How Candidate Grouping Works Without an Absolute Threshold
Continuity-photo grouping follows two paths in a fixed order. Photos that already have a human-assigned element label are grouped by that label first. Only completely unlabeled photos enter the image-feature candidate path. That second path does not decide “this is the same prop.” It looks for the closer whole-image feature match in later takes. There is no absolute similarity threshold in this matching step.
Candidate grouping answers “which images should I review together?” It does not answer “has the subject already been confirmed as the same?”
Human labels take priority over visual similarity
Once a photo is linked to an element name, it does not compete in the unlabeled similarity path. The element label becomes the grouping key, and photos with that label enter the same tagged group. Other asset categories are not used as the grouping rule.
This keeps two kinds of evidence separate. A human label is an explicit structured record. Image-feature distance is only a candidate-ordering mechanism. They should not be described as equivalent automated judgments.
Tagged-group order also has a UI boundary. The current collection traversal does not guarantee stable alphabetical ordering by element name. Recalculation can change the order in which groups appear, so screen position should not be read as element priority.
Unlabeled photos are compared by whole-image feature distance
For unlabeled photos, SlateX extracts a whole-image feature representation. Grouping begins from images in the lowest-numbered take and uses those images as seeds. For each later take, the routine looks among unused candidates and chooses the one with the smaller feature distance.
The word “smaller” is relative. The routine finds the closer available candidate, but it does not then compare that distance with a fixed acceptance constant. If enough images are collected to form a group, the candidate group can be created.
That means the software is not saying “these images passed a similarity standard.” It is saying “among the candidates considered here, this one is closer to the seed.” If every image in the set is a poor match, the relatively closer image can still be chosen.
No threshold means the group label carries no verdict
A direct consequence is that images of different real-world subjects can still land in one candidate group. The generated group identifier only gives the candidates a place to be reviewed together; it does not certify the content.
The starting point is constrained as well. Images from the lowest-numbered take act as seeds. Images from later takes do not independently start the same kind of new seed process. That affects which images can become the basis of a candidate group.
If an image does not produce a usable feature representation, or a distance operation does not yield a usable result, it can also be skipped. “Not grouped” therefore does not prove “not similar.”
Relatively closer is not the same as passing a threshold. Without a threshold, grouping itself cannot carry the continuity decision.
Whole-image features do not mean “this prop”
The current similarity path works on the whole image. It does not perform object detection and it does not let the routine compare a user-selected region as the grouping subject. Background, framing, people, and overall composition can all influence feature distance.
So candidate grouping should not be described as finding the same prop or understanding a scene change. Two images with similar composition but different target objects can still be close in whole-image feature space. The reverse is also possible: the same object photographed with large framing or background changes should not be assumed to group together.
Image alignment does not change that boundary. Alignment can transform image geometry when the result passes its own checks, but it does not add object semantics or erase camera-position differences into a verified common viewpoint.
The final action is human visual review
The purpose of a candidate group is to put images that deserve comparison into one review context. The reviewer can switch between images in the group or use the flash-view behavior to look for changes in costume, props, position, action, or another continuity detail.
The useful action is not “check whether the software was right.” Instead, move through the candidate images and inspect the specific continuity detail that matters. If needed, correct the organization with an explicit human element label.
The software provides candidates and a review order. The continuity judgment remains with the person reviewing the images.
FAQ
Do photos with an element label still use the visual-similarity path?
Not in the same unlabeled candidate path. Once an element name has been explicitly attached, that label takes priority as the grouping key.
Do candidate images have to pass a similarity score before they form a group?
No absolute similarity threshold is used in this step. The routine chooses the relatively closer candidate from later takes, and a group can form once the at-least-two-images condition is met.
Does a candidate group mean the photos contain the same prop?
No. The comparison uses whole-image features without object detection or region semantics. A group only says which images were placed together for review.
What should the reviewer do with a candidate group?
Switch or flash through the images and inspect the specific continuity detail that matters, then make the judgment manually. The group identifier itself is not the decision.
Axiom One LLC — SlateX. Figures current as of 24 September 2026.