How we test ClipMend
Most repair tools advertise a "success rate" with no method behind it. We'd rather show our work. ClipMend is tested on footage from real cameras, through the real engine — not synthetic demos — and we measure recovery against the originals byte-for-byte.
What "verified" means here
- Pixel-identical rescue. On GoPro (HEVC), DJI (H.264) and multi-slice footage, recovered frames decode pixel-for-pixel identical to the source — in display order, with rotation preserved.
- Byte-exact B-frame order. A phone HEVC clip (HLG 10-bit, rotated) rebuilt to 10,792 / 10,792 frames with the elementary stream byte-identical to the original.
- Byte-exact audio. Recovered AAC packets compared against the source come back identical, packet for packet (e.g. 879/879 on a truncated GoPro clip), and PCM audio is recovered bit-exact.
- Color & HDR preserved. iPhone HLG stays 10-bit BT.2020; SDR tone-mapping lands on correct BT.709.
- Batch & merge. A card of broken segments rescues with an honest "N of M" result; compatible clips merge losslessly (frames 1:1, decode-clean).
The method
- Real source corpus. Footage from GoPro, DJI, Sony, Panasonic, iPhone, dashcams, OBS/screen recorders, plus RAW stills — the cameras people actually use.
- Break it the way it breaks in the wild. 19 corruption classes: truncation, wiped/shifted headers, missing index (moov), fragmented MP4 (moof), garbage tails, broken timestamps.
- Run the real engine, not a mock. Every function is exercised through the same code the app ships — measured with
ffprobeagainst the original (frame counts, fps, stream bytes, color tags). - Compare, don't assume. Recovered streams are diffed against the source. "Recovered" means the bytes match — not "it opened".
The limits we won't hide
Honesty is part of the product. Rescue only works when the frames are physically present in the file. Zero-filled, empty or header-only files can't be recovered by any tool — and ClipMend's free Analyze says so before you spend time on it. Some classes are still partial or unsupported (certain AAC chunk layouts fall back to video-only; DJI O3 is no-promise; photo repair currently covers JPEG). We list these rather than paper over them.
FAQ
Why not just publish a single "success rate" percentage?
Because it's meaningless without the test set. A number is easy to inflate by choosing easy files. We'd rather tell you exactly which cameras and corruption classes we verify, and where we fall short.
Can I verify a rescue myself?
Yes — compare the output's frame count and duration to the original with ffprobe, and play it. ClipMend also shows a thumbnail of the recovered frame right in the result, so you can see the picture came back.
See it on your own footage
ClipMend for macOS — free public beta. The Analyze step is free and shows what's recoverable before anything else.
⬇ Download ClipMend