Skip to main content

​
Use key terms, not conversational phrases

Jumper works best when you use direct, descriptive keywords rather than conversational language.

❌ Avoid

“show me all the frames with a man holding a skateboard on a ramp”

✅ Better

“man holding a skateboard on a ramp”
The model focuses on the visual content described in your query, so extra words like “show me all the frames with” don’t help and can sometimes reduce accuracy.

​
Keep searches focused

When searching for multiple unrelated items, perform separate searches rather than combining them with “or”.

❌ Avoid

“a polar bear or a red car”

✅ Better

Search 1: “polar bear”
Search 2: “red car”
Separate searches give you better results because the model can focus on one concept at a time. You can always add both sets of results to your timeline if needed.

​
Simplify searches

If a search doesn’t return the results you’re looking for, try simplifying it by removing specific actions or details.

If this doesn't work...

“a couple kissing on bench”

...try this instead

“a couple sitting on bench”
Sometimes removing specific actions (like “kissing”) and using more general terms (like “sitting”) can help the model find what you’re looking for. You can always refine your search after getting initial results.

​
Add shot and scene details

You can include cinematographic details and scene characteristics in your search queries to find specific types of shots or settings. The model will often recognize these cues and return more relevant results. Examples:
  • “red car close-up”
  • “red car wide shot”
  • “person medium shot”
  • “building drone shot”
  • “city street at night”
  • “beach sunrise”
  • “forest daytime”
  • “interior daylight”
  • “mountain landscape dusk”

​
Search for text in footage

Jumper’s models have OCR-like capabilities, meaning they can read text that appears in your footage. You can search for signs, labels, text overlays, or any written content visible in your video. For example, searching for “STOP” might find traffic signs, or searching for a product name might find shots where that text appears on screen. It might help to add “text” or “sign” to your search to help the model find what you’re looking for.

“Acmecorp text”

“Acmecorp sign”
There’s no guarantee that the model will find all occurrences of text in your footage.

​
Avoid negative searches

Avoid using negative terms like “not” in your queries. The model may not properly interpret the negation and could return the opposite of what you want.

❌ Avoid

“not a red car”

✅ Better

“blue car”

​
Narrow the search context

The models can struggle if you are searching in large amounts of media. If you experience degraded accuracy, you might want to use the context picker to narrow down the search context.
If you know what files or bins you want to search in, its recommended to search in that specific context instead of everything.

​
Temporarily hide results

You can shift-click a thumbnail in the search results to temporarily remove it from view. This is especially handy when you want to add all results to the timeline except for one or a few false positives.