures”
The model is text-only. Pick one that can see. The vision check sends a real image — a solid red square — and asks what colour it is.
Pixtral returns 404
pixtral-large-latest is not on Mistral’s model list any more (checked 2 September 2026). Use mistral-medium-latest, which can see and is free on Mistral’s tier.
The lettering lands in the wrong place
The model proposes, you drag. With no vision model set up, the free Find a quiet spot places the box in the calmest part of the picture.
Files
My file ends in .jpg, not .png
That is the real format. Google and Cloudflare send JPEG, and the app names a file for what it is. Older versions wrote .png on JPEG data, which confuses WordPress uploads and strict tools.
The app
Windows says “Windows protected your PC”
Expected. The app is not signed with a paid code-signing certificate. Click More info, then Run anyway. Once only.
“Image Forge could not start”
It could not open its local port (47821–47825). Close any other copy of Image Forge, or whatever else is using those ports.
“Remove your Image Forge data as well?” — did it remove anything?
In 1.0.0, no: it pointed at a folder that never existed, whatever you answered. From 1.0.1 it removes %APPDATA%\image-forge when you say Yes, and a silent uninstall always keeps your data.
A picture was made but the row says failed
Fixed. The row now stays done and a separate message says writing to your folder is what failed, because on a paid engine “failed” invites paying twice.
The paid dialog said far more pictures than I was paying for
Fixed in 1.0.1. A queue mixing paid and free rows was described as if every row were paid. It now counts only billed rows and says how many are free.
Text in a picture is gibberish
Most image models cannot spell. Use a Nano Banana model, or add the words with the Letterer, which uses real fonts.
“Every key is resting”
The whole pool hit its limit. Rows re-queue automatically when the cooldown expires. Or switch engines and run again.
Still stuck
Open an issue. Include what you pressed and what it said.
Never paste an API key into an issue. The key check in Settings describes a key without revealing it.