# Higgsfield Image / Video / Voice Model Selection v0.1

Status: active benchmark policy for CartoonOS / Luminautas.
Date: 2026-09-08.

## Principle

**Do not select one global best model. Select the best model for the exact production task.**

CartoonOS benchmarks three independent modalities:

- image
- video
- voice

Each uses different units, hard gates and economics.

## 1. Image model track

Current live Higgsfield image catalog includes, among others:

### Google / Gemini Nano Banana family
- `nano_banana` — Nano Banana
- `nano_banana_pro` — Nano Banana Pro
- `nano_banana_2` — Nano Banana 2
- `nano_banana_2_lite` — Nano Banana 2 Lite

Observed live preflight examples on 2026-09-08:
- Nano Banana: ~1 credit/image
- Nano Banana Pro 2K: ~2 credits/image
- Nano Banana Pro 4K: ~4 credits/image
- Nano Banana 2 2K: ~2 credits/image
- Nano Banana 2 Lite 1K: ~1 credit/image

Pricing must be stored as a dated snapshot because provider pricing can change.

Other relevant image models include:
- GPT Image 2 / 2.5
- Seedream 4.5 / 5.0
- FLUX.2 / Flux Kontext
- Recraft V4.1
- Cinema Studio Image
- Soul / Soul Cinema
- Kling O1 Image
- Grok Image

### Image benchmark archetypes

- character sheet
- turnaround
- expression pack
- transparent cutout
- background plate
- storyboard frame
- educational diagram
- start/end video frame

### Image KPIs

Quality:
- visual quality
- exact character identity consistency
- anatomy consistency
- silhouette consistency
- palette/material consistency
- expression readability
- prompt adherence
- text/diagram accuracy when required
- clean-background/transparency quality

Efficiency:
- credits per requested image
- first-pass acceptance
- credits per accepted image
- generation latency p50/p95
- QA/rework time
- rejected-image rate

Character assets have a hard identity floor >=95 before economics may influence selection.

## 2. Video model track

Relevant live models include Seedance, Kling, Cinema Studio, MiniMax, Wan, Gemini Omni, Veo, FLUX Video and other current catalog entries.

### Video archetypes
- character close-up
- character dialogue
- multi-character
- cinematic establishing
- short B-roll
- Clarity Island
- motion transfer

### Video KPIs
- visual quality
- character identity
- anatomy
- prompt adherence
- motion/physics
- temporal continuity
- audio/lipsync where applicable
- first-pass acceptance
- provider success
- retries/rework
- credits/output second
- credits/accepted second
- accepted seconds/hour
- end-to-end latency

**Primary economics metric:** `credits_per_accepted_second`, not raw credits per generation.

## 3. Voice model track

The canonical character voice is distinct from the TTS engine.

Benchmark engines only against an already-approved mascot voice identity.

Current live TTS engine families include:
- Seed Audio 1.0
- Qwen Audio 3.0 TTS Flash
- Text to Speech V2 variants: ElevenLabs, MiniMax, Seed Speech, Vibe Voice, Cozy Voice
- other compatible engines from the live catalog

### Voice KPIs
- canonical voice identity similarity
- pronunciation accuracy
- ES-419 fit
- age/personality fit
- emotional readability
- naturalness
- pacing consistency
- pitch consistency
- audio artifact rate
- lipsync compatibility
- generation latency
- cost per accepted dialogue minute
- first-pass acceptance

Hard gates for canonical mascot voices:
- identity similarity >=95
- canonical lexicon pronunciation >=98
- emotional readability >=90
- lipsync QA pass where required

## 4. Canonical voices

Every mascot has one active versioned voice identity:
- Lumi Voice v1
- Tiko Voice v1
- Nova Voice v1
- Marina Voice v1
- Rin Voice v1
- Arqueo Voice v1
- Pepe Voice v1

Exact provider `voice_type + voice_id` values remain unbound until audition and explicit approval. Never invent them.

Shorts and flagship episodes must reuse the same active mascot voice binding.

A provider/engine migration requires matched-sample regression testing and explicit version approval.

## 5. Model selector logic

For each generation request:

1. classify modality;
2. classify archetype;
3. determine exact character versions;
4. apply hard quality floors;
5. filter models by live capability snapshot;
6. use matched benchmark cohort only;
7. require enough sample confidence;
8. estimate expected total accepted cost, not raw request cost;
9. account for latency and reliability;
10. select best eligible configuration;
11. retain alternatives for fallback;
12. store decision + reason + benchmark snapshot version.

## 6. Confidence policy

- `<5` matched samples: `insufficient_sample`
- `5–19`: `provisional`
- `>=20`: `decision_ready`

No low-sample model is declared a permanent winner.

## 7. Expected-total-cost formula

A useful production estimate is:

`expected_total_cost = request_cost / first_pass_acceptance_rate + expected_rework_cost + QA_cost`

For video, normalize further to accepted seconds/minutes.

For voice, normalize to accepted dialogue minutes.

For image, normalize to accepted production assets.

## 8. Nano Banana evaluation plan

Benchmark all four active Nano Banana variants against the same controlled asset briefs:

1. Lumi six-view turnaround
2. Tiko six-view turnaround
3. Lumi expression pack
4. Tiko expression pack
5. clean 16:9 background
6. Curiosity Graph educational diagram
7. 9:16 Short hook frame
8. start/end keyframe pair

Evaluate exact identity, anatomy, consistency across panels, text/diagram quality, cost and latency.

Do not crown Nano Banana based on price alone.

## 9. Production recommendation policy

The dashboard should output recommendations like:

- `best_for_character_turnaround`
- `best_for_expression_pack`
- `best_for_clean_background`
- `best_for_diagram`
- `best_for_character_dialogue_video`
- `best_for_multi_character_video`
- `best_for_fast_short_broll`
- `best_for_lumi_voice`
- `best_for_tiko_voice`

Avoid a single `best_model` field.

## 10. North-star production economics

Track at system level:

- cost per accepted second
- cost per accepted dialogue minute
- cost per approved image asset
- finished minutes per production hour
- first-pass acceptance
- late rework rate
- total production cost per finished Short
- total production cost per finished flagship minute
- quality-adjusted production cost

The goal is **high-quality accepted output per dollar and production hour**, not lowest nominal API price.