Full 90-second dialogue scene between two original characters, consistent faces throughout, zero real actors [YouTube]

by @hiroshinguyen_vfx in c/character-animation · · Full Scene

Full 90-second dialogue scene between two original characters, consistent faces throughout, zero real actors [YouTube]
Two OCs, a tense kitchen conversation, 14 shots, both characters hold across all of them. LoRA+InstantID per character, Act-One for the performances, 3 weeks. The over-the-shoulder shots were the consistency nightmare. Challenge entry.
youtube.com

Two OCs, a tense kitchen conversation, 14 shots, both characters hold across all of them. LoRA+InstantID per character, Act-One for the performances, 3 weeks. The over-the-shoulder shots were the consistency nightmare. Challenge entry.

1,177 points · 8 comments

Comments

  • @miaparkcreative · · 99 points

    the PERFORMANCES are what elevate this above a tech demo. the beat where character B looks away before answering — that's acting, that's Act-One carrying a real human choice through to an AI face. the scene has subtext. most AI dialogue has none.

  • @averytorrespro27 · · 71 points

    subtext is the final frontier. anyone can make two faces say words now. two faces having a conversation where what they DON'T say matters — that's filmmaking, and it's rare in this medium. you did it.

  • @mialee3930 · · 66 points

    the subtext is what separates this from a tech demo. the character is THINKING, not just talking. rare and hard.

  • @sophiakim_art9391 · · 97 points

    a 90-second TWO-character scene with both holding across 14 shots is genuinely state-of-the-art for an individual creator right now. two characters doubles every consistency problem AND adds the two-shot framing where they both have to be on-model in the same frame. this is a real achievement.

  • @soodavislab · · 100 points

    the over-the-shoulder shots being the nightmare makes total sense — now you need character A's back/hair consistent AND character B's face consistent in the SAME shot, and the OTS framing is exactly where identity methods are weakest (partial face, back of head). how'd you solve it?