AI avatar video with a transparent background: Azure webm vs Sume
Azure's avatar batch API can output VP9 webm with an alpha channel. Sume's avatar video returns an MP4 with a scene behind the avatar.
Azure's text to speech avatar can output a transparent background: batch synthesis in webm can use the VP9 codec, and Microsoft says only vp9 can contain an alpha channel. Sume's avatar video does not advertise a transparent output; it returns an MP4 and puts the avatar in a scene you direct with a prompt or a photo.
If you need a cut-out presenter to place over your own page, that difference decides the vendor.
How does Azure do transparency?
According to Microsoft's overview, batch synthesis can use H264, HEVC or AV1 when the format is mp4, and VP9 or AV1 when it is webm. Only vp9 can contain an alpha channel. The real-time codec is H264, so the alpha option is a batch feature in this table of video avatars.
Browser support for alpha video is your problem to check; the Microsoft page covers the encoder, not the player.
What does Sume return?
The talking-video job returns public media.sume.com video artifacts, and Sume's docs describe a final MP4. The background is part of the request: scene can be a prompt or a photo, and each video_inputs scene can have a background prompt or image. Current execution expects scene backgrounds to resolve to one shared scene.
So the Sume way to get a clean backdrop is to ask for one, for example a flat studio color, not to expect an alpha channel.
| Question | Azure avatar batch | Sume avatar video |
|---|---|---|
| Container options | mp4 or webm | MP4 |
| Alpha channel | Yes, only with vp9 | Not documented |
| Background control | Background color and other properties | scene prompt or photo; per-scene background |
| Real-time output | H264 only | Not applicable; job-based |
Can I cut out the background myself on Sume?
Sume documents media tools that work on a hosted clip: Video filter applies an allowlisted pixel graph and returns a new MP4, and Timeline compose places a still and a video in one frame. Neither is described as an alpha matte, so do not count on a transparent cut-out from them. Test with a sample before building on it.
What backdrop should I request?
Ask for a plain, evenly lit scene in the prompt, such as a solid-color studio wall, and review the preview stills before rendering. Sume's avatar video previews show the first frame without spending a full render, and changing quality afterward does not need a new preview.
If true alpha is a hard requirement, pick the vendor whose docs promise it, and keep Sume for clips with a designed scene.
How do I check whether an output really has alpha?
Do not trust the file extension. Download a short sample and inspect the stream: a webm file with VP9 and an alpha channel reports it in its pixel format, and a quick test is to play it over a colored page in your target browser. An MP4 never has a transparent region, so a clean flat background that you key out later is the only workable route there.
If you build the overlay yourself, keep the avatar clip's background a single flat color with no shadow, and record that choice in the request prompt, so every clip in a series matches.
Sources
Related posts
More in Sume Avatar 1.0
- AI spokesperson: build the avatar from a prompt, profile or photo?
Sume makes an avatar three ways: prompt, profile traits or a photo. Which one suits a brand spokesperson, what each costs, and the rights each one raises.
- AI spokesperson video price per second: standard, plus, max
Sume Avatar 1.0 talking video bills per second: $0.184 standard, $0.245 plus, $0.55 max, with a product image adding a little. A 4-second floor example.
- Avatar 1.0 API routes: canonical paths vs legacy aliases
New Avatar 1.0 code should call /v1/avatar-1.0/generate and /v1/avatar-1.0/talking-video. Older model-run and legacy aliases still work with the same body.
- Avatar 1.0 quality: standard for drafts, max for the final weekly cut
Sume Avatar 1.0 takes quality standard, plus or max. plus is the default. Use standard to check a script, then re-render the keeper at max.
Written by Sume