Correction, 2026-10-01: An earlier version said this message appears for every failure; a few provider errors (Kling's "Image recognition failed", Runway's messages) are shown as written.
Correction, 2026-09-29: the first version of this page told you to contact support to get credits back and warned that every retry is another charge. That was wrong. On 8frame the credits for a failed generation come back automatically, in the same step that marks it failed. It also said size and format limits cause most failures; our logs say content filters do. Both are fixed below.
Most failed generations have one of five causes: a model's content filter, an input the model cannot use, a provider outage, a retired model, or a regional block. On 8frame you pay nothing for any of them, because the credits are returned automatically. What a failure does cost is the wait, so it is worth finding the cause instead of retrying.
TL;DR
- Content filters cause most failures we see. The single biggest pattern: Seedance rejecting a reference image with a close-up face.
- Credits come back on their own. Every failed generation on 8frame in the 120 days to 2026-09-28 was refunded automatically.
- One message covers almost every cause on 8frame: "Sensitive content detected. Please try changing your prompt or image." It appears for outages and oversized files too (only a few provider errors, such as Kling's "Image recognition failed", are shown as written), so do not read it as a verdict on your content. Use the checklist below.
- Images that are too big or in an odd format are fixed for you. Images that are too small are not.
- A retired model fails permanently. Sora's provider shut down on 2026-09-24.
What our failure logs show
Ranked from the most common, over the 120 days to 2026-09-28:
- Seedance with a reference image. By far the largest group. Seedance runs with a reference image failed more than twice as often as runs without one. The trigger we confirmed in testing is a face in close-up; a waist-up shot of the same person passes. Full explanation in Seedance "flagged as sensitive" (E005).
- Other content filters. GPT Image, Nano Banana and Kling each refused some prompts or reference images. What each refusal message means is in AI generation refusal messages, decoded.
- Provider errors. "Downstream service error" and similar: the provider had a problem, not your input.
- Files a provider would not take. A handful of uploads over a provider's 10 MB limit.
- Prompts over a model's limit. Midjourney rejects prompts of 2,000 characters or more. Limits for every model are in AI prompt length limits by model.
Check the input
Too small is the one hard stop. Each model has a minimum image size, and nothing is enlarged automatically: Veo 3.1 needs 1280 by 720, most others 300 to 512 pixels a side. The full table is in upload limits by model.
Too large is handled for you. The canvas takes images up to 50 MB and resizes them for models with lower limits, so you rarely need to shrink a photo yourself.
Format. The canvas takes PNG, JPG, WebP, GIF, AVIF and BMP, and converts them for models that want something else. It does not take HEIC, the iPhone's default photo format. Set the iPhone to Settings, Camera, Formats, Most Compatible, or convert the file to JPEG.
Video input length. Video is not trimmed for you. Gemini Omni Flash takes up to 10 seconds of uploaded video, per Google's documentation; Grok Imagine up to 15; Happy Horse up to 60.
Duration and resolution combinations. Wan 3.0 supports 2 to 30 seconds. Asking for something outside a model's range fails, so check the selector rather than assuming.
Regional restrictions
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.
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. See Sora 2 alternatives for where to land.
Provider outages
Every model on the canvas runs on a provider's infrastructure. When that provider has an incident, generations fail through no fault of your prompt or input.
How to tell: the same prompt failing on one model while a different model works is usually a provider problem. Switch models or wait; your credits for the failed attempt are already back.
Content filters
Every model has its own filter, and they do not agree. Anything involving people, violence or named individuals is where refusals cluster. Rewording often works for a prompt; for a reference image, change the image.
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
When a generation fails, 8frame returns the credits it cost automatically, as part of marking it failed. You do not need to ask. In the 120 days to 2026-09-28, every failed generation on 8frame was refunded this way. Details in do failed AI generations cost credits.
What is not refunded: a generation that completed but that you do not like. That is a finished run, not a failure.
One pricing default is worth knowing, 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. Prime has the same default and costs 189 at it, so set the resolution on either.
The diagnostic order
- Check the input. Above the model's minimum size, an accepted format, video within the length limit.
- 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 or a filter specific to the first model.
- Change the reference image if one is attached, especially a close-up face on Seedance.
- Reword the prompt if you suspect a content filter.
- Check whether the model is still current. Retired models fail permanently.
- Contact support only if a failed run's credits did not come back.
FAQ
Do I get credits back for a failed generation? Yes, automatically. On 8frame the credits return in the same step that marks the generation failed.
Why does 8frame say "Sensitive content detected" when nothing was sensitive? It is the message 8frame shows for almost every failure, including provider outages and file problems; a few provider errors, such as Kling's "Image recognition failed" and Runway's own messages, are shown as the provider wrote them. Work through the diagnostic order above to find the actual cause.
Why does my iPhone photo fail to upload? Most likely it is HEIC, which the canvas does not take. Convert it to JPEG or set the iPhone camera to Most Compatible.
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 change the reference. The 8frame canvas is free and unlimited; generation is paid from $19/month.