AI Video Burns Garbled Subtitles Into the Frame: What Stops It and What It Costs
Credits and specs follow TWno1's live billing settings. Last checked: 2026-10-05
Put dialogue in a video prompt and the model may burn misspelled subtitles into the frame; in September 2026, 11 prompt variations failed to stop it for us.
This was not an occasional glitch. On one video model, all 8 shots that had dialogue came back with burned-in text. With another, an entire batch of 19 clips did. Below: why "no subtitles" in the prompt does not work, the check-and-fallback routine we used on our demo drama, what that routine costs, and the separate problem of garbled text on signs and props.
Three kinds of unwanted text
| Type | What you see | Where it comes from | Fix direction |
|---|---|---|---|
| Burned-in dialogue subtitles | A line of text at the bottom that looks like the script, with wrong or made-up characters | Dialogue text in the prompt | Detect and regenerate; if that fails, take the dialogue out of the prompt |
| Text on objects | Company names, signs, document titles drawn as nonsense glyphs | Image models write dense Chinese text poorly, and add text the script never asked for | Allow only text the script specifies; correct any misspelling in post |
| Logos or watermarks nobody asked for | A mark in a corner | BytePlus's guide says these can appear unprompted | Regenerate, or crop in post |
Each has a different fix. Treating them as one problem wastes renders.
What did not work
The trigger is simple: the dialogue text is in the prompt. Our guess is that the model sees a quoted line and treats it as text that should be visible. None of the vendor docs we read confirm that mechanism.
We tried 11 phrasings. Stronger "no subtitles" wording. The instruction in English. Not mentioning subtitles at all. None of them stopped it. We also asked for black bars at the top and bottom, hoping the text would land there instead of on the picture. It did not.
BytePlus, to its credit, says as much in its Seedance 2.0 prompt guide. The model cannot reliably prevent subtitles. Writing "no subtitles, no text" only lowers the odds. The same guide notes that horizontal output shows fewer subtitles than vertical.
That last detail matters for short drama, which is mostly 9:16 vertical: the orientation that gets more stray subtitles.
There is also a real trade-off here. If you want the model to generate spoken dialogue itself, the docs tell you to put the lines in the prompt. BytePlus's API page says dialogue in double quotes gives better audio; Google's Veo docs say to put speech in quotation marks. Native voice and burned-in text come from the same instruction.
The routine we used on our demo drama
- Sample 8 frames from every clip and run OCR (optical character recognition: software that reads text in an image).
- If it finds text that should not be there, regenerate with a new seed. Up to 2 times.
- Still burned in after two tries? Generate a version with no dialogue in the prompt. In our small test, 0 of 3 such clips had burned-in text.
- That version has no native speech, so the voice is added separately and lip-sync is applied afterward.
The worst case is 4 generations for one shot: 1 original, 2 retries, 1 dialogue-free version.
| Step | Generations so far | What you get |
|---|---|---|
| First render | 1 | Native dialogue, may have burned-in text |
| Two retries with new seeds | 3 | A chance at a clean take |
| Dialogue-free version | 4 | Clean in our 3-clip sample; voice and lip-sync done separately |
Our opinion: if an episode is wall-to-wall dialogue, skip straight to step 4 rather than gambling three renders per shot first. That only holds if your separate voice and lip-sync steps are dependable, and it is reasoning, not something we measured.
Text on signs, papers and name badges
The second kind of garbage text lives on objects. Company names come out as nonsense because image models handle dense Traditional Chinese poorly. Worse, models add text on their own, like a room number or a name badge the script never mentioned.
The rules we used on the demo drama:
- Text in a frame was allowed only if the script spelled out the exact words; anything else the model wrote was rejected.
- The shorter the text, the better. If it still came out garbled, the correct words were laid over it in post.
The point is to never let the model write freely. Text it invents is often garbled, and it may not match the story.
Vendors say their models can write. How far does that go?
Some vendors do claim text rendering. BytePlus says Seedance 2.0 can render common characters, slogans, subtitles and speech bubbles, and advises avoiding rare characters and symbols. Google says Omni Flash renders readable on-screen text when you spell out the exact wording. Those are vendor claims, not independent measurements.
| Source | Claim or finding | Kind |
|---|---|---|
| BytePlus Seedance 2.0 guide | Renders common characters, slogans, subtitles; avoid rare characters | Vendor claim |
| Google Omni Flash docs | Spell out the text and it will be readable | Vendor claim |
| Alibaba Cloud Wan 3.0 launch post | On-screen text rendering accuracy is still maturing | Vendor-stated limitation |
| VTR-Bench (Oct 2026 preprint) | 11 models, 300 prompts: widespread difficulty rendering scene text correctly | Research, not peer reviewed |
| VidScribe (Sep 2026 preprint) | 803 samples, 11 systems; image-to-video keeps text best, video-to-video text editing is the main bottleneck | Research, not peer reviewed |
| AVGen-Bench (Apr 2026 preprint) | Text rendering keeps failing even where visuals look good | Research, not peer reviewed |
How we read it: "can render text" describes the case where you ask for simple text on purpose. Whether the model adds text you did not ask for is a different question, and for drama it is the one that hurts.
Why subtitles belong in post
Even perfectly spelled burned-in subtitles fail a real spec. Netflix's English guide caps a line at 42 characters and adult reading speed at 20 characters per second; its Traditional Chinese guide caps a line at 16 characters, two lines, and 9 characters per second. The delivery bar we used on our demo drama adds: each subtitle on screen for at least 1 second, placed above the player's bottom controls, and in sync with the voice to within 100 ms per line.
A video model controls none of that. Post-production subtitles can be corrected, translated and moved. Burned-in text can only be re-rendered. TWno1's public spec says the same thing: text on signs or subtitles in generated video may warp, so add subtitles in post.
Limits: where this routine falls short
- Eight frames can miss things. A subtitle that flashes for half a second between two samples will slip through. That is an inference; we have not measured the miss rate.
- 0 of 3 is a tiny sample. The dialogue-free result could change with a different model or genre.
- The fallback loses native dialogue. Bad lip-sync can pull viewers out of the scene more than a misspelled line would.
- Not a good fit when the text is the point. If a close-up of a letter must be readable, overlay the words in post instead of hoping the model gets them right.
If subtitles go on in post, the footage has to leave the tool cleanly. TWno1 downloads finished shots as ordinary video files, so you can subtitle them in whatever editor you use, to the 42-character English line or the 16-character Chinese one. Fix any burned-in shot before you download. Once text is baked into the picture, no editor can lift it out.
Related reading: Why Your AI Video Camera Won't Stay Still, and How to Rewrite the Prompt; Per-Token vs Per-Second AI Video Pricing: How to Calculate Seedance, Omni Flash and Veo Costs; Shot Sizes in AI Short Drama: Rules From a 244-Frame Storyboard With 4 Wide Shots.
FAQ
Why does AI video burn garbled subtitles into the frame?
The most common trigger is dialogue text inside the video prompt. Once a line is in the prompt, it often shows up as visible text, frequently misspelled. We tried 11 prompt variations, including stronger wording, English instructions and not mentioning subtitles at all, and none of them stopped AI video from burning in subtitles.
Does writing "no subtitles" in an AI video prompt work?
It lowers the odds but does not prevent it. BytePlus's official Seedance 2.0 guide says the model cannot reliably prevent subtitles, that "no subtitles, no text" only reduces the chance, and that vertical video shows more stray subtitles than horizontal. Checked 2026-10-05.
What is the cheapest way to handle burned-in subtitles in AI video?
The workflow we used on our demo drama sampled 8 frames per clip and ran OCR. If it found burned-in text, we regenerated with a new seed up to 2 times, then fell back to a version with no dialogue in the prompt and added the voice and lip-sync separately. Worst case, one shot takes 4 generations.
How do you fix garbled text on signs and documents in AI video?
Allow only text the script spells out word for word, and keep it short. If it still renders as garbled glyphs, lay the correct words over it in post. On our demo drama we discarded any room numbers or name badges the model added by itself.
Should AI short drama subtitles be burned in or added in post?
Add them in post. Subtitles have specs a video model cannot control: Netflix allows 42 characters per line in English and 16 in Traditional Chinese, two lines maximum, and the text must stay in sync with speech. Burned-in text cannot be edited, only re-rendered.
References
- BytePlus ModelArk: Seedance 2.0 series prompt guide (checked 2026-10-05)
- BytePlus ModelArk: Create a video generation task (API) (checked 2026-10-05)
- Google Gemini API: Omni Flash (checked 2026-10-05)
- Alibaba Cloud: Wan 3.0 general availability post (checked 2026-10-05)
- VTR-Bench (arXiv 2610.01499, preprint) (checked 2026-10-05)
- VidScribe (arXiv 2609.36598, preprint) (checked 2026-10-05)
- AVGen-Bench (arXiv 2604.08540, preprint) (checked 2026-10-05)
- Netflix: English timed text style guide (checked 2026-10-05)
- Netflix: Traditional Chinese timed text style guide (checked 2026-10-05)