Does color grading or blur need a YouTube AI label? Sume video filter
YouTube lists beauty filters, color and lighting adjustment, background blur and upscaling as minor edits needing no AI label. One $0.02 filter job does most.

No. YouTube's disclosure page says minor edits do not need an AI disclosure, and it defines minor edits as ones that are primarily aesthetic and do not alter the content in a way that could mislead the viewer about what actually happened. Its list includes beauty filters, color adjustment or lighting filters, special effects filters such as background blur or vintage effects, caption creation, and video sharpening, upscaling or repair.
Sume's video filter does this kind of pixel pass on one clip for $0.02 per encode job, and it returns a new MP4 without touching the source.
What the page exempts
Keep the intent test in view: aesthetic, not misleading. The table maps YouTube's minor-edit list to what a filter job can do.
| YouTube minor edit | Sume video filter route |
|---|---|
| Color adjustment or lighting filters | eq, colorbalance, hue, or the dim op |
| Background blur or vintage effects | gblur, vignette, noise in filtergraph |
| Sharpening | unsharp or cas in filtergraph |
| Beauty filters | No named beauty feature; use tone and blur filters in filtergraph |
| Caption creation | Video captions, $0.20 per job |
Check, then run
The filter API has a free check endpoint, POST /v1/video-filter/check, that validates the program and the source without creating a job. Then the encode call needs an Idempotency-Key. This program warms the picture slightly and adds a light vignette:
curl -X POST https://api.sume.com/v1/video-filter \
-H "Authorization: Bearer $SUME_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: grade-001" \
-d '{
"video_url": "https://media.sume.com/artifacts/artf_demo/talk.mp4",
"filtergraph": "eq=contrast=1.06:saturation=1.1,vignette=PI/10"
}'Limits that matter
- The source must be 300 seconds or shorter; longer sources fail with
output_duration_exceeded. - The filtergraph is capped at 2048 characters and 32 filters, and only allowlisted pixel filters are accepted.
- The default mode is async; poll
GET /v1/jobs/:id/resultor sendmode: syncto wait up to 30 seconds.
Why a separate pass is worth it
Because the filter job never rewrites the source, you can keep the untouched clip as your reference and publish the graded copy. If a viewer, a brand partner or a platform ever asks what was changed, the answer is a short, boring list: a tone curve, a vignette, a crop. That is the same shape of answer YouTube's page wants, where the edit is aesthetic and the facts on screen are unchanged.
Run the free check first when you are experimenting. An unknown filter name fails up front with the allowlist in the error, before any compute is reserved, so a typo costs nothing.
Caveat
A grade is only minor while it stays aesthetic. Relighting a scene so it shows a different time of day, or blurring a sign that changes what happened, can move you into 'alters footage of a real event' territory. If YouTube's systems still label the upload, you can usually change the answer under Attributes, except for content made with YouTube's tools, with C2PA metadata, or labeled after manual review.
Sources
Related posts
More in Media tools
- Douyin ad first frame: black at most 60%, check with stills
Douyin in-feed ads need a first frame that is at most 60% black. Pull stills from your render with Sume video inspect and measure the first frame before upload.
- Edits carousels: four matching images per call
Instagram Edits now supports carousels, per secondary reports. Sume image models with an n range of 1-4 return up to four matching images per request.
- Email-sized preview image from an avatar clip: video-frames max_edge
Pull the first frame of a Sume avatar clip as a 600-pixel jpeg for an email or a link card: one video-frames call with at [0] and max_edge 600.
- Faithful vs generative upscale: which to pick
The key question for any upscaler is whether it may invent detail. Compare Real-ESRGAN and SeedVR scale ranges, then run a test through Sume's upscale tools.
Written by Sume