Keep your AI character recognisable from scene to scene in Google Flow.

Keep your AI character recognisable from scene to scene in Google Flow. 96 clips tested: see the reference sheet, saved Character and face-visible start frame that helped. You'll see the whole setup in a real session, then a twelve-scene test run three times. All prompts are on screen word for word. What you'll learn • Why words alone fail: four takes, four different women • Step 1: a three-panel reference sheet made with Nano Banana in Flow (0 credits) • Step 2: save her as a Character with a voice, then just type @Alex • Step 3: start every shot with the face visible and well lit • The twelve-scene test: run 1 turned the T-shirt yellow (16 of 24 takes), run 2 fixed the shirt but the face drifted, run 3 held both • Voice in English and Danish, two characters in one shot, an outfit change, slow vs fast camera, and five tips Alex's verdict (her editorial call against a crew filming a real presenter on location) • Words only, no reference: Not yet (four takes, four different women) • The twelve-scene test: Not yet (at least one take clearly her in 11 of 12 scenes; the office scene missed at 64 and 65) • Same voice in English and Danish: As good • Two characters in one shot: As good (both lines word for word) • A new outfit: Not yet (face match 48) • Fast camera: Not yet (face match fell to 51 in the whip pan; the slow dolly was fine) Chapters 0:00 We kill jobs 0:09 Twelve worlds, one character 0:26 Alex introduces the video 0:41 Why words alone fail 1:32 Step one: build a reference sheet 2:27 Step two: save her as a Character 2:59 Step three: the start frame 3:52 The twelve-scene test 5:56 Does the voice hold? 6:28 Two characters, and a new outfit 7:01 Slow camera, fast camera 7:29 Five tips 8:04 Alex sums up How we measured A face-recognition model (OpenCV YuNet + SFace) compares every clip with Alex's reference portrait and gives a similarity score from 0 to 100 (median per take). It's our comparison score, not an identity test: in these tests, takes that scored 70 or more looked clearly like her. The 96 clips: 72 in the twelve-scene test (3 runs × 12 scenes × 2 takes) and 24 in the other tests (words only, start frames, single-scene fixes, English and Danish voice, two characters, outfit change, slow and fast camera). Each run of 24 clips cost 240 Flow credits; images cost 0 credits. Previous: how Claude connects to Google Flow: [LINK TO VIDEO 1] Next: Omni 1.1 vs Seedance 2.5, same prompts: [LINK] Want to build your own Alex and automate things with people who do this every day? Join The SIKY Way: siky.ai Made with Claude (Anthropic) running Google Flow in Chrome · Google Flow (Omni 1.1 Flash, Nano Banana 2.1, Characters) · Alex's voice: ElevenLabs · Alex on camera: HeyGen #GoogleFlow #AIcharacter #AIvideo