Help
Why generations can fail, what each error means, and what you can do about it.
These categories are based on real FlashBoards production failures. Most failed generations fall into moderation, provider busy/unavailable, provider processing errors, and timeouts.
If a generation fails for any reason, you are not charged. Your balance is fully restored once the failure is confirmed. In most cases this happens within seconds, but some stuck tasks may take up to an hour to resolve through our background reconciliation process.
For more on how billing works, see Generation Billing.
🚫Content blocked by safety filter
38.6% of failed generations
Why it happens. Your prompt, image, video, or audio triggered a provider moderation rule. This is the single biggest failure category in production, and different models enforce different safety policies.
What to do. Rephrase the prompt, use a different input, or try another model. Common triggers include real people, celebrities, faces in restricted models, explicit content, and sensitive topics.
⚠️Invalid input or unsupported media
0.4% of failed generations
Why it happens. Something about the input did not match what the model expects. Real examples from production include images over pixel limits, mismatched frame rates or aspect ratios, videos that are too long, and reference media that does not meet the model's format rules.
What to do. Try resizing or recompressing the media, shortening the video, using matching aspect ratios, or switching to a cleaner source file. If the model requires a clear face or body framing, use an input that matches that requirement.
⏳Service temporarily unavailable
26.5% of failed generations
Why it happens. The provider is overloaded, rate-limited, under maintenance, or having temporary account-capacity issues. This is one of the most common production failure buckets, especially on popular video models.
What to do. Wait a bit and retry. If the same model stays busy, switch to another model or come back later. These failures are usually temporary and are not charged.
⏱️Request timed out
4.3% of failed generations
Why it happens. The provider took too long to create, run, or return the generation. Timeouts happen more often on longer video jobs and on models that are already under heavy load.
What to do. Try again. Shorter durations, fewer inputs, or simpler prompts can help. If the request may have continued in the background, check Generation History before running it again.
🛠️Provider error during generation
20.0% of failed generations
Why it happens. The model accepted the request but failed internally while processing it. In production this often appears as messages like 'failed to process task', 'generation failed', 'model response exception', or other provider-side internal errors.
What to do. Retry once or twice. If the same model keeps failing, switch models. If the failure repeats with the same message, contact support with the model name and error text so we can inspect the provider-side failure.
⚠️Upload or media fetch failed
0.6% of failed generations
Why it happens. The input media could not be uploaded or fetched reliably. This can happen when a remote URL fails, the source file is too large, the provider cannot access the media, or the browser has already cleared a local file.
What to do. Re-add the file, try a smaller or simpler input, or upload the media again from your device. If you used a remote image or video URL, download it locally first and then upload it to FlashBoards.
AI models are hosted by external providers, and each has different reliability levels. Newer or more popular models tend to have higher failure rates because of limited capacity and frequent updates on the provider side.
Video generation models (Sora, Kling, Veo, Wan) are more complex than image models and generally take longer, time out more often, and are more sensitive to prompt content. This is normal for current AI video technology.
In the last 60 days, the biggest production failure buckets were moderation/safety blocks first, provider busy or unavailable second, and provider-side processing failures third. Browser-side network failures exist, but they are a much smaller share of the total.
Half my Sora generations failed. Why is the success rate so low?
Some models fail more often than others. Large commercial models like Sora, Veo, and Seedance tend to have stricter content restrictions and higher rejection rates. Combined with capacity limits during peak hours, this leads to more failures overall. You are not charged for any of those failed attempts.
The generation "succeeded" but the output has artifacts or frozen frames. Is that a failure?
If the model delivered output, it counts as a successful generation even if the quality is poor. AI video models can produce artifacts, ghosting, or frozen frames. Try adjusting your prompt, using a different model, or lowering the duration.
I got "Content blocked by safety filter" but my prompt seems fine.
Safety filters are automated and sometimes overly strict. Try rephrasing your prompt, removing specific names or references, or using a different model. Each model has its own content rules.
A generation has been processing for a very long time. Should I cancel it?
Video generations can take several minutes. If a generation looks stuck for more than 10 minutes, it will usually be resolved later by the provider or by our background reconciliation process. Only workflow-based generations currently support real cancellation. Most direct-model generations cannot be stopped once they have been submitted to the provider, and removing an item from the board does not stop the provider-side job. If the run ultimately fails, it will not be charged.
Percentages above are based on recent production failures and are shown as a share of all failed generations. They do not add up to exactly 100% because smaller edge cases and user-cancelled runs are not broken out as their own sections here.