Synthesis Rig Assessment — “The Ferret Incident”

Author: rho-techlead | Step 1 Prep | 2026-05-20

Locked Constraints (from Step 0)

Zero on-screen physical contact between bellhop and ferret
Zero fluid camera movement — locked tripod, whip-cut transitions only
Each shot is a static tableau with max 2 subjects
Wes Anderson aesthetic: meticulous symmetry, pastel palette, flat wide-angle compositions

Optimal API Selection

Image Generation (Steps 3-4: Characters, Settings, Storyboard)

Primary: gemini-3-pro-image-preview (Nano Banana Pro) — default for genmedia-image generate

Best for: Character reference chains, composite sheets, storyboard frames
Why: Supports --reference-image chaining (critical for character consistency). Supports all aspect ratios including 16:9.
Use for: All character headshots, body sheets, scene tests, composite sheets, setting references, and storyboard start/end frames.

Secondary: imagen-4.0-generate-001 (Imagen 4) — via genmedia-image imagen

Best for: High-fidelity environment/setting references where reference chaining isn’t needed
Why: 2K resolution option gives sharper detail for setting reference images (hotel lobby, corridors, desk compositions)
Use for: Master setting references (empty hotel environments) — these don’t need character reference chaining.

Video Generation (Step 5: Principal Photography)

Primary: veo-3.1-fast-generate-001 (Veo 3.1 Fast) — default model

Best for: Base shot generation (4-8s clips)
Why: Supports from-frames (start+end keyframe interpolation), generates audio, fastest turnaround
Use for: All standard tableau shots. The locked-tripod, static-composition approach means from-image is ideal — animate one keyframe with minimal motion described in the prompt.

From-Frames vs From-Image Decision:

From-Image (genmedia-video from-image): Use for shots with minimal visual delta (character holds pose, clock ticks, static composition with subtle ambient motion like curtain sway). This is the MAJORITY of our shots given the static tableau mandate.
From-Frames (genmedia-video from-frames): Use for shots with defined visual change (e.g., start: tidy desk → end: toppled lamp/scattered papers). The start/end frames lock the transition.

Extend model: veo-3.1-lite-generate-001 (Veo 3.1 Lite)

Only model supporting extend. Each extend adds exactly 7 seconds.
Use sparingly — most Anderson-style shots are short (4-6s). Reserve for any establishing shot or montage beat that needs >8s.

Voice Generation (Step 6: Narration & Dialogue)

Model: gemini-3.1-flash-tts-preview via genmedia-voice generate

Voice selection TBD after character profiles are finalized
Narrator voice: Need dry, clinical, measured — candidates: Fenrir (authoritative, measured), Enceladus, or Schedar. Will audition in Step 3.
Bellhop voice: Need uptight, precise — shorter lines
Inspector voice: Need pompous, clipped — brief appearances only
800-char limit per call — split longer narration passages

Music Generation (Step 6: Score)

Primary: lyria-3-pro-preview (Lyria 3 Pro)

~2:30 duration — sufficient for full musical arc segments
Use for: Main score (quirky plucked strings / glockenspiel motif per editor’s musical arc)

Secondary: lyria-3-clip-preview (Lyria 3 Clip)

~30s clips — use for specific SFX stingers or bridge cues

Wes Anderson Prompt DNA

Every image and video prompt MUST include these Tone Anchor keywords (pending final Tone Contract from rho-idea, these are my technical recommendations):

Wes Anderson style, meticulous symmetry, centered composition, 
pastel color palette, flat wide-angle lens, locked tripod shot, 
warm soft lighting, whimsical, deadpan, precise staging

Anti-Drift Keywords (Genre Counterbalance)

Models will try to make this dramatic/moody. Fight it with:

bright even lighting, NO shadows, NO dramatic contrast, 
NO cinematic depth of field, flat perspective, 
saturated pastels, theatrical staging, comedy

Ferret-Specific Prompts

To combat anthropomorphism bias:

realistic white ferret, no cartoon features, no googly eyes, 
natural animal proportions, literal ferret behavior

Reference Budget Strategy (Per Shot Type)

Shot Type	Ref 1	Ref 2	Ref 3
Bellhop solo	bellhop_sheet.png	hotel_setting.png	(available)
Ferret solo	ferret_sheet.png	hotel_setting.png	(available)
Aftermath/environment	hotel_setting.png	(available)	(available)
Bellhop + Inspector	bellhop_sheet.png	inspector_sheet.png	hotel_setting.png
Insert (bell, clock, ledger)	object_ref.png	hotel_setting.png	(available)

Budget is comfortable for every shot type. Never exceeds 3.

Production Efficiency Notes

Python orchestration — all batch generation scripts will use /workspace/tools/genmedia.py wrapper, never raw bash loops (quote-safe prompts with special characters).
Overhang Principle — every shot generated with +4s (2s pre-roll + 2s post-roll) for trim flexibility.
QA Helper delegation — batch generation execution and verify-dailies checks will be delegated to a qa-helper agent per the Rule of 10.
Skip-existing — all scripts use skip_existing=True (default) for safe re-runs after partial failures.
Shot manifest — shot-manifest.json generated alongside principal photography scripts for verify-dailies gate.