Stop putting your AI characters straight into scenes 🫵😠👎 Here's what you do: give them a screen test first! 🎬 And here's why this is useful: the model reads your character's face and builds their voice from it, even when you tell it the accent...but let's go in depth: 🗣️ THE MODEL CASTS THE VOICE: Seedance reads your character's appearance and builds how they sound from it. Age, build, skin tone, ethnicity, the wear on their clothes. Before you write a word of dialogue, as soon as Seedance reads your character sheet, it's already building the voice. The screen test is where you hear it, and where you lock it. Imagine you're the casting director and your character is auditioning. If you want more control over who gets the role, follow these steps: 🎥 THE SETUP: Here's how I do mine: plain backdrop. Character alone on a mark. One slow push-in from a wide to a tight close-up. A short monologue straight to camera about who they are. The push-in does double duty. The wide checks the costume. The close-up checks the face. If anything drifts from the sheet, you catch it early and adjust the reference sheet. That might mean adding close-up detail shots of their clothing to the sheet. 🔒 LOCK IT: Epoo's screen test runs on ONE reference: his character sheet. No voice file. No audio reference. The voice you hear came from how the model read him. Happy with the voice? Extract the audio and make it your audio reference in every clip for that character. ✂️ Trim the clip down to the strongest stretch of dialogue, then extract the audio. An AI agent can do both in seconds: hand it the video and ask for the trimmed audio. Keep it short enough for Seedance 2.0 or 2.5 to accept as an audio reference, especially if you're using multiple audio references. Skip this and one character can sound like three different people in the same film. BONUS: the screen test also works as a video reference for how your character moves and performs in later clips... 🌎 ACCENTS: Seedance is surprisingly good at accents, and a screen test is where it shines. For this test we used "British English". With a plain backdrop and nothing else in the frame, the model has nothing to juggle. No crowd, no set, no camera tricks. All of its attention goes to one person: the face, the voice, the timing, the accent. That's why the performance in a screen test often comes out better than in a busy scene. Strip the setting down and you give the model room to act. ❌Wrong voice? Rewrite the accent line and run the test again at 480p; no need to upscale, since you're only using it for audio and possibly a performance reference. It's one cheap clip versus a whole film of the wrong voice. Screen test first to 'audition' your characters. 'Cast' them for your full movie second. Full prompt below 👇
A ragged man speaks to camera in a dark void, holding an old map, beside a character reference sheet.
- 3.9K views
- 76 likes
- 7 reposts
- 11 replies
- 76 bookmarks
This video is God's Way Foundation’s work, shown from their public post. See it on X.
