Fix color banding in AI video: the deband filter API

Visible steps in skies and gradients can be softened with the deband filter in Sume's video filter, with noise as a fallback. It returns a new MP4. Check free.

4 min readSume
All posts

Put deband in the filtergraph of POST /v1/video-filter. It is on Sume's allowlist, and the encode returns a new MP4 at the listed $0.02 per job. If deband leaves visible steps, adding a little noise can hide them. Neither changes the color range; this is a cleanup, not an HDR conversion.

Runway's changelog (Ruby HDR conversion, Aug 20, 2026, read 2026-10-01) describes turning SDR video into HDR with updated brightness and color range. Banding in an SDR clip is a different problem, and the plain filter route is what this post covers.

Which filters help with banding?

Banding shows as visible steps in smooth gradients such as skies and walls. Two allowlisted filters address it, and format is on the list as well.

Banding-related filters on the video filter allowlist, from the code read 2026-10-01.
FilterRoleNote
debandSmooths steps in flat gradientsStart here
noiseAdds grain that hides stepsVisible grain is the cost
formatSets the pixel formatAllowlisted; not needed for most clips

How do I run it?

Check, then encode. The check is free, takes no Idempotency-Key, and does not boot a box.

curl -X POST 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_demo/sky.mp4",
    "filtergraph": "deband,noise=alls=2:allf=t"
  }'

What does Sume hand back?

A new MP4; the source is untouched. The docs do not describe a bit depth or HDR for this output, and the video trim docs name libx264 and yuv420p for exact re-encodes. Treat the result as a standard delivery file. For more on what Sume outputs, see HDR and EXR: what Sume outputs.

Will compression bring the banding back?

It can. Platforms re-encode uploads, and a clean gradient can pick up steps again. That is why the noise fallback is worth testing: a perfectly smooth ramp is the first thing compression flattens, and grain tends to hold. The source clip must be 300 seconds or shorter and on media.sume.com.

Sources

Related posts

More in Use cases

All Use cases posts

Written by Sume