Adrian photo avatar: White and Transparent background settings have no effect in Foundry Playground

The grey/beige area immediately around Adrian is part of the 512×512 source portrait, so the Playground’s white or transparent setting changes only the video canvas behind that portrait; Microsoft does not document a GUI or API option that segments Adrian from, or replaces the background within, his source photo.

Adrian is specifically listed as a standard vasa-1 photo avatar, and photo-avatar output is 512×512—matching the downloaded file you observed. ^1^ ^2^

Playground behavior

The documented Playground workflow allows selecting a background, but it does not state that this performs foreground extraction or removes pixels from a photo avatar’s original image. ^3^

Consequently:

• White: changes transparent or surrounding canvas pixels to white; it does not recolor Adrian’s beige/grey portrait background.
• Transparent: can make the area outside the square portrait transparent, but the portrait itself remains opaque.
• 512×512 WebM: is the documented resolution for photo avatars, not evidence that the side areas from the wider preview should be included in the download.
• The behavior described is therefore consistent with the documented photo-avatar model rather than evidence that Chrome is incorrectly displaying the alpha channel.

There is no documented GUI setting that removes Adrian’s portrait background. To obtain Adrian isolated from that background, post-processing/subject masking would be required, or a suitable custom photo avatar would need to be created from an appropriately prepared source image.

API transparency settings

The batch API can request transparency for the output canvas, but it does not document subject segmentation for Adrian. Transparent batch output requires all three of these values:

• videoFormat: webm
• videoCodec: vp9
• backgroundColor: #00000000 or transparent

Only VP9 can carry an alpha channel in this avatar output. ^4^

For Adrian, the relevant avatarConfig is:

{
  "avatarConfig": {
    "talkingAvatarCharacter": "adrian",
    "photoAvatarBaseModel": "vasa-1",
    "videoFormat": "webm",
    "videoCodec": "vp9",
    "backgroundColor": "#00000000",
    "subtitleType": "none"
  }
}

As a standard photo avatar:

• Include photoAvatarBaseModel: "vasa-1".
• Do not set customized: true.
• Do not supply talkingAvatarStyle; the batch documentation says that property is omitted for a standard photo avatar. ^4^

The request is submitted through the asynchronous batch-synthesis REST operation, then polled until its status is Succeeded; the output URL is returned in outputs.result. ^5^

Even with those settings, expect alpha only where the generated canvas is transparent—not removal of the opaque beige/grey pixels embedded in Adrian’s portrait.

If the white setting also fails outside the portrait

That would be a different symptom. Generate a clip with a visibly contrasting color and inspect pixels outside the 512×512 portrait area. If that selected color is not applied there, report it through Azure support with:

• Speech resource region and SKU
• Foundry portal version
• Avatar and base model (Adrian, vasa-1)
• Selected background value
• Job timestamp
• Downloaded output and screenshot of the preview

Do not include the resource key or the private text used for synthesis.

References

1. Supported standard text to speech avatar
2. What is Text to speech avatar?
3. Text to speech avatar quickstart - Speech service - Foundry Tools
4. Batch synthesis properties for text to speech avatar
5. Use batch synthesis for text to speech avatar (programming-language-rest)