People ask for a number: how many faces can you blur at once? Product pages like round caps because servers bill per unit. A client-side tool does not sell you “up to 40 faces.” What actually happens is a detector proposes boxes, you accept or fix them, and you download a raster with those regions covered. The detector’s accuracy falls as faces get smaller, more numerous, more occluded, and less frontal. The session’s performance falls as the bitmap gets huge. Neither of those is a secret “max 25 faces” policy. Both are reasons a group photo is a slower, more manual job than a two-person snapshot.
HideShot can auto-suggest faces and always lets you draw Blur, Pixelate, or Black Box by hand. Crowd-oriented covering is pixelate a face in a crowd. A general people-blur walkthrough is blur people in a photo. This page is the capacity and reliability question: why dense photos miss, why that is not a quota, and how to finish a class photo or an event frame without pretending the detector saw everyone.
Guide
Face detection in the browser is typically a lightweight model (the family that includes Haar cascades in older demos and SSD / BlazeFace-class networks in newer ones). Those models score windows of the image for “this looks like a face.” They are tuned on faces that occupy a reasonable fraction of the frame, often near-frontal, reasonably lit. They are not a census. When a wedding dance floor or a graduation bleacher packs fifty heads into a few megapixels, many heads occupy too few pixels, sit at profile, or sit behind another shoulder. Confidence drops. The UI shows fewer boxes than people you can name. That gap is the whole topic.
What actually limits a local session
Miss rate, not a counter. There is no HideShot ledger that stops you at face 30. You can keep drawing. The tool will not refuse a 41st oval because of a license tier. If a detector run returns twelve boxes on a photo with forty recognizable faces, the other twenty-eight are your job with rectangle, oval, or lasso.
Face size in pixels. A detector that is reliable on a 200-pixel-wide face will be unreliable on a 24-pixel-wide face in the same JPEG. Zoom the image. If you can still recognize a person at 100% zoom, they are in scope even if the model skipped them. Distant blobs that are already unreadable as faces do not need boxes that wreck the skyline — but “unreadable” must be true at 100%, not at a tiny preview.
Pose, occlusion, and lighting. Hats, hands, microphones, backlight, and side views are standard miss conditions. Event photos are made of those conditions. Class photos with everyone facing the camera are easier and still miss the kid who turned. Sports sideline shots miss helmets. Do not use a successful selfie detect as evidence the model will finish a crowd.
Device memory and image size. A 40-megapixel group photo in a phone browser is a heavy bitmap. The tab may stutter. That is RAM and CPU, not a face quota. If the page struggles, work on a reasonably sized export (still large enough that small faces remain coverable), or mark in two passes on crops, then confirm the full frame.
Why manual mark is part of the design
Auto-detect is a time-saver on simple frames. It is a first pass on hard frames. Privacy claims are about the export, not about the model’s recall. If you publish a photo where the detector got the front row and missed the child at the edge, you published the child. Click-to-cover exists so the claim can still be true. Oval follows a head better than a huge rectangle that also smears a chalkboard. Lasso follows a hairline. After manual adds, download and zoom — detection boxes can also be slightly small, leaving an ear or a jaw.
For a dense, faceless mass where the job is a pattern of heads rather than a roster, pixelate a face in a crowd matches that visual. For covering people without a crowd-specific brief, blur people in a photo is the general pass. Neither page invents a maximum headcount. Use this page when the question is “will auto-detect finish a class photo?” The honest answer is: probably not completely; plan to walk the rows.
Practical sequence for a large-group still
- Open the JPEG or PNG you will actually share — after crop. Covering a wide original and then posting a tighter crop of an unmarked duplicate wastes the work.
- Run detect if the tool offers it. Accept useful boxes. Do not assume completeness.
- Zoom to 100%. Scan left to right, front row to back, then the edges and reflections (glass, instruments, trophies). Mark every recognizable face that should not travel with the post.
- Use Pixelate or Blur on distant heads so the scene still reads; use stronger cover on close faces. Black box is fine when identity matters more than aesthetics.
- Download. Zoom the PNG. If a face is still a person you could describe, enlarge that region. If the detector later offers a second pass, still re-walk; models do not get a second opinion unless you look.
Yearbook and class-style frames are a useful mental test: everyone is meant to be identifiable. If your goal is to share the banner and the trophy but not a roster of minors, you will mark a lot of heads. That is tedious and correct. A concert pit where faces are already a texture can take a lighter pass on the unreadable back. Decide per face at 100%, not per genre slogan.
Mistakes that treat detect as a census
Posting because the tool drew a lot of boxes. Lots is not all. Walk the back row.
Assuming a number like “up to 50 faces” exists and was hidden. The limit is misses plus your time. Inventing a cap would be a marketing lie.
Covering only faces the model found and leaving a sharp profile at the edge. Profiles are people.
Working only on a phone-sized preview of a 6000-pixel-wide team photo. Small faces exist only when you zoom. Zoom.
Marking one frame of a burst and posting the unmarked near-duplicates. Each file is a new set of faces. Detection will not follow the burst for you.