avgblur, boxblur, gblur or smartblur: which blur on Sume?

Sume's video filter allows four blur filters: gblur, boxblur, avgblur and smartblur. For a general soft blur, start with gblur, the Gaussian blur controlled by a sigma value. boxblur and avgblur average pixels inside a window, which gives a flatter softness, and smartblur softens while trying to keep edges. All four go in the filtergraph field, cost $0.02 per encode and can be checked first for free.

How do the four differ?

FFmpeg describes boxblur as making each pixel the average of the pixels in a box around it, and avgblur as an average blur whose window is set by sizeX and sizeY. Which looks best is a taste call, so compare on a frame rather than trusting the names.

Which should I choose for a caption background?

Use gblur with a moderate sigma and pair it with the dim op so white text stays readable. A box or average blur works too, but can look blockier around bright highlights. Run each on the same five-second clip and pull a frame from each before you commit a full batch.

avgblur is the one to reach for when you want the blur to be directional, for example a wide horizontal window with a small vertical one.

How do I send one?

Write the filter with its options as name=key=value and no spaces. Send it to the free check, then the encode. The server wraps the graph as input and output, so you never write [0:v].

Blurring only part of a frame needs a split and overlay graph, which fits inside the allowlist but is easier to review with the check endpoint first.

curl -s https://api.sume.com/v1/video-filter/check \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "video_url": "https://media.sume.com/artifacts/artf_EXAMPLE.mp4",
    "filtergraph": "avgblur=sizeX=20:sizeY=2"
  }'

What does it cost and what are the caps?

The check is free. The encode is $0.02 flat for a source of 300 seconds or less, and the output keeps the source geometry, frame rate and audio. A graph may hold 32 filters and 2,048 characters. Unknown filters are refused with invalid_filtergraph, which lists the allowlist.

Because each encode returns a new MP4 and never overwrites the source, you can try all four filters on one clip and keep the best. A practical routine is to check four graphs for free, encode the two you like for $0.02 each, pull one frame from each, and then run the winner across the rest of the batch. Keep the idempotency key stable per clip and filter pair so a retry cannot queue a duplicate paid encode.