When Seedance refuses a prompt, it is almost always the model's content check, not a bug. On SeedVideo AI, 1,226 of 11,168 text-to-video and image-to-video jobs failed between September 15 and October 14, 2026, and 1,056 of those failures (86%) were content-check refusals. The reason named most often was copyrighted or trademarked material, and it was usually spotted in the generated video rather than in the prompt: brands, logos and well-known characters. Image jobs were refused far more often than text jobs, 13.4% against 4.5%. Real people were rarely the cause, at 19 of 6,214 image jobs. Describe an original subject instead of a named brand or character. A refused job is not charged.
The numbers
| Text-to-video and image-to-video, Sept 15 – Oct 14, 2026 | Count |
|---|---|
| Jobs that finished or failed | 11,168 |
| Failed | 1,226 (11.0%) |
| Failed because the content check refused them | 1,056 (86% of failures, 9.5% of all jobs) |
| Text-to-video jobs refused | 226 of 4,989 (4.5%) |
| Image-to-video jobs refused | 830 of 6,179 (13.4%) |
| Image jobs refused with a message that names real people | 19 of 6,214 submitted (0.3%) |
The other 170 failures were not about content: provider timeouts, images below the minimum size, source files the provider could not download, and similar technical errors.
Why Seedance refused: the reasons it gave
Since September 29, 2026, the model provider has returned a reason with each refusal, and the site shows it. Of the 573 refusals from then on that came with a reason:
| Reason the model gave | Refusals | Share |
|---|---|---|
| The generated video may involve copyrighted or trademarked content | 277 | 48% |
| The input image was flagged by content moderation | 125 | 22% |
| The generated video was blocked by content moderation | 72 | 13% |
| The prompt was flagged by content moderation | 38 | 7% |
| The input may contain copyrighted or trademarked content (logos, brands, characters) | 35 | 6% |
| The input contains a real person | 17 | 3% |
| The generated audio may involve copyrighted music, or was blocked | 6 | 1% |
| A general safety-filter message | 3 | under 1% |
Two patterns stand out.
Copyright is the biggest single cause. Counting the video, the input and the audio together, 317 of the 573 refusals (55%) were about copyrighted or trademarked material.
Most refusals happen after the prompt was accepted. 355 of the 573 (62%) were raised on the generated video or its audio, not on what you typed or uploaded. That is why a prompt that never names a brand can still be refused: if the clip the model produces looks like a famous character, logo or film scene, the output check stops it. The fix is to steer the model towards something original, not to find other words for the same thing.
Text prompts against images
Text-only jobs were refused 4.5% of the time; jobs that started from an image, 13.4%. The image adds a second thing to check, and the reasons differ:
| Reason | Text to video (127 with a reason) | Image to video (446 with a reason) |
|---|---|---|
| Generated video looks copyrighted or trademarked | 87 (69%) | 190 (43%) |
| Input image flagged | — | 125 (28%) |
| Generated video blocked | 18 (14%) | 54 (12%) |
| Input looks copyrighted (logo, brand, character) | — | 35 (8%) |
| Prompt flagged | 21 (17%) | 17 (4%) |
| Real person in the input | — | 17 (4%) |
| Audio, or a general safety message | 1 (1%) | 8 (2%) |
Which models refuse most
Refusals as a share of jobs submitted. Models with fewer than 100 jobs are small samples.
| Model | Image to video refused | Text to video refused |
|---|---|---|
| Seedance 2.5 | 203 of 1,372 (14.8%) | 4 of 89 (4.5%) |
| Seedance 2.0 Mini | 611 of 4,388 (13.9%) | 220 of 4,725 (4.7%) |
| Seedance 2.0 Fast | 5 of 71 (7.0%) | — |
| Seedance 2.0 | 6 of 100 (6.0%) | 1 of 39 (2.6%) |
| Seedance 1.5 Pro | 3 of 81 (3.7%) | 1 of 67 (1.5%) |
| Wan 3.0 | 1 of 73 (1.4%) | — |
| Kling 3.0 | 0 of 63 (0%) | 0 of 41 (0%) |
| Seedance 1.0 Pro Fast | 0 of 45 (0%) | 0 of 59 (0%) |
Seedance 2.5 and Seedance 2.0 Mini carry almost all the volume, and both refused about one image job in seven. If a picture keeps being refused on Seedance, Kling 3.0 in the same model picker refused none of its 63 image jobs over the same month.
How to rewrite a refused prompt
The rule behind every rewrite: describe an original subject, setting and motion, and drop names you do not own. Asking for a lookalike of a famous character or product usually fails at the output check anyway, because the output check judges the clip, not your wording.
These are illustrative rewrites written for this guide, not prompts taken from users, and the content check can still refuse any prompt.
| Refused pattern | Rewrite |
|---|---|
| A named cartoon or game character: "Pikachu running through a field of sunflowers" | An original character: "A small fluffy orange fox cub with oversized ears runs through a field of sunflowers, low tracking shot, warm afternoon light" |
| A brand in a product shot: "A Coca-Cola can on ice with the logo turning to camera" | A plain product with the look you want: "A red aluminium soda can with a plain label, beaded with condensation, turning slowly on crushed ice, macro lens, studio backlight". If it is your own brand, add the logo in your editor afterwards. |
| A famous film scene: "The Matrix bullet-time shot of Neo dodging bullets" | The camera technique with your own subject: "A dancer freezes mid-leap as the camera orbits 360 degrees around her, water droplets hanging in the air, rooftop at dusk" |
| A named public figure: "[celebrity name] giving a speech on stage" | A fictional person: "A fictional keynote speaker in a black T-shirt walks across a dark stage under a single spotlight" |
| Graphic wording: "A zombie tears into a victim, gore everywhere" | Suggest instead of show: "A shadowy figure lurches out of the fog in an abandoned hospital corridor, lights flickering, the camera backing away" |
| A product photo with someone else's logo | Crop the logo out, turn the product so the logo is small, or use your own artwork as the first frame. |
If it is still refused
- Change one thing at a time. Swap the subject, or the image, or one phrase, then run it again, so you learn which part tripped the check.
- Try relaxed content mode. On Seedance 2.5 and the Seedance 2.0 models, the composer's Advanced settings include a relaxed mode that reduces some false blocks. It costs 10% more credits, still enforces the usage policy, and does not guarantee success. Over the same month, 276 of the 384 jobs run in relaxed mode finished (72%). When people switched to relaxed mode within 30 minutes of a standard-mode refusal, 20 of 49 of those next jobs finished; some of them may also have changed the prompt.
- Try another model. Kling 3.0 refused none of its 104 text and image jobs in the period, though it was used far less.
- For a refused photo, change the picture, not just the words. Use a different shot, or a stylized version that turns the subject into a clearly fictional, illustrated character while keeping the pose and framing.
- Nothing is charged for a refusal. A job that ends in a refusal returns its credits, so trying again costs only the new job.
What does not usually get refused
Faces and ordinary photos of people. Only 19 of 6,214 image jobs (0.3%) were refused with a message that names real people, and 17 of those said so explicitly. Portraits, family photos and selfies as the first frame are supported. Naming a celebrity or other public figure in the prompt is a different matter, so describe a fictional person instead.
Ordinary action and drama. Fights, chases and tense scenes run when they are not graphic. The prompt check flagged only 38 of 573 refusals (7%).
How we counted
- Source. SeedVideo AI's own production job log, read only.
- Window. Jobs submitted from September 15 to October 14, 2026 (UTC).
- Jobs counted. Every text-to-video and image-to-video job that finished or failed, including draft previews and Premium unlimited-mode jobs. Reference-video and video-editing jobs are left out. Per-model rates use all jobs submitted, including the few still running when the data was read.
- Refusal. A failed job whose error code is the provider's content-policy code. The reason breakdown covers only refusals from September 29, when the provider began returning a reason; the 483 earlier refusals said only that generation failed.
- Refresh. The counts are rerun monthly with the same query.
FAQ
Why does Seedance refuse my prompt?
Almost always because of the model's content check rather than a fault. On SeedVideo AI, 86% of failed jobs between September 15 and October 14, 2026 were refusals, and the most common reason was copyrighted or trademarked material, such as a brand, a logo or a well-known character. Describe an original subject instead and run it again. A refused job is not charged.
Why was my image to video generation refused?
Image jobs were refused 13.4% of the time, against 4.5% for text. The most common reasons were a generated clip that looked copyrighted (43%) and an input image flagged by moderation (28%). Crop out third-party logos, use your own artwork, or try Kling 3.0, which refused none of its 63 image jobs in the period.
My prompt doesn't name a brand. Why does it say the video may involve copyrighted content?
Because the check also runs on the result. 62% of refusals with a reason were raised on the generated video or audio, after the prompt and image had passed. If the clip resembles a famous character, logo or film scene, it is stopped. Steer towards an original subject rather than a lookalike.
Does Seedance block faces or photos of real people?
Rarely. 19 of 6,214 image jobs (0.3%) were refused with a message that names real people. Portraits and family photos as the first frame are supported. Naming a celebrity in the prompt can be refused, so describe a fictional person instead.
Am I charged when Seedance refuses a generation?
No. A refused or failed job returns its credits on SeedVideo AI, so only the job you run next is charged.
What is relaxed content mode?
An option in the composer's Advanced settings for Seedance 2.5 and the Seedance 2.0 models that reduces some false blocks. It costs 10% more credits, still enforces the usage policy, and does not guarantee success: 72% of relaxed-mode jobs finished over the month measured.
Which video model refuses the least?
Kling 3.0 refused none of its 104 jobs between September 15 and October 14, 2026. Among Seedance models, Seedance 1.5 Pro refused 3.7% of image jobs, against about 14% for Seedance 2.5 and Seedance 2.0 Mini.
Try your rewrite
Run the rewritten prompt in the text to video generator, or start from your own picture in the image to video generator. To know how long the retry will take, see how long Seedance 2.5 takes, measured on 724 jobs.
