Alias ArchiveArchive in progress
Archive in progress

Before Nine: Across Rooms and Languages

AI Short Drama Trials03 / SERIES

Four characters and five generated passages in a 116-second short, moving from a morning walk into a meeting and back outside.

Filed
Field
Other
Edition
EN / Reading copy
Approximately 1:56. Chinese and Japanese dialogue with the original embedded Chinese subtitles and sound. The web copy is 720p; no English subtitle track has been added.

Two people walk to work and into a meeting. Chinese becomes Japanese; conversation gives way to formal speech and an unspoken inner voice. After the meeting, they walk out together again.

The film keeps its original production number, EP05, and appears third in this selected series. It runs approximately 116.3 seconds, with four characters across five generated passages.

Separate space, identity, and sound

All five passages used Seedance 2.5 at 1080p, followed by manual assembly by the owner. Rather than passing a hard first frame down the sequence, the route used setting references and separate face, full-body, and necessary short voice references.

Setting images controlled topology and materials rather than dictating a camera or pose. Face anchors controlled identity; full-body images supplied proportions and clothing. Some shots required an additional side-face reference. More references made explicit responsibilities more important: an image should not take control of aspects outside its assigned job.

Establish people across shots

The meeting needed four characters to be present in the story without forcing every frame to prove all four identities at once. Relationship shots and singles establish who speaks, who listens, who drifts away, and which line brings attention back.

One table relationship shot omitted a character whose presence was established later. That result was retained for its narrative readability in this film, not as permission to skip seating and off-screen-space checks.

Inner voice and language switching

An inner voice needs more than a “thinks” label. Three conditions must coincide: closed lips, a voice outside the room’s live dialogue, and a body that continues acting. Otherwise the result can become spoken narration or a frozen portrait.

Multilingual speech likewise needs a speaker, language, and intention for each line. A voice reference can constrain timbre without directing the purpose and pauses of every exchange. Naturalness needs continuous listening across the surrounding lines.

Keep the interaction that works

Further constraints and split versions of the ending introduced proportion changes and stiff interaction. The final choice returned to an earlier long-passage route, keeping a natural pause and shoulder interaction. A cut slightly later than the written plan was also accepted.

This preserves one particular performance. It does not establish automatic state transfer between independent generations or remove the need to check multiple characters and languages. The review and accounting approach appears in Keep Rework Ahead of the Final Render.

Related reading

02 / LINKS
C01

Staff Badge: Before the Handover

A 63-second office conversation driven by a packing box, a staff badge, and a departure story that does not hold up.

Read post ↗
C02

Keep Rework Ahead of the Final Render

Three review stages for AI short drama: verify inputs, test the risky passage, and inspect the final delivery independently.

Read post ↗