Remove Burned-In Subtitles, Watermarks or Text from a Video

Erase Burned-In Text Free

Upload the footage, say what to remove, watch it disappear from the picture. No credit card required.

How to Remove Burned-In Text from a Video

1. 1
   Upload the footage

   Drag in any MP4, MOV, MKV or WebM up to 14 GB or 3 hours. Valmera runs a one-time analysis — word-level transcript, silence and shot detection, and labeled frame tiles the agent reads directly, so it can already see the text on screen before you describe it.
2. 2
   Ask where the text is (optional)

   Type "where is there text burned into this video?". You get back measured rectangles as fractions of the frame, each labeled captions, watermark or text, with the seconds it is visible and how many sampled frames it appeared in. This step is worth taking when there is more than one mark and you only want some of them gone.
3. 3
   Type the removal

   "remove the burned-in subtitles" clears every caption band in one pass. For one specific mark: "erase the watermark in the top-right corner", or "remove the username at the bottom left between 0:12 and 0:40". You can name a time window; regions apply to the whole video otherwise.
4. 4
   Check the measurement, then export

   The reply carries an ink reading for each rectangle before and after the repaint, and the agent looks at the frames it produced. If a region still measures dirty it widens it or repaints the whole rectangle instead of just the letter strokes, rather than reporting success. When the preview looks right, export an H.264 MP4.

Detection samples frames by seeking, so it is quick. The repaint itself decodes and re-encodes every frame of the full-resolution original — the one edit in Valmera that is bounded by video length rather than by your patience.

Blur, Cover, Repaint — Three Different Jobs

Most tools that advertise watermark removal do one of the first two. A blur keeps every pixel of the text in the frame and low-passes it, which is why the result is a conspicuous rectangle of mush that a viewer reads as "something was removed here". A solid box is more honest about itself but destroys the picture behind the mark as well. Both leave the video obviously altered in the exact spot you wanted nobody to look at.

Inpainting is the third option: work out what the picture behind the text was, and put that back. For thin, high-contrast ink over footage that keeps moving — which describes subtitles almost perfectly — this is not a hard problem, because the information is genuinely there. The words change every couple of seconds, so a pixel covered by the stroke of an E at 3.0s is clean background at 3.4s. The reconstruction is a photograph, not a hallucination.

Valmera ships all three. Blur, pixelate and black-box regions exist for censoring; the erase exists for removal. Ask for the outcome and the agent picks — say "hide his name badge" and you get a censor, say "get rid of the subtitles" and you get a repaint.

Finding the Text: Measured, Not Guessed

The first half of the job is knowing exactly where the text is, and getting it wrong is worse than doing nothing — a rectangle in the wrong place censors the picture and leaves the text. So Valmera measures it off the frames rather than asking a vision model to eyeball coordinates.

The scan samples 28 frames spread across the video (never the very first or last, which are the most likely to be a fade or a title card). In each one it runs a morphological stroke response — a top-hat and a black-hat, maximised together, so light text on dark footage and dark text on light footage both register, and white text with a black outline registers twice. Glyphs are merged into lines, and lines are filtered on geometry: a text line is between 1.2% and 28% of the frame height, at least 3.5% of its width, and wider than it is tall. That geometry is what stops foliage, brickwork and a crowd from voting as text. Pixels that fall inside a text line in enough of the samples become a region, padded outward because outlines, drop shadows and antialiasing live just outside the ink and a rectangle that clips them leaves a visible ghost.

Each region is then classified without asking a model anything: the crop is compared between samples. Content that changes is captions — wide, centred, new words every line. Content identical in every frame is a watermark — a handle or a logo. Everything else is text: a headline, a UI label, a lower-third. That is what lets "remove the subtitles but keep the channel logo" work as a sentence.

If the structural scan finds nothing and you insist there is something, the agent looks at four frames itself and names what it sees — and then the box it names is snapped to the measured ink inside it before anything is repainted. The model is allowed to point; only the pixels are allowed to measure. And if that finds nothing either, the agent says so and asks where the mark is instead of inventing a rectangle.

How the Repaint Works

The removal runs in two passes over the full-resolution original. The first builds the plate: sample frames across the region's time span, and for each pixel take the median of every sample in which that pixel was not covered by ink. The plate is only trusted where enough clean samples agree and the shot is steady enough for them to line up, measured as the mean absolute deviation between samples — a handheld pan gives a plate that would not register correctly, so it is not used there.

The second pass streams every frame of the video. Where this frame has ink, the plate is pasted in. Pixels the plate never saw fall back to boundary-inward reconstruction, which is excellent on thin strokes and imperfect on busy detail. Then two finishing touches that decide whether the result is invisible or merely text-free: the high-frequency noise of the untouched pixels in the same band is measured and matched into the fill, because a perfectly smooth patch reads as plastic against real sensor noise; and the mask is feathered, so no rectangle edge appears in the output.

Two details handle the awkward styles. The backing bar: news lower-thirds and one popular mobile-editor caption preset sit on a solid strip. Repainting only the letters would lift the text off the bar and leave a grey rectangle floating in the shot — which measures as "no ink" and would otherwise be reported as success. So the region is tested against its own surroundings for colour distance and flatness, and when it is a bar the rectangle grows to the bar's real extent, measured across the whole span and unioned, because a bar measured on the shortest caption leaves both ends of the longest one standing. Strokes versus whole rectangle: the default repaints only the letter strokes and keeps the picture between them, which is right for captions and handles. For a solid object, a sticker or a logo shape, the whole rectangle is repainted instead — a harder problem, because an object sits in every frame and no plate can be built for it. That case has its own page: removing an object or a person from a video.

The repainted file keeps the source's duration, frame rate and audio exactly, which is what makes it a drop-in replacement: your transcript, shot list, cuts, captions and every timestamp in the edit still point at the same moment. Erasing text at any point in a session does not invalidate the editing you already did.

The Part Nobody Else Does: Measuring the Result

After the repaint, Valmera measures the ink response inside each rectangle again — on the file that will actually be rendered — and compares it to the reading taken before. The agent sees something like ink 41 -> 2 — gone, or ink 41 -> 27 — STILL VISIBLE. On the second reading it is told to widen the rectangle or repaint the whole box instead of the strokes, and explicitly told not to claim the text was removed until it measures clean.

This matters more here than anywhere else in the editor. Inpainting is the one operation where a confident-sounding "done — the watermark is gone" is easy to write and expensive to trust, because you find out it was wrong at export. Valmera's replies are verified server-side against the edits actually recorded, so the agent cannot claim work it did not do; for the erase specifically, the verification goes down to the pixels. It also renders a preview and looks at the frames it produced, so if the repaint left a soft patch it is expected to tell you rather than let you discover it.

When It Goes Wrong, and What to Do

The general rule: if the repaint looks worse than the mark, put it back. Ask to restore the region and the source is re-derived from your untouched original without it, and then reach for a blur or a black box, or crop the band out of frame entirely — reframing a 16:9 video to 9:16 often loses a corner watermark for free.

How People Do This Without Valmera

After Effects Content-Aware Fill. The professional answer, and a good one: mask the object, track it, and After Effects generates replacement pixels using surrounding frames — the same underlying idea as the plate described above, with far more control over the mask and a reference-frame option for hard cases. The costs are a Creative Cloud subscription, a slow and RAM-hungry generate step, and a manual mask per shot.

DaVinci Resolve Studio. Patch Replacer samples a clean area of the frame and blends it over the mark; Fusion's Object Removal does neural reconstruction. Both sit on the paid Studio side rather than the free build, so check which edition you have before planning around them. Studio is a one-off licence rather than a subscription, and if you already own it this is a strong option for a single stubborn shot.

ffmpeg delogo / removelogo. Free, scriptable, one command, no install beyond ffmpeg itself. Be clear about what it does: delogo interpolates the box from the pixels immediately outside it — a controlled smear, not a reconstruction — and it has no idea where the text is, so you supply x, y, w and h yourself. removelogo takes a bitmap mask image marking the logo and fills those pixels in from their neighbours instead of a plain rectangle. On a small static station logo over a soft background this is genuinely fine and costs nothing; over detail, or over a subtitle band, it is visibly wrong.

Open-source subtitle removers. Projects built on video inpainting models can produce very good results on the hardsub case. The price is a Python environment, model weights, a GPU worth using, and drawing the box yourself. If you enjoy that, it is the best free result available.

Online watermark removers. A large category, and mostly blur or smart-looking smudge behind an AI label, with upload caps and — often — a watermark of their own on the output. Check the result at 100% before you trust the marketing.

Valmera's claim against that field is narrow and specific: it is the option where you never draw a box, never open a masking UI, and get told in numbers whether the text actually came out. It is not the option with the most control over a single hero shot — that is still After Effects.

The Line Worth Drawing: Your Text vs Someone Else's

Two requests that look identical to a tool are not the same act.

Removing burned-in captions from your own footage, taking a client's old lower-third out at their request, clearing a caption band so you can re-caption in a different font, stripping a handle from a video you own after rebranding — routine post-production, and the reason this feature exists at all.

Removing someone else's watermark so their footage can be reused is a different thing. The mark identifies the owner and exists precisely to prevent that reuse. Stock previews from libraries like Getty and Shutterstock are licensed as previews; stripping the mark breaks the licence, whatever the footage is then used for. And in the United States, removing copyright management information is prohibited under DMCA section 1202 as a separate wrong from infringement itself — so "my use was fair" is not automatically an answer to "you removed the notice".

Valmera does not inspect your footage and has no way to tell whose it is. The tool will run; the responsibility for what you run it on is yours. None of this is legal advice, and the rules differ by country — if a project turns on the question, ask a lawyer rather than a video editor.

Honest Limits

Length. The repaint decodes and re-encodes every frame at full resolution, so it is capped by total pixel work: about 10 minutes of source, and roughly 900 megapixel-seconds — near enough 7 minutes at 1080p and under 2 minutes at 4K. Above that the agent refuses honestly and offers what fits rather than starting something that cannot finish. Every other edit still works on the full 14 GB / 3-hour upload.

No motion tracking. An erase region is a fixed rectangle, optionally limited to a time window. A watermark that drifts around the frame needs a rectangle covering its whole route, which repaints much more picture than the mark occupies. Corner marks — the overwhelming majority — are the good case.

Large translucent overlays. A big diagonal watermark spread across the whole picture is not thin ink on a background; there is no clean background to recover, and repainting the frame is repainting the video. Inpainting does not save that footage, and no honest tool will tell you otherwise.

The export path changes slightly. Normally Valmera exports from your original file at source quality. Once anything is erased, the render reads a repainted full-resolution copy of that original instead — a high-quality H.264 re-encode. Your upload itself is never modified, every erase is re-derived from it rather than from a previous repaint, and undoing one restores the untouched pixels.

No subtitle extraction. Burned-in subtitles cannot be turned into an SRT here — Valmera has no SRT or VTT import or export at all. It can transcribe the speech itself and burn a fresh caption track once the old one is gone.

And our own mark. On a page about removing watermarks it would be odd not to say it: free-plan exports carry a small Valmera mark in the top-left corner, and every export closes with a brief end card after your video. Creator, Pro and Frontier exports carry no watermark at all, and upgrading re-renders an already-marked export clean. Previews are never marked on any plan.

What People Use It For

Part of the Same Chat

Erasing text is rarely the whole job. The same conversation can cut the video, add word-level captions, score it with music that ducks under speech, reframe it for another platform, and grade it — one request can carry several: "erase the old subtitles, caption it in Bebas Neue, and make me a 9:16 version". Browse everything in the tools hub, or read how censoring and effects work in the effects docs.

The same erase tools are exposed over Valmera's MCP server, so Claude can measure and repaint burned-in text from inside a conversation without you opening the studio at all.

Take the Text Out of the Picture

Measured, repainted, and checked against the pixels before anyone claims it worked.

Frequently Asked Questions

Remove Burned-In Text from Your Video

The agentic AI video editor — describe the edit, review the preview, download the export. 50 free credits, no card.