A failed generation usually has a mundane cause: an input outside a model's limits, a region restriction, a provider that is down, or a model that has been retired. Working out which one you hit takes a minute and prevents you from retrying something that will fail identically.
TL;DR
- Input size and format limits cause most failures. Several models cap uploads at 10MB
- Duration and resolution combinations are not all valid on every model
- Regional restrictions exist and are hard blocks, not settings
- A retired model fails permanently. Sora's provider shut down on 2026-09-24
- If credits were deducted for a failed generation, contact support. Do not keep retrying
Check the input first
The most common cause, and the easiest to fix.
File size. Several models on this canvas accept a maximum of 10MB for an input image, which is smaller than a photo from a modern phone. Resize before uploading.
Format. Accepted formats are typically jpg, jpeg, png, webp, gif and avif. A HEIC file straight from an iPhone is a common failure.
Video input duration. Gemini Omni Flash caps uploaded video input at 10 seconds, per Google's own documentation. Longer input fails rather than being trimmed.
Duration and resolution combinations. Wan 3.0 supports 2 to 30 seconds; asking for something outside a model's supported range fails. Check the selector rather than assuming.
Regional restrictions
These are real and they are hard blocks rather than settings you can change.
Gemini Omni Flash: editing or extending uploaded video is unavailable in the EEA, Switzerland and the UK. From Google's model documentation. If you are in one of those and trying to edit uploaded video on that model, it will not work regardless of what you change.
Generation from text is unaffected. It is specifically the editing and extending path.
A retired model
Sora's provider shut down on 2026-09-24. OpenAI's deprecations page lists that date for the Videos API and every sora-2 model, announced 2026-03-24, with no recommended replacement.
Anything routed to Sora after that date fails, and retrying will not help. If you had a workflow depending on it, see Sora 2 alternatives for where to land.
More generally, model rosters change. A model that worked last month can be delisted by its provider.
Provider outages
Every model on this canvas runs on a third-party provider's infrastructure. When that provider has an incident, generations fail through no fault of your prompt or your input.
How to tell: the same prompt failing repeatedly on one model while a different model works is usually a provider problem. Try a different model; if that succeeds, wait rather than retrying.
Content filters
Models have their own safety filters and they are not always predictable. A prompt can be rejected for reasons that are not obvious, particularly anything involving people, violence, or named individuals.
Rewording usually resolves it. If you are trying to generate a real person's likeness, that refusal is working correctly and you should not route around it: likeness generation requires the person's documented consent, and public figures are off limits entirely.
About your credits
If credits were deducted for a generation that produced nothing, contact support rather than retrying repeatedly. Each retry is another charge, and five retries of something that fails deterministically is five charges for nothing.
One thing worth knowing about how pricing works here, because it explains an unexpected charge that is not a failure: Wan 3.0 base defaults to 1080p, its most expensive tier. A 5-second clip costs 24 credits at 480p and 95 at 1080p. If a generation succeeded but cost four times what you expected, that is the resolution default rather than a billing error. Prime has the same default and costs 189 at it, so set the resolution on either.
The diagnostic order
- Check the input. Size under the model's cap, an accepted format, duration within range.
- Check for a regional restriction if you are in the EEA, Switzerland or the UK and editing uploaded video.
- Try a different model with the same prompt. Success means a provider problem; failure means your input or prompt.
- Reword the prompt if you suspect a content filter.
- Check whether the model is still current. Retired models fail permanently.
- Contact support if credits were taken. Stop retrying.
FAQ
Do I get credits back for a failed generation? Contact support with the details rather than retrying. Repeated retries of a deterministic failure cost more.
Why does my iPhone photo fail to upload? Likely HEIC format or file size. Convert to jpg or png and check it is under the model's cap, which is 10MB on several models.
Why does editing uploaded video fail for me in Europe? Gemini Omni Flash does not support editing or extending uploaded video in the EEA, Switzerland or the UK.
Why does Sora fail now? Its provider shut down on 2026-09-24. There is no replacement and retrying will not help.
Check the input, try another model, then ask support. The 8frame canvas is free and unlimited, and generation is paid from $19/month.