An AI avatar can match every syllable and still feel wrong. This 94-second explainer looks beyond the mouth: expression, head movement and gestures.
Using labelled official Kandinsky S2V examples and original animated diagrams, I explain the photo + speech + prompt workflow, five-second chunks and visual memory—then show a simple two-pass viewing test.
00:00 Lip sync is only the beginning
00:15 Photo + speech + prompt
00:26 Four things to watch
00:39 Five-second chunks and visual memory
00:54 What selected demos can—and cannot—show
01:09 The two-pass test and takeaway
At 01:21, three seconds of silence are intentional. The same excerpt then repeats with its original audio at 01:24. Elsewhere, the English narration is commentary, not a dub of the example characters.
Sources and credits:
• Official demonstration excerpts: Kandinsky Lab
https://kandinskyavatar.github.io/
• Paper: Kandinsky S2V 1.0: Audio-Driven Avatar Video Generation
https://arxiv.org/abs/2610.19290
• Code and implementation details:
https://github.com/kandinskylab/kandinsky-s2v
This is an explanatory film, not an independent model benchmark. The examples were selected by the researchers; I did not run Kandinsky S2V for this video. Realistic example footage is AI-generated. Narration is synthetic. The diagrams, editing and quiet background score were created for this production. The conceptual cover artwork is AI-generated, not a Kandinsky result.
Produced with Framewright:
https://github.com/markoblogo/framewright
Which detail breaks the illusion first for you: lips, expression, head movement, or hands?
#AIAvatars #Kandinsky #GenerativeAI
YouTube / ABVX