Nano Banana 2: The 14-Image Reference Workflow Explained
A quick 300-word overview on the essential techniques for nano banana 2: the 14-image reference wo...
Nano Banana 2 is the reference consistency protocol used by professional AI filmmakers on Anamorpha. Named informally after the iterative "peeling back" of a character reference until the model fully internalizes it, the workflow uses a structured set of 14 images to establish a character so thoroughly that the model can generate them consistently from almost any angle, in almost any lighting, across an entire production.
This guide breaks down every step of the protocol.
The Core Principle
Most AI consistency problems come from underspecifying the character. Text descriptions, no matter how detailed, leave vast amounts of visual information to the model's interpretation. Two words like "brunette hair" can produce a hundred different hairstyles. A photograph of a specific brunette hairstyle produces one.
The Nano Banana 2 protocol solves this by flooding the reference with 14 carefully curated images that cover every dimension of visual variation the model is likely to encounter in production.
The 14-Image Set: What They Are
Organize your 14 reference images into five categories:
Category 1: Core Views (4 images)
- Front, neutral expression, full body, studio white background
- 3/4 left, full body, studio white background
- Profile right, full body, studio white background
- Back, full body, studio white background
These four are your character turnaround (see the companion turnaround guide). They are non-negotiable.
Category 2: Facial Expression Range (4 images)
- Neutral / resting
- Subtle smile
- Concern / furrowed brow
- Surprise / open mouth
These teach the model what the character's face does under emotional variation. Without these, the model defaults to the same expression for every emotion state.
Category 3: Lighting Conditions (3 images)
- Bright daylight, front-lit
- Golden hour / warm practical light, 3/4 lit
- Night / low-key, rim-lit
These calibrate the model to how the character's skin, hair, and outfit materials behave under different lighting. Critical for productions that cross multiple lighting environments.
Category 4: Clothing Variations (2 images)
- Primary outfit (what they wear most)
- Secondary outfit (scene-specific — casual, formal, or action-appropriate)
Category 5: Context Integration (1 image)
- The character in a scene with actual environment behind them — not studio white. This teaches the model to place the character correctly within a world.
Building the Reference Set
You don't need real photographs. Generate every image in this set using Image Studio with seed-locked prompting. The front view is your anchor; generate each subsequent image referencing the front view at a high reference strength.
The order matters: 1. Generate front view → lock seed 2. Generate 3/4 and profile using front as reference 3. Generate back using profile as reference 4. Generate expression variations using front as reference (crop to face) 5. Generate lighting variations using 3/4 as base, modify lighting in prompt 6. Generate outfit variations using front as base 7. Generate context integration last, using 3/4 or profile as anchor
Uploading to Character Studio
In Anamorpha's Character Studio, create a new character and upload all 14 images. The studio automatically tags each by category based on background color and composition. Review the auto-tags and correct any misclassifications before saving.
Once saved, the character is available in the model selection panel when generating video or images. Select it as the "Character Reference" rather than providing an image manually.
How the Workflow Improves Consistency
Standard single-image reference typically yields 40–60% consistency across a production (measured by whether the generated character is recognizably the same person across 20 shots). Nano Banana 2 protocol consistently achieves 80–90% consistency on tested productions.
The improvement comes from multiple directions:
- The model sees the character from all angles, eliminating guesswork
- The expression range prevents emotional drift (where the character always looks the same regardless of scene context)
- The lighting references prevent material drift (where skin or fabric looks completely different in a new lighting setup)
Prompt Practices When Using the Protocol
When you've uploaded the full 14-image set, shorten your character descriptions in generation prompts. The model no longer needs detailed description — it has the visual data.
Instead of: a woman with long brown hair, green eyes, wearing a tailored navy blazer and white shirt, sharp features...
Use: [Character Name] from reference sheet, 3/4 view, in the coffee shop interior
Let the reference carry the visual load. Use prompt tokens for camera, lighting, and action instead.
Common Protocol Mistakes
Using AI-generated images with artifacts: Every image in the set must be clean. A morphed ear or distorted hand in the reference set corrupts the model's understanding of the character.
Inconsistent lighting in core views: All four turnaround views should use the same neutral studio lighting. Introducing lighting variation here confuses the model about the character's base appearance.
Skipping expression range: Productions that skip expression images generate characters with one expression regardless of the scene's emotional content. The model defaults to whatever expression appears most in the reference.
Uploading images at low resolution: All reference images should be at least 1024×1024. Low-resolution references produce blurry, low-detail character consistency.
Takeaways
- The 14-image set covers: 4 turnaround views, 4 expressions, 3 lighting conditions, 2 outfits, 1 context
- Generate all 14 using seed-locked prompting from a single master front-view reference
- Upload to Character Studio before starting production — not mid-production
- Shorten character descriptions in generation prompts once the set is uploaded
- Expect 80–90% visual consistency across shots when the full protocol is followed
Continue exploring
- Voice-First AI Directing: Director Studio, from Scene to Screen
- AI Scene Continuity: Inside the Persistent 3D Scene Rig
- The 180 Degree Rule in AI Video: Catch Axis Crossings Before Render
- Character Subtext in AI Video: Render the Tell, Not the Emotion Label