THE GLASS TESTSCENE 01
LIVE-ACTION R2V LAB

GET BUG FOR REAL / MODEL STUDY 01

One scene.
Five models.
No alibis.

A controlled head-to-head test of element/reference-to-video generation using the same live-action cast, the same glass-divided geography, and the same multi-shot dramatic chunks.

Open comparison Read the test design
Approved live-action portrait reference for David
DAVIDIDENTITY AUTHORITY
Approved live-action portrait reference for Youxin
YOUXINIDENTITY AUTHORITY
WARM BEDROOM CLINICAL OBSERVATION
03multi-shot chunks
05model lanes
01locked reference pack
00renders submitted

CONTROL THE INPUT

The reference pack is the experiment.

Each reference has one job. Appearance, geography, and finish are separated so we can tell whether a model follows instructions—or simply averages every image together.

01

Live action is non-negotiable. Natural skin, practical light, photographic depth, restrained performance. No watercolor, illustration, animation, or painterly rendering.

02

Geography survives the medium change. David’s bed is beside a single large plate-glass divide. Youxin occupies the clinical side. They initially feel co-present; the reveal makes the separation legible.

03

The glass has a dramatic function. It separates, receives both hands, then becomes Youxin’s mirror when David’s side goes black.

THE UNIT OF COMPARISON

Multi-shot chunks, not isolated shots.

Each model must solve editorial thought: a beginning, an internal transition, and a resolved end image.

SYNCHRONIZED REVIEW

Five lanes. One playhead.

00:00.0

FAIRNESS PROTOCOL

What stays locked.

  1. 01
    Same dramatic brief

    Identical story beats, spoken intent, geography, casting, aspect ratio, and requested duration.

  2. 02
    Native syntax only

    The content is locked; prompt syntax is adapted to each provider’s documented reference mechanism.

  3. 03
    First viable result

    Record the first complete render before any retries. A retry becomes a separately costed attempt.

  4. 04
    Receipts, not estimates

    Cost and render time come from actual generation receipts and timestamps, then remain editable for audit.

  5. 05
    Two distinct tracks

    True multi-reference video models rank together. FLUX.3 and LTX 2.5 remain visible but are not misrepresented as equivalent R2V APIs.

DECISION SURFACE

Turn impressions into production guidance.