Generatability Audit — Team Rho

Auditor: rho-techlead (Director of Photography)
Date: 2026-05-20
Methodology: Constraint Matrix (Character Count, Interaction Complexity, Continuity Strategy, Safety Pre-Check)

Spark 1: “The Gilded Teacup” — Magical Realism / Fantasy Drama

Character Count: LOW RISK ✅

1 human (girl) + 1 object (teacup/tree)
Reference budget: 1 character sheet + 1 setting reference + 1 object reference = 3. Exactly at Veo’s limit.
The tree is an evolving object (seed → sprout → full miniature willow → storm-threatened). Will need 2-3 reference states for the tree, used at different story beats rather than simultaneously.

Interaction Complexity: LOW-MEDIUM ⚠️

Girl interacts with the teacup (drops seed, shields tree). These are proximity interactions with an object, not character-to-character contact.
The seed-drop moment is the riskiest — hand + small object + precise placement. Mitigation: Use a macro/ECU insert shot of the hand releasing the seed. Cinematic Obscurity keeps this clean.
The “protection” action (shielding the tree from rain) can be framed as the girl hovering her hands above the teacup — no contact needed.

Continuity Strategy: RECURSIVE SYNTHESIS ANCHORS

Single setting (windowsill/interior room) keeps environment stable.
The tree evolves through growth stages — chain reference images at each stage. Generate tree-state references in Step 3 alongside the character sheet.
Weather shift (golden hour → storm) changes the lighting. Generate a setting reference for each lighting condition: warm golden and dark storm.
Character stays constant — one girl, one outfit. Standard reference chain handles this.

Safety Pre-Check: LOW RISK ✅

“Young girl” — recommend prompting as “young woman” or “teenage girl, age 16” with clear stylistic context (anime/Ghibli) to avoid ambiguity.
Magical/glowing elements: model-safe. Fantasy genre anchors this clearly.
Rainstorm: model-safe.
No realism triggers.

Generatability Verdict: HIGH ⭐⭐⭐⭐

Single character, one primary location, object-focused narrative. Ghibli/anime aesthetic is well-supported by Nano Banana. Primary challenge is the tree’s evolving appearance — manageable with staged reference images. The emotional arc (wonder → danger → resolution) maps well to visual storytelling without relying on dialogue.

Spark 2: “The Standing Stones” — Folk Horror / Psychological Thriller

Character Count: LOW RISK ✅

1 human (woman in red coat) + environment (moor, standing stones)
Reference budget: 1 character sheet + 1 moor setting + 1 stone formation = 3. At budget but clean.
The red coat is a strong visual anchor — high contrast against grey moor means the model can’t lose her.

Interaction Complexity: VERY LOW ✅

Zero character-to-character interaction. The woman observes, walks, and reacts. The stones are environmental.
Camera work is slow zooms and static shots — this is what Veo does best.
The “eerie stillness” the production notes mention is literally the default mode of AI video generation. The concept weaponizes a model limitation as an aesthetic feature.

Continuity Strategy: FROM-FRAMES MOTION PRIORITY

Slow, deliberate camera moves (zooms, pans, push-ins) map perfectly to from-frames interpolation. Generate start frame + end frame, let Veo fill the motion.
The stones deliberately shift between shots — this relaxes continuity requirements on the environment. We only need the moor and fog to be consistent, not the exact stone positions.
Woman’s red coat provides an unambiguous continuity anchor across any setting variation.
Fog/atmospheric haze naturally masks minor inconsistencies between shots.

Safety Pre-Check: MEDIUM ⚠️

“Horror” genre label could bias model toward gore/violence. Mitigation: Prompt as “atmospheric mystery” and “psychological suspense”, never use “horror” in generation prompts.
Woman alone in a threatening landscape — frame as “exploration” and “discovery”, not “danger” or “threat.”
Stones “closing in” — describe as “the landscape shifting around her” rather than anything physically threatening.
1970s grain aesthetic — LOW RISK, purely stylistic.
Overall manageable with consistent prompt sanitization.

Generatability Verdict: VERY HIGH ⭐⭐⭐⭐⭐

This is the most technically feasible spark. It directly exploits AI video strengths: atmospheric environments, slow camera work, single character, minimal physical action. The 1970s film grain aesthetic actively masks artifacts. The fog obscures edges where models typically fail. The concept turns AI limitations into aesthetic virtues. From a pure synthesis standpoint, this is a slam dunk.

Spark 3: “The Ferret Incident” — Deadpan Indie Comedy

Character Count: MEDIUM-HIGH RISK ⚠️

2 humans (bellhop + hotel inspector) + 1 animal (ferret)
When bellhop and inspector share the frame: 2 character sheets + 1 setting = at budget, no room for the ferret reference.
The ferret is a primary narrative element but can’t be referenced alongside two humans. Must alternate: bellhop+ferret shots and bellhop+inspector shots, never all three.

Interaction Complexity: HIGH RISK ❌

The core narrative action is a human chasing and catching a small, fast-moving animal. This is contact interaction at its most challenging.
Veo will struggle with: ferret running across surfaces, bellhop lunging/grabbing, any direct human-animal contact.
Comedy timing depends on precise physical choreography — the exact thing generative video handles worst.
Mitigation (Cinematic Obscurity): Could decompose into reaction shots (bellhop’s face → cut to ferret alone → cut to aftermath). But this undercuts the visual comedy timing that makes Anderson-style humor work. The mitigation fights the concept.

Continuity Strategy: RECURSIVE SYNTHESIS ANCHORS (DIFFICULT)

Wes Anderson symmetry is a double-edged sword. Models CAN generate symmetrical compositions per-shot, but maintaining identical spatial geometry (same desk, same wallpaper pattern, same prop placement) across many shots is extremely hard.
The ferret’s appearance — small white animal — will drift significantly across shots. Anchoring a ferret with reference images is unreliable; models struggle with small animal consistency.
The pastel color palette + flat lighting helps somewhat (fewer shadows to go wrong).
Will require heavy reference chaining and likely multiple regeneration cycles.

Safety Pre-Check: LOW RISK ✅

Hotel setting, comedy genre, animal character — all model-safe.
No realism triggers, no violence, no sensitive archetypes.

Generatability Verdict: MEDIUM-LOW ⭐⭐

The concept is charming but technically the hardest of the three. The core comedy depends on physical interaction (catching the ferret) which is the weakest capability in current video models. Maintaining ferret consistency and Anderson-style spatial precision will burn significant regen cycles. The reference budget is tight when all three characters share scenes. Risk of spending most of production time fighting the tools rather than telling the story.

Summary & Recommendation

Spark	Gen. Score	Character Risk	Interaction Risk	Continuity	Safety
1. Gilded Teacup	⭐⭐⭐⭐ HIGH	Low	Low-Med	Recursive (manageable)	Low
2. Standing Stones	⭐⭐⭐⭐⭐ VERY HIGH	Low	Very Low	From-Frames (ideal)	Medium (manageable)
3. Ferret Incident	⭐⭐ MED-LOW	Med-High	High	Recursive (difficult)	Low

My recommendation: Spark 2 (“The Standing Stones”) is the strongest from a generatability standpoint. It plays directly to every AI video strength and turns model limitations into aesthetic features. The From-Frames pipeline maps perfectly to the slow, deliberate camera work. The safety concerns are medium but fully manageable with prompt sanitization.

Spark 1 (“The Gilded Teacup”) is a strong second — highly feasible with a beautiful aesthetic that Nano Banana handles well. Slightly more continuity management needed for the evolving tree object.

Spark 3 (“The Ferret Incident”) is technically the most ambitious. The comedy depends on precise physical choreography that current models can’t reliably deliver. I’d steer away from this unless the team is prepared for heavy regen cycles and a potential concept pivot mid-production.

I’m comfortable executing either Spark 1 or Spark 2. Happy to discuss trade-offs.