Draft 3 Planning

1. Prompt Structure

The Recommended Order

GPT Image 2 rewards structure over adjectives. Words like "stunning" or "beautiful" do very little. What works is organized, specific description in this order:

• Material/Medium Block — What the image is made of (torn paper, gold foil, gouache, watercolor paper)
• Use Case / Format — What the image is for (children's book illustration, full bleed 3:2)
• Scene/Setting — The environment, background, lighting conditions
• Subject(s) — Each character described individually with full anchors
• Environment Details — Specific visual attributes, sea life, props
• Rendering / Palette — Colors, light direction, depth zones
• Constraints — What to avoid or preserve

Why Order Matters

The model allocates more attention to earlier tokens in the prompt. Putting materials and format first tells the model "this is a mixed-media collage" before it starts imagining the scene, which fundamentally changes how it renders everything that follows. Burying the style at the end means the model may have already committed to a photorealistic rendering before it sees "watercolor."

Prompt Length

GPT Image 2 follows long prompts more reliably than most other models. Our best results used 300-500 word prompts. If you're getting generic results, the answer is usually more specificity, not less. Be explicit about every element you care about.

Concrete Example (from our Spread 01 prompt)

Dry torn paper coral with visible fiber texture exists deep
underwater alongside photographic reef life.
Gold foil metallic stipple shimmers on surfaces submerged
in real ocean water.
Gouache painted characters swim through photographic
Hawaiian reef. Heavy watercolor paper base.

Full bleed 3:2 children's book illustration. Mixed media
collage blending with underwater photography.

SCENE: A dramatic steep Hawaiian reef drop-off descending
into the deep...

OLLIE: A cartoon baby octopus. [full character anchor with
colors, proportions, anti-patterns]...

DOT: Solid natural Tahitian pearl. [full character anchor
with eye shape, anti-patterns]...

ENVIRONMENT: Rich shallow coral reef dropping into deep...

RENDERING: Warm golden god-ray light transitioning to deep
navy. [hex palette]. Full bleed 3:2 aspect ratio.

Text Rendering

GPT Image 2 achieves ~99% accuracy on text (vs ~60% for DALL-E 3). Place exact text in quotation marks or ALL CAPS, specify font style and placement explicitly. Best at 1-5 words per text element. Not relevant for our book illustrations (we avoid in-image text).

2. Texture Artifacts and Watermarking

What the Artifacts Are

Shortly after GPT Image 2 launched on April 21, 2026, users reported a consistent visual artifact: generated images were layered with persistent tiling textures and grime artifacts:

• Muddy smears or dirt overlays on photorealistic outputs
• Tiling noise patterns visible across interiors, landscapes, and other scenes
• Subtle grime textures that appear embedded at the pixel level
• Noise pattern amplification where artifacts compound across successive generations

What Causes It

Primary cause: Steganographic Watermarking. OpenAI embeds a dual-layer watermark into every generated image: (1) C2PA metadata and (2) an imperceptible pixel-level watermark with no public detector yet. The pixel watermark is baked into the generation process itself, creating visible texture patterns especially on smooth, uniform surfaces. Downstream workflows like upscaling or compositing can amplify these artifacts.

Secondary cause: Context-dependent noise accumulation. The model propagates noise from previous generations within the same conversation. Artifacts carry over and compound with each successive image in a chat thread. OpenAI fixed a noise amplification bug, but some baseline texture persists.

Tertiary cause: Training data contamination. GPT Image 2 has been observed generating images containing visible Gemini branding and other AI-model artifacts — evidence that AI-generated content in the training data introduces learned artifacts.

Which Images Show Artifacts MORE

• Photorealistic renders with smooth surfaces (skin, fabric, plastic)
• Sky gradients and large uniform color fields
• Interior scenes with walls, floors, ceilings
• Organic landscapes with dense foliage (forests reveal a distinct synthetic rendering pattern)
• Any image where you zoom in on micro-textures (skin pores, fine detail)

Which Images Show Artifacts LESS

• Heavily textured subjects where noise blends in (collage, mixed media)
• Watercolor and gouache styles (intentional texture masks artifact texture)
• Subjects with inherent visual complexity (dense coral reefs, busy compositions)
• Stylized/illustrated looks rather than photorealistic

Community Workarounds

• AI eraser spot cleanup: Export the image, use an AI eraser tool to brush over specific artifact areas — faster than regenerating the whole image
• Upscale-then-downscale pipeline: Upscale with an external tool (Real-ESRGAN, Topaz), then downscale back. The upscaler smooths over watermark artifacts during reconstruction
• Controlled grain language: Specify intentional grain ("35mm film aesthetic," "Kodak Portra 400 look") to replace artifact grain with controlled, aesthetically pleasing grain
• Fresh API calls: Each generation should be a fresh API call with no conversation history
• Positive surface language: Describe what you want: smooth matte finish, clean flat color
• Use quality: "high" for finals — lower quality settings show more compression artifacts

Our Batch Analysis (May 5-6, 2026)

Analysis of our 5-spread OpenAI batch revealed likely artifact amplifiers in our prompts:

• Global texture language: Phrases like "gold foil metallic stipple shimmers on surfaces" gave the model permission to add texture everywhere, including areas that should be smooth. Fix: target texture to specific elements ("gold foil stipple on Ollie's skin only")
• "Maximum density" language: Spread 05 used "maximum density" which likely worsened artifacts by encouraging the model to fill negative space with noise/texture
• Mixed photorealistic/illustrated language: Asking for "photographic Hawaiian reef" alongside "gouache painted characters" forces a hybrid rendering mode where photorealistic elements are more susceptible to watermark artifacts
• Our style inherently helps: The mixed-media collage look (torn paper, watercolor, gouache) naturally masks artifacts that would be obvious on photorealistic surfaces

3. Resolution Capabilities

Actual Specifications

Parameter	Value
Supported sizes	1024×1024 (1:1), 1536×1024 (3:2), 1024×1536 (2:3), auto
Our output	1536×1024 (3:2 landscape)

Key finding: GPT-Image-2 only supports 4 fixed sizes: 1024×1024 (1:1), 1536×1024 (3:2 landscape), 1024×1536 (2:3 portrait), and auto. For print-quality book illustrations, generate at 1536×1024 (3:2) and upscale externally (Real-ESRGAN, Topaz Gigapixel).

4. Character Consistency

What Works

• Detailed character anchors repeated in every prompt — full physical description with colors, proportions, materials. Our Ollie anchor is ~100 words; Dot is ~80 words. Both must appear in EVERY prompt
• Explicit anti-pattern lists — "NO mouth, NO ears, NO antenna, NO hardware" prevents the most common character drift
• Reference images labeled by role — "Image 1: character sheet. Image 2: approved style reference." Up to 16 reference images supported
• Iterative refinement language — "Maintain the same composition and character but change [specific element]"

What Does Not Work

• Vague descriptions — "a cute octopus" produces different characters every time
• Relying on model memory — the model does NOT remember characters between separate API calls
• Short character blocks — fewer details = more creative freedom for the model = more drift

Our Batch 1 Issues

• Dot rendered with mouths, crescents, and arcs rather than the specified glowing U-shapes. Needed more aggressive anti-patterns and ALL-CAPS emphasis on the eye shape
• Ollie occasionally got ears despite "NO EARS" in the prompt. Repeating "NO EARS" in the rendering block as well as the character block improved compliance
• Scene concepts came through well (reef, depth zones, color palette), but character-level detail was the weak point. The model allocates attention broadly across long prompts, so character-critical details need emphasis (CAPS, repetition, anti-patterns)

5. Scene Accuracy and Anatomical Errors

The Multi-Species Problem

GPT Image 2 struggles with complex multi-species underwater compositions. When a prompt lists many species, the model often:

• Morphs multiple species together (fish with wrong fin configurations, hybrid creatures)
• Generates blob-like creatures that do not match any real species
• Adds extra fins, limbs, or appendages to animals
• Gets species colors wrong (yellow tang rendered as blue, etc.)

How to Fix

• Limit unique species to 4-6 per scene — fewer species = more attention per species = better anatomy
• Describe each creature individually with real anatomy: "a Moorish idol with its distinctive tall triangular dorsal fin, black-and-white vertical banding, and trailing yellow caudal filament"
• Be explicit about count and placement: "three yellow tangs swimming left-to-right through the midground"
• Separate foreground from background creatures: give detailed anatomy only to foreground species; background species can be silhouettes or "schools of small fish"
• Use real species names with key identifying features: the model renders more accurately when given the correct name plus 1-2 anatomical features

Our Batch Findings

Spread 05 requested maximum density with 10+ named species and showed morphed fish and anatomically incorrect creatures. Spread 01 and 08, which had fewer species described more carefully, produced better individual creature accuracy. For future batches: prioritize quality of each creature over quantity of species.

6. Editing and Reference Images

How Reference Images Work

GPT Image 2 processes all reference images at high fidelity automatically with no adjustable knob.

Multi-Image Input

• Label each by number and role: "Image 1: character reference. Image 2: style reference. Image 3: background scene."
• Describe how they interact: "Apply the style from Image 1 to the character in Image 2, placed in the setting of Image 3."
• Downscale references to what the task actually needs

The Edit Pattern

Change: [exactly what should change]
Preserve: [face, identity, pose, lighting, framing, background, geometry, text, layout]
Constraints: [no extra objects, no redesign, no logo drift, no watermark]

7. API Parameters

Setting	Use Case	Cost (1024x1024)
low	Fast drafts, thumbnails	~$0.006
medium	General use	~$0.053
high	Final assets, dense layouts	~$0.211
auto	Let the model decide	varies

Output Format

Format	Notes
png	Default. Lossless. Best for illustration work.
jpeg	Faster than PNG. Good when latency matters.
webp	Smallest file size. Good for web delivery.

8. GPT Image 2 vs DALL-E 3

Feature	GPT Image 2	DALL-E 3
Text rendering accuracy	~99%	~60%
Texture/photorealism	More natural, contextually grounded	Good but less precise
Character consistency	Strong with proper anchoring	Inconsistent
Prompt following	Follows long, structured prompts well	Better with shorter prompts
Max resolution	Up to 2K (2048x2048)	1024x1024

9. Recommendations for Ollie and Dot

Avoiding Texture Issues

• Fresh API call for each image — never iterate within a conversation
• Use quality: "high" and png output format for all finals
• Target texture language to specific elements, not globally
• Keep "maximum density" language out of prompts — specify exact counts and placements
• Post-process with AI eraser for spot artifacts rather than regenerating entire images
• Our mixed-media style is a natural advantage — lean into it

Improving Character Accuracy

• Repeat character anchors verbatim in every prompt (100+ words per character)
• Use ALL-CAPS for critical anti-patterns: "NO EARS," "NO MOUTH"
• Repeat critical constraints in both the character block AND the rendering block
• Attach 2-3 approved reference images labeled by role

Improving Scene Accuracy

• Limit to 4-6 unique species per scene
• Describe each creature with real anatomical features
• Specify exact counts and positions: "three yellow tangs in the midground"
• Use "schools of small colorful fish" for background density

Iteration Workflow

• Concept phase: quality: "low", rapid iteration on composition and pose
• Refinement phase: quality: "medium", nail down details and expressions
• Final render: quality: "high", png format, full resolution
• Post-processing: AI eraser for spot artifacts, external upscaling for print resolution

10. Sources

• OpenAI Developer Community — GPT-image-generator 2.0 issues, bugs, and workarounds thread. Primary source for texture artifact reports, noise amplification fix, community workarounds
• Startup Fortune — GPT Image 2 grime artifacts and watermark strategy analysis. Steganographic watermark analysis, pixel-level provenance embedding
• OpenAI Help Center — C2PA in ChatGPT Images. Official C2PA metadata documentation
• OpenAI Cookbook — Image Generation Models Prompting Guide. Official prompt structure recommendations
• fal.ai — GPT Image 2 Prompting Guide. Structured prompting best practices, edit patterns
• a2a-mcp.org — GPT Image 2 Prompts Guide 2026. Advanced prompt techniques, character consistency
• Framia — GPT Image 2 Resolution analysis. Native 2K ceiling, resolution specifications
• PixVerse — GPT Image 2 Review 2026. Prompt guide, use cases, quality comparison
• PixNova — GPT Image 2 48-Hour Stress Test. Post-processing workflows, AI eraser technique
• ImagesPlatform — GPT Image 2 In-Depth Technical Review. 4K claims vs reality, character consistency
• Our own generation experiments — Batch 1 (5 spreads, May 5-6 2026): prompt analysis of spreads 01, 05, 07, 08, 13

1. Midjourney --sref: How It Works

Mechanism

Midjourney's --sref (Style Reference) is a parameter that transfers the aesthetic of a reference image or a numeric style code to a new generation. It operates in two modes:

• Image URL mode: --sref [image URL] — Midjourney analyzes the image and attempts to transfer its style.
• Numeric code mode: --sref [code] — A numeric identifier mapping to a pre-defined internal style within Midjourney's latent space.

What --sref Captures

• Drawing/collage techniques and rendering method
• Photography styles (if photographic)
• Color palettes and contrast levels
• Light and shadow ratios
• Compositions and layouts
• Texture and surface quality
• Overall mood/atmosphere

Internal Implementation

Midjourney has not published their architecture, but based on community analysis:

• Style codes are points in a learned style latent space. Each numeric code maps to a specific region producing a reproducible aesthetic.
• Image-based --sref likely uses a vision encoder (possibly CLIP-based) to extract style features, then conditions the diffusion process on those features.
• The system separates content from style. --sref transfers aesthetics without subject matter; --cref does the opposite.
• Style versions (--sv) indicate different encoder architectures. The current default is --sv 6.

--sref vs --cref

Feature	--sref (Style Reference)	--cref (Character Reference)
Purpose	Keep the "camera"/art direction the same	Keep the "actor"/character the same
Captures	Colors, textures, lighting, composition, rendering style	Facial features, hair, body proportions, clothing
Weight param	--sw (0-1000, default 100)	--cw (0-100, default 100)
Metaphor	Same cinematographer, different scene	Same actor, different scene

Why It Works Well

• Dedicated style encoding — Style is extracted as a separate conditioning signal
• Numeric reproducibility — Same code always produces the same aesthetic
• Weight control — The --sw parameter lets you blend style influence
• Multiple references — You can provide multiple style images to blend

Limitations

• Illustrations get more creative reinterpretation than photographs
• Not available via API for external tooling (Discord-only or Midjourney web)
• Style codes are opaque; cannot engineer a code for a specific style without trial and error

2. Nano Banana 2 / Gemini Style Capabilities

Model Overview

• Nano Banana Pro = Gemini 3 Pro Image — highest quality, up to 5 character + 6 object references (11 total)
• Nano Banana 2 = Gemini 3.1 Flash Image — faster, cheaper, up to 4 character + 10 object references (14 total)

Style Reference Image Support

Capability	Nano Banana Pro (3 Pro)	Nano Banana 2 (3.1 Flash)
Character reference images	Up to 5	Up to 4
Object fidelity images	Up to 6	Up to 10
Total reference images	Up to 11	Up to 14
Supported formats	PNG, JPEG, WebP, HEIC, HEIF	PNG, JPEG, WebP, HEIC, HEIF

Important: The 14-image limit is NOT freely allocatable. Character and object quotas are independent.

Best Approach for Style Consistency

• Provide 2-4 approved style reference images alongside each generation prompt
• Use explicit referential language: "Use the same watercolor style, color palette, and line weight as the reference images."
• Maintain a character sheet and use it as reference for all subsequent generations
• Text prompts should precede reference images in the API call
• Recommended order: Subject > Composition > Action > Location > Style > Editing instructions

3. GPT-Image-2 Style Reference Capabilities

GPT-Image-2 supports up to 16 reference images (JPEG, PNG, WebP, under 30MB each) via the images.edit endpoint.

How Style Matching Works

• Analyzes provided reference images for visual features at high fidelity automatically
• Applies style, palette, edge treatment, and silhouette language from references
• Maintains character consistency across variations, style transfers, and partial edits

Best Prompt Structure for Style Consistency

Image 1: Character reference sheet showing Ollie the fox
Image 2: Approved illustration showing target color palette
Image 3: Scene composition reference

Generate a new illustration of Ollie walking through a forest.
Apply the watercolor style and color palette from Image 2.
Maintain Ollie's exact appearance from Image 1.
Use similar composition depth as Image 3.

Cost Structure

Quality	Approx Cost per Image (1024x1024)
Low	~$0.006
Medium	~$0.053
High	~$0.211
Batch (50% discount)	Half of above

4. Automated Style Extraction Approaches

Vision Models as Style Analyzers

Both Gemini and Claude can analyze an approved illustration and produce a structured style description, identifying:

• Color palette: Dominant colors, accent colors, approximate hex values
• Rendering technique: Watercolor, gouache, digital paint, pencil
• Line quality: Weight, softness, color of outlines
• Texture: Paper grain, brush strokes, noise patterns
• Lighting: Direction, warmth, contrast ratio, shadow softness
• Composition patterns: Typical framing, depth, character-to-background ratio
• Character proportions: Head-to-body ratio, stylization level

Practical Workflow

• Select 3-5 approved "gold standard" illustrations
• Send each to a vision model with a structured analysis prompt
• Aggregate and normalize the extracted attributes into a single style profile
• Use the style profile as a prompt prefix for all subsequent generations
• Periodically re-validate by comparing new generations against gold standards

Academic Research

• StyleTokenizer (ECCV 2024): Trains a dedicated style feature extractor on 30,000 images across 300+ style categories. Extracts style from a single reference image and injects it into a diffusion model.
• StyleBrush (2024): Extracts and transfers style from a single image using a diffusion model trained on Wikiart. Demonstrates single-image style extraction is viable for production.
• Semantic Guidance (Nature, 2025): Uses semantic segmentation to decouple multiple styles from reference images with parallel decoupling adapters.

Key takeaway: For our purposes, using a vision model to extract a structured text description is more practical than training custom models. It is free/cheap, fast to iterate, and the output (text) can be directly used as prompt input for both NB2 and GPT-Image-2.

5. Professional AI Illustration Studio Best Practices

Tier 1: LoRA Fine-Tuning (Most Reliable, Most Effort)

• Train a custom LoRA model on 15-30 character reference images
• Produces 95%+ consistency for books over 20 pages
• Not applicable since NB2 and GPT-Image-2 don't support custom LoRA injection

Tier 2: Reference Image + Prompt Engineering (Our Sweet Spot)

• Create a canonical character sheet (front view, side view, expressions, outfit variations)
• Build a style guide with palette hex codes, line weight, shading rules
• Use a base prompt template with immutable character traits separated from variable scene elements
• Append scene-specific details to the template for each illustration

[STYLE BLOCK - immutable]
Watercolor children's book illustration. Soft pencil outlines
with warm earthy palette.

[CHARACTER BLOCK - immutable per character]
Ollie: small orange fox with oversized round head, bright
curious eyes (#4169E1 blue), wearing a green scarf (#2E8B57).

[SCENE BLOCK - variable]
Ollie standing at the edge of a meadow, looking up at
fireflies. Golden hour lighting from the left.

Quality Control Process

• Automated similarity scoring — SSIM and LPIPS metrics vs reference standards
• Human review — Side-by-side comparison with approved references
• Targeted regeneration — Failed pages get img2img refinement, not full recreation

6. Recommended Approach for Ollie and Dot

Strategy: Vision-Extracted Style Profile + Multi-Reference Generation

• Step 1: Select 3-5 gold standard illustrations from existing corpus
• Step 2: Extract structured style profiles via vision model (Gemini/Claude)
• Step 3: Create three-block prompt templates (Style + Character + Scene)
• Step 4: Generate with 2-3 gold standards + 1-2 character sheets as reference images
• Step 5: Automated QC loop — vision model scores consistency, regenerate if score < 7

Factor	Rationale
No LoRA needed	Works natively with NB2 and GPT-Image-2 APIs
Low setup cost	Vision-based extraction is essentially free
Iterative	Profile can be refined as more illustrations are approved
Cross-model	Same style profile works for both NB2 and GPT-Image-2
Automated QC	Vision models can score consistency without human bottleneck

7. Experiment Plan

Phase 1: Style Profile Extraction (Cost: ~$0-1)

Select gold standards, run style extraction, merge into canonical profile, validate with text-only generations.

Phase 2: Reference Image Strategy Testing (Cost: ~$2-5)

Test 0, 1, and 3 reference images on both NB2 and GPT-Image-2. Identify optimal consistency-to-cost ratio.

Phase 3: Prompt Template Optimization (Cost: ~$2-3)

Test prompt ordering variations: style-first, scene-first, and interleaved. Measure style fidelity.

Phase 4: Automated QC Validation (Cost: ~$0)

Test vision-model-based quality scoring. Validate against human judgment (>80% agreement target).

Phase	Estimated Cost
Phase 1	$0.50-$1.00
Phase 2	$2.00-$5.00
Phase 3	$2.00-$3.00
Phase 4	$0.00
Total	$4.50-$9.00

Estimated generations: 50-70 images total across all phases.

8. Proposed Style Profile Schema

A 10-category JSON schema covering medium, color palette, line work, lighting, texture, character rendering, background treatment, mood/atmosphere, negative constraints, and prompt templates. The schema supports cross-model adaptation (NB2 and GPT-Image-2) with model-specific formatting differences.

1. Prompt Structure Best Practices

Subject-First Ordering

NB2 weights the start of the prompt heavily. Put your subject first, then style/environment details. Starting with style descriptors produces softer, less focused results.

Do: "A small brown dog wearing a red cape, standing on a hilltop at sunset, digital illustration style"

Don't: "Digital illustration style, sunset lighting, a small brown dog on a hilltop"

Natural Language Over Tag Soup

NB2 is a thinking model that understands intent, physics, and composition. Write prompts as if briefing a human illustrator. Avoid comma-separated tag lists.

JSON-Structured Prompts

Community benchmarks show JSON-structured prompts improve consistency by ~25% over plain text. Structure with labeled fields:

{
  "subject": "A small brown dog named Ollie wearing a red cape",
  "pose": "standing heroically, looking to the right",
  "environment": "grassy hilltop at golden hour",
  "style": "soft watercolor children's book illustration",
  "lighting": "warm golden backlight with soft shadows",
  "camera": "low angle, medium shot"
}

Regional Prompts

For different content in different parts of an image, divide the canvas into regions with a prompt for each. Useful for complex scenes with foreground characters and detailed backgrounds.

2. Reference Image Strategy

Multi-Reference Conditioning

NB2 supports up to 10 reference images simultaneously. When using uploaded images, clearly define each role: "Use Image A for the character's pose," "Use Image B for the art style," "Use Image C for the background environment."

Character Sheets (360-Degree)

The most reliable consistency method:

• Generate 2-3 images of your character within a single frame showing multiple angles (front, left, right, back)
• Use this sheet as reference input for all subsequent generations
• The character sheet becomes the "ingredient" that maintains identity across scenes

Image Search Grounding (NB2 Exclusive)

NB2 can retrieve real-world reference images via Google Search during generation. This is unique to NB2 (not available in Pro) and helps ground outputs in real-world visual context.

3. Temperature and Parameter Optimization

Temperature Settings

• Range: 0 to 2 (default: 1)
• Low (0-0.5): More deterministic, predictable. Best for character sheets and batch generation.
• Medium (0.8-1.2): Balanced creativity and adherence. Good default for most illustration work.
• High (1.5-2.0): More diverse, surprising. Good for exploration, less predictable.

Thinking Mode (NB2 Exclusive)

• Minimal: Fastest generation, good for rapid iteration
• High: Best quality, slower. Use for final/hero images
• Dynamic: Model decides the appropriate level per request

Quality Keywords and Seed Control

Include terms like "HD," "4K," or "HDR" for better clarity. Start every prompt with a fixed seed for reproducibility. Combine with weighted descriptors like "(same face:1.3)" to enforce consistency.

4. Character Consistency Approaches

• Naming characters: Assign distinct names; model tracks named entities better across multi-turn sessions
• Consistency anchors: "same character," "maintain facial features," "preserve proportions," "consistent with previous image"
• Multi-turn continuity: Remind NB2 to "use same character identity as the previous image"
• Style locking: Prompt "same art style" across sequences. Define style guide once, reference in every prompt
• Iterative editing: If 80% correct, edit don't regenerate. NB2 excels at conversational edits
• Negative prompts: Combine positive traits with negatives to prevent common failure modes

5. NB2 vs Nano Banana Pro

Feature	Nano Banana 2	Nano Banana Pro
Architecture	Gemini 3.1 Flash	Gemini 3 Pro
Speed	4-6 sec at 1K	10-20 sec at 1K
Quality	~95% of Pro	Best absolute quality
Textures/Lighting	Very good	Superior
Unique Features	Image Search Grounding, Thinking Mode	N/A
Cost	Significantly cheaper	Higher per image
Best For	Speed, batch generation, stylized work	Precision, identity preservation

6. NB2 vs GPT-Image-2

Capability	NB2	GPT-Image-2
Photorealism	Superior	Strong
Text/Typography	Good	Superior (best in class)
Anime/Illustration	Superior	Good
Structural Control	Moderate	Superior
Speed	Faster	Slower
Batch Consistency	Good	Superior
Cinematic Lighting	Superior	Good
Cost (API)	$60/M output tokens	$30/M output tokens

Choose NB2 for: stylized illustration, photorealism, cinematic scenes, rapid iteration, anime. Choose GPT-Image-2 for: text-heavy designs, precise layouts, batch production, structural accuracy.

7. Community Tips

• Prompt like a Creative Director — describe the scene, mood, intent
• Use the 80/20 edit rule — if mostly right, edit don't regenerate
• Build character sheets first — invest upfront, save time across the project
• Test with 1-2 images before scaling
• Use specific color palettes: "ochre, sap green, dusty blue" beats "warm earth tones"
• Limit core colors to 3-5 plus occasional accents
• Lock your style guide in writing — define and reuse

1. Why Multi-Model?

Single-model approaches force compromise. Every model has blind spots. The 2026 best practice is to route each task to the model that handles it best.

For children's book illustration: GPT-Image-2 excels at structural accuracy, text rendering, and batch consistency. NB2 excels at stylized illustration, cinematic lighting, and rapid iteration. Combining them produces results neither can achieve alone.

2. Model Strengths Comparison

Capability	NB2	GPT-Image-2	Best Choice
Photorealism	Excellent	Very Good	NB2
Stylized Illustration	Excellent	Good	NB2
Text/Typography	Good	Best in class	GPT-Image-2
Precise Layout	Moderate	Excellent	GPT-Image-2
Cinematic Lighting	Superior	Good	NB2
Batch Consistency	Good	Excellent	GPT-Image-2
Character Identity	Good (improving)	Strong	GPT-Image-2
Speed	4-6 sec/image	Slower	NB2
Cost	$60/M output	$30/M output	GPT-Image-2
Conversational Editing	Excellent	Good	NB2
Multi-Reference	Up to 10	Limited	NB2

3. Pipeline Architectures

Architecture A: GPT-Image-2 for Composition, NB2 for Style

Flow: GPT-Image-2 (structure) → NB2 (style transfer)

• GPT-Image-2 generates the base composition with precise character placement and layout
• NB2 receives the output as reference and applies stylistic treatment: watercolor textures, warm lighting, soft edges
• Best for scenes with specific spatial requirements

Architecture B: NB2 for Exploration, GPT-Image-2 for Production

Flow: NB2 (ideation) → GPT-Image-2 (final production)

• NB2 generates rapid concept variations at 3-5x the speed
• Select best concepts from NB2 explorations
• GPT-Image-2 produces final version with batch consistency
• Best when still figuring out composition/mood

Architecture C: Parallel Generation with Best-Of Selection

Both models generate independently on the same prompt. Compare and select the best per scene. Best when you have prompt budget and want maximum quality.

Architecture D: Layered Composition

• NB2 generates backgrounds/environments (superior cinematic lighting)
• GPT-Image-2 generates characters (better identity preservation)
• Composite layers, then final pass for unified style
• Best for complex scenes where both background quality and character accuracy are critical

4. Ollie & Dot Recommended Pipeline

Phase 1 — Character Design (GPT-Image-2): Generate definitive character sheets. Lock proportions, colors, features. GPT-Image-2's batch consistency ensures reliable reference sheets.

Phase 2 — Scene Exploration (NB2): Rapid scene iteration with character sheet references. Explore compositions and lighting at 3-5x speed. Multi-reference conditioning with character + environment + style references. Generate 4-6 variants per scene.

Phase 3 — Final Production (Model-Dependent): Text scenes → GPT-Image-2. Atmospheric/stylistic scenes → NB2. Precise character placement → GPT-Image-2. Wide environmental scenes → NB2.

Phase 4 — Consistency Pass: Review all finals. Use NB2 conversational editing for minor adjustments. Use GPT-Image-2 for structural corrections.

5. Cost Optimization

• Cheap models for iteration (NB2 for rapid exploration)
• Expensive models for delivery (GPT-Image-2 for finals where its strengths matter)
• Don't iterate at scale on the expensive model — validate with 1-2 test images first

6. Pipeline Orchestration Tools

• ComfyUI: Open-source visual workflow builder. Chain multiple models, apply controls, automate batches.
• MindStudio: Visual builder for chaining 200+ models. Manages credit costs and routing.
• Custom Python pipeline: Maximum control. Call GPT-Image-2 API for base, pass to NB2 via Gemini API as reference, automated QC via Gemini text model.

7. Character Consistency Across Models

• Shared character sheet as anchor for both models
• Detailed character description block prepended to every prompt
• Style guide document with palette, lighting, texture rules
• Post-generation review via Gemini text model to flag inconsistencies
• One model for character, one for environment to avoid cross-model drift

8. Key Takeaways

• Don't pick one model — route by task
• Use NB2 for speed/exploration, GPT-Image-2 for precision/production
• Character sheets are the bridge between models
• NB2's conversational editing is the best finisher
• Automate the pipeline with ComfyUI or custom scripts
• Always test with 1-2 images before scaling

1. Upscaling GPT-Image-2 Outputs

Pro Upscaler Overview

Leonardo AI's Pro Upscaler is purpose-built for AI-generated images, capable of scaling up to 105 megapixels. Key differentiator: while most upscalers are trained on real photographs, Leonardo's understands synthetic grain and AI textures, cleaning them rather than amplifying them.

Performance

• 1536×1024 to print resolution: At 2x (3072×2048), results are highly usable. At 4x (6144×4096), minor artifacts but still outperforms traditional upscalers.
• AI-generated image handling: Specifically handles noise patterns and stylistic grain from AI models. Standard upscalers misread these as defects.
• Structural integrity: Maintained under intense magnification — important for character consistency.

Verdict: Strong candidate. 105MP ceiling far exceeds print needs (12×8 inch spread at 300 DPI = ~8.6MP).

2. Outpainting (3:2 to 2:1 Aspect Ratio)

Leonardo's Canvas mode includes Inpainting/Outpainting tools for extending images beyond original boundaries.

How It Works

• Place generation box overlapping ~60% existing image, ~40% empty canvas
• Keep original prompt or leave blank
• AI extends the scene into empty space

Suitability for 3:2 to 2:1

Converting 1536×1024 (3:2) to 2048×1024 (2:1) requires ~512px per side extension — moderate and within the tool's sweet spot. Works best when extending backgrounds/environments rather than adding character elements.

3. Other Post-Processing

• Alchemy v4: Proprietary pipeline with Hyper-Realism and Abstract Concept modes. Abstract mode more relevant for Ollie & Dot.
• Background removal: Available as post-processing action
• Image-to-image: Style refinement passes
• Canvas editor: Manual touch-up and detail refinement

4. API and Pricing

Plan	Price	Tokens/Month	API Access
Free	$0	150/day	No
Apprentice	$12/mo	8,500	Limited
Artisan	$30/mo	25,000	Yes
Maestro	$48-60/mo	60,000	Full + priority

Token costs: Standard generation 2-25 tokens, upscaling ~2x generation cost. Cost optimization: generate low-res, only upscale selections (saves 40-50%). Artisan plan handles ~500-1,000 upscale operations. $5 free API credit for testing.

5. Comparison with Other Upscalers

Tool	Best For	API	Pricing	AI-Image Aware
Leonardo Pro	AI-generated images	Yes	Token-based	Yes (trained on synthetic)
Real-ESRGAN	General-purpose, free	Self-hosted	Free	No
Topaz Photo AI	Photographers	No	$199 one-time	No
Magnific AI	Creative upscaling	Yes	Subscription	Partial (over-hallucination risk)

Recommendation: Leonardo Pro is the strongest candidate: purpose-built for AI images, API available, token-based pricing, outpainting and upscaling in one platform. Runner-up: Real-ESRGAN as free fallback. Avoid: Magnific — tendency to hallucinate new details could compromise illustration consistency.

1. Latest Model: V8.1 Alpha

Version Timeline

Version	Release	Status
V6	Late 2024	Legacy
V7	2025	Stable, was default
V8.0 Alpha	March 17, 2026	Superseded
V8.1 Alpha	April 14, 2026	Current latest

V7 Improvements Over V6

• Much smarter text prompt interpretation
• Significantly better coherence for hands, bodies, objects
• Draft Mode: half cost, 10x speed for iteration
• Omni Reference replaced --cref with a dedicated reference tab

V8 Improvements Over V7

• ~5x faster rendering
• Native 2K resolution with --hd flag
• Sharper photorealism with physically grounded lighting
• Much better text rendering in images
• More accurate handling of complex prompts

V8.1 Improvements Over V8.0

• HD now default (native 2K without --hd flag)
• HD mode 3x faster and 3x cheaper
• Standard resolution 50% faster, 25% cheaper
• Returns to V7's consistent aesthetic, fixing V8.0's over-polished look
• Super-stable moodboards and srefs — the major V8.0 weak point fixed

2. Artifact and Morphing Status

Hands/bodies/objects dramatically better from V7 onward. V8 further improves physical coherence. V8.1 corrects V8.0's over-processed default look. Some legacy features still missing (image prompting, in-painting). Reduced ability for abstract/surreal compositions compared to V7.

3. Style References (--sref)

How Style References Work

The --sref parameter extracts aesthetic qualities from reference images: color palette, lighting, texture treatment, composition tendencies. Unlike regular image prompts, it focuses on style elements rather than content.

SREF System

• Versioned system with 6 sub-versions via --sv
• --sv 6: Default (latest)
• V8.1 srefs are "super-stable" — major V8.0 fix
• Backward compatible with V7 profiles
• Style codes are model-version dependent

Style Creator Tool

New in V8: shape aesthetics visually. Enter a prompt, get aesthetic options, select preferences, receive a unique reusable style code. Web-only (midjourney.com), not Discord.

Moodboards

Curate collections of images to define persistent style. Upload images, name and save for reuse, apply across generations.

4. Cross-Model Style Transfer

Recommended workflow:

• Generate style exemplars in Midjourney using crafted prompts + style codes
• Feed those images as references to NB2 or GPT-Image-2
• Or: use a vision model (GPT-4o, Gemini) to analyze the Midjourney outputs and produce structured style descriptions
• Convert descriptions into prompt components for other models
• One-time investment, reusable indefinitely

5. API Access

No public API. Enterprise dashboard users only, must apply. Unofficial APIs (PiAPI, APIFRAME, ImagineAPI) violate ToS and risk account bans. Midjourney cannot be reliably integrated into an automated pipeline.

Best used as a manual tool for style exploration and reference generation, not batch production.

6. Strategic Recommendation

Use For (Manual, Occasional)

• Style exploration via Style Creator and moodboards
• Reference image generation (high-quality style exemplars)
• Style extraction via vision model analysis
• Quality benchmarking against GPT-Image-2 outputs

Do NOT Use For

• Batch production (no reliable API)
• Automated workflows (too fragile)
• Character consistency at scale (better via GPT-Image-2)

7. Pricing

Plan	Price	Images/Month
Basic	$10/mo	~200
Standard	$30/mo	~900
Pro	$60/mo	~1,800 fast

For our use case (style references, not production), Basic at $10/mo would suffice.

8. Summary Scorecard

Capability	Score
Image quality (V8.1)	9/10
Style consistency (sref/moodboards)	8/10
Artifact issues	8/10
API for pipeline	2/10
Cross-model style transfer	7/10
Cost efficiency	8/10
Overall pipeline fit	6/10 (style reference tool only)

1. Vision-Extracted Style Profiles: Best Practices

What Is a Vision-Extracted Style Profile?

A structured text description of an image's visual style, produced by sending approved reference images to a vision model (Gemini, Claude, GPT-4o) and asking it to analyze and catalog style attributes. The output is a reusable text block for generation prompts.

How Many Images to Analyze

Count	Pros	Cons
1 image	Simple, fast	Overfits to one scene's lighting/composition
2 images	Better than one	Still too few to separate scene from style
3-5 images	Sweet spot: enough diversity to find common thread	Recommended
6-10 images	Very thorough	Diminishing returns, may average too aggressively
10+ images	Comprehensive	Descriptions become generic

Key principle: Images should be diverse in SCENE but consistent in STYLE. Forces the vision model to identify style (what stays the same) vs. content (what changes).

The Extraction Prompt

A 10-category extraction covering: (1) Rendering Technique, (2) Color Palette with hex codes, (3) Line Work, (4) Lighting and Shadow with contrast ratios, (5) Texture and Surface, (6) Character Rendering proportions, (7) Background Treatment, (8) Composition and Framing, (9) What This Style Is NOT, (10) One-Sentence Test. Forces specificity, anchors to observable details, includes negatives.

Structuring Output

• Run extraction on each of 3-5 gold standard images separately
• Per-image profiles: keep each one
• Intersection profile: attributes in ALL profiles = core style
• Conflict resolution: note ranges, decide canonical values
• Final canonical profile: merged, validated, human-approved

JSON Style Guides

The 2025-2026 standard: 70-78% of professional AI creators use JSON-formatted style guides. Structured prompts improve accuracy by 60-80% for complex scenes. Eliminates ambiguity, machine-readable, reproducible. GPT-Image-2 responds especially well to structured input.

2. Replicating Midjourney --sref

What --sref Actually Does

Midjourney uses a dedicated internal style encoder (likely CLIP-based) to extract style features from a reference image and inject them as a separate conditioning signal. Style is kept separate from content.

No Direct Equivalent Exists

Feature	Midjourney --sref	NB2/GPT-Image-2
Style encoding	Dedicated encoder, separate channel	General vision encoder, mixed with content
Style vs content separation	Explicit	Implicit (must be prompted)
Reproducibility	Deterministic codes	No equivalent
Weight control	--sw 0-1000	No direct param

Practical Approximations

• Approach 1: Reference images + explicit style instructions (available now, works with our models)
• Approach 2: IP-Adapter via ComfyUI (closest open-source, requires Stable Diffusion)
• Approach 3: exactly.ai (managed LoRA, tied to their platform)
• Approach 4: Custom LoRA fine-tuning (SD only)

3. Academic Techniques

Technique	Can Use with NB2?	Can Use with GPT-Image-2?	Post-Processing?	Open Source?
StyleTokenizer (ECCV 2024)	No	No	Potentially	Yes
StyleBrush (2024)	No	No	Yes (promising)	Yes
InstantStyle (2024)	No	No	Yes (via ComfyUI)	Yes
IP-Adapter (2023-2025)	No	No	Yes (via ComfyUI)	Yes
Z-STAR (CVPR 2024)	No	No	Potentially	Yes

Bottom line: None can be plugged directly into NB2 or GPT-Image-2 APIs. All require Stable Diffusion. However, they CAN be used as post-processing layers, and their design principles (separate style encoding, targeted injection, JSON prompts) inform better prompt writing.

4. Style Layering as Post-Processing

Generate a base image with GPT-Image-2 (composition, character consistency) then run through a style transfer layer for the specific aesthetic.

Options Evaluated

• Option A: Neural Style Transfer — Not recommended. Produces "painterly filter" effect, looks obviously processed.
• Option B: IP-Adapter via ComfyUI — Recommended investigation. Best balance of quality, control, accessibility. Uses ControlNet for composition preservation.
• Option C: StyleBrush — Promising if accessible. Adjustable style strength, better separation than IP-Adapter.
• Option D: Img2Img with low denoising — Quick and dirty. Good for experimentation, risky for production.

Current recommendation: Don't add post-processing yet. Focus on making GPT-Image-2/NB2 produce the right style natively. Post-processing is the fallback. If needed later, start with Option B.

5. Why It Hasn't Worked Before

Problem 1: Descriptions Too Generic

"Colorful watercolor children's book illustration with warm tones" describes half of all children's books. Fix: Use the detailed extraction prompt with hex codes, ratios, named techniques.

Problem 2: Not Enough Reference Images

Single image = profile captures scene-specific attributes instead of style attributes. Fix: Use 3-5 images with different scenes but same style. Intersection = style.

Problem 3: Wrong Prompt Placement

Style buried in middle/end of prompt. Fix: Style FIRST, clear section markers, repeat critical elements at end.

Problem 4: Content Described, Not Style

Vision model defaults to "what is in the image." Fix: Explicitly say "Do NOT describe what is depicted. ONLY describe HOW it is depicted."

Problem 5: Missing the "Ineffable" Quality

Measurable attributes correct but images still look "off." Fix: Always pair text with reference images. Use one-sentence descriptor for emergent quality. Use negative constraints. Iterate.

Problem 6: Model Cannot Produce the Style

Model's style distribution bias. Fix: Try the other model, use fine-tuned SD, apply post-processing, or adjust target style.

6. Practical Workflow

STEP 1: STYLE EXTRACTION (One-time, ~30 min)
- User provides 3-5 gold standard images
- Run extraction prompt on each via Gemini/Claude
- Aggregate into canonical JSON style profile
- User reviews and approves

STEP 2: PROMPT TEMPLATE CREATION (~15 min)
- Build three-block template: [STYLE] + [CHARACTER] + [SCENE]
- Test with 2-3 generations
- Adjust style profile based on test results

STEP 3: BATCH GENERATION (Per illustration, ~5-10 min)
- Fill in [SCENE] block
- Select 2-3 reference images
- Generate 3-5 candidates
- Automated QC: vision model scores consistency
- Present top 2-3 to user

STEP 4: ITERATIVE REFINEMENT (As needed)
- Fix one attribute at a time
- Re-run with updated profile

Cost Estimate

Step	Images	Est. Cost
Style extraction (vision, text only)	0	$0
Template validation	3-5	$0.30-$1.00
Per illustration (3-5 candidates)	3-5	$0.30-$1.00
Automated QC (vision, text only)	0	$0
Full 24-page book (12 spreads)	36-60	$3.60-$12.00

7. Key Principles

• Start with extraction, not generation
• Reference images are not optional — always include 2-3
• The profile is a living document — refine after test generations
• Separate style from content in prompts with clear section markers
• JSON > prose for style descriptions
• Iterate on one attribute at a time

1. NBP Model Overview

Nano Banana Pro is Google DeepMind's premium image generation model, built on Gemini 3 Pro. It is a "Thinking" model that understands intent, physics, and composition rather than simply matching keywords.

Feature	NBP (Pro)	NB2 (Flash)
Quality	Highest quality, richer textures, more natural lighting	~95% of Pro quality
Speed	Slower (standard generation time)	3-5x faster than Pro
Cost	$0.134/image (Google AI Studio)	Free tier available
Text Rendering	Best-in-class legible text	Improved but inferior to Pro
Lighting/Composition	Superior complex lighting, depth, perspective	Good but less nuanced
Reasoning	Advanced semantic understanding of spatial relationships	Fast but less compositional reasoning

2. Prompt Structure

NBP processes prompts as semantic structures, reasoning about subject relationships, spatial positions, and instructional requirements before generating.

Optimal Ordering (Most Important First)

• Reference image role assignment
• Core subject/character
• Action/composition
• Location/environment
• Style/medium
• Constraints (aspect ratio, what to avoid)

Anti-Patterns

• Tag soups: "dog, park, 4k, realistic, beautiful"
• Abstract quality language: "museum quality," "impossible beauty"
• Contradictory specifications in the same prompt
• Exceeding 10 numbered changes per edit prompt
• Rewriting prompts from scratch with bloated character bibles

3. Reference Image Strategy

Supports up to 14 reference images. Practical optimum: 5-6 for best consistency. Above 8-10, consistency degrades.

Slot	Type	Purpose
1	Style Reference	Visual quality, lighting, material feel
2	Composition Reference	Scene layout, camera angle
3-4	Character Reference(s)	Identity lock for recurring characters
5	Quality Anchor	Best previous generation for this scene
6	Brand Bible (optional)	Color palette, typography rules

Critical: Every prompt must open with explicit role assignment. Character refs should include 3 views (frontal, 45-degree, side). Minimum 1024x1024 resolution.

4. Character Consistency (Three Pillars)

Foundation Image Method

• Create one strong foundation image of the character
• Reference it in every subsequent generation
• Change only camera angles and context

Trait Locking

• Use identical descriptive words in every generation
• "Emerald eyes" in panel one = "emerald eyes" in all panels (never switch to "green eyes")
• Lock facial structure markers: eyes, jawline, distinct marks

Session-Based Memory

• Define character comprehensively in first prompt of session
• Subsequent prompts only need action or emotion
• Supports up to 5 consistent characters in a single workflow

5. Temperature Strategy

Goal	Temperature	Notes
Maximum consistency	0.8-1.0	Best for character-critical scenes
Balanced (default)	1.0	Gemini 3 default
Creative variation	1.0-1.4	Good for exploring compositions
Too random (avoid)	>1.5	Degrades quality and consistency

Community workflow: Generate 4 variants at temperatures 0.85, 0.95, 1.05, 1.15. Pick best. Edit pass at 0.9. Final polish at 0.85.

6. Edit vs Fresh Generation

Edit when:

• Image is 80%+ correct
• Specific changes to composition, color, or character details needed
• Character pose or expression adjustment in otherwise good scene

Generate fresh when:

• Starting completely new scene
• Previous attempts accumulated too many edit artifacts
• After 3+ failed edit attempts (quality degrades)

Key rules: Never chain edits. Always go back to best original source. Max 10 changes per edit. Short, numbered, concrete visual instructions.

7. Known Limitations

• Hands/fingers: Still commonly distorted in complex poses
• Small faces: Struggles with accurate rendering at distance
• Facial hallucination: May alter features after edits
• 4K timeout failures: High failure rate (TPU resource allocation)
• Quality degradation: Reports of silent model version switches (April 2026)
• IMAGE_SAFETY: False positives on legitimate content (Jan 2026)
• Edit quality decay: Noticeable after 3+ sequential edits

Ollie & Dot Workarounds

• Always state aspect ratio explicitly: "Full bleed 16:9"
• Limit edits to 1-2 passes max, then regenerate fresh
• Use NB2 for rapid exploration, NBP only for final refinement
• Keep character detail in BOTH text and reference images
• If pearl iridescence is lost, regenerate (cheaper than editing back in)

1. Model Specs

Spec	NBP (Pro)	NB2 (Flash)
Base Model	Gemini 3 Pro Image	Gemini 3.1 Flash Image
Max Resolution	4K (4096x4096)	4K (4096x4096)
Max Reference Images	14 (optimal: 6)	10 (optimal: 5-6)
Max Character Tracking	5 simultaneous	3-4 simultaneous
Generation Speed (1K)	10-20 sec	4-6 sec
Cost (2K)	~$0.134/image	~$0.08/image
Text Rendering	Superior	Good (can blur small text)

2. Quality Comparison: Mixed-Media Collage

Criterion	Winner	Notes
Torn paper fiber texture	NBP	Richer detail, more natural fiber separation
Gold foil metallic rendering	NBP	Superior metallic sheen, realistic stipple
Gouache paint texture	NBP	More natural brushwork, visible paint body
Watercolor transparency	NBP	Better wet-on-wet bleeding, truer pigment mixing
Photo-real caustics	NBP	Superior photorealistic integration
Overall material fidelity	NBP	NB2 is 90-95% of Pro quality

3. Character Consistency

Ollie (Cartoon Octopus)

• NBP: Reliable tentacle count, consistent gold foil freckle-dots, stable eye proportions
• NB2: Occasionally adds/removes tentacle, sometimes renders gold as painted yellow, slight drift without strong reference

Dot (Pearl)

• NBP: Consistent shape/glow, reliably respects "no mouth," rich nacre iridescence
• NB2: Glow intensity varies, occasionally adds mouth-like marks, flatter rainbow effect

Verdict: NBP is superior for character consistency, particularly for maintaining specific material qualities (gold foil, nacre) that define the characters.

4. Style Drift Rates

Model	Drift Pattern	Frequency
NBP	Over-refinement (too polished vs handmade)	~10%
NBP	Unwanted photorealistic elements	~5%
NB2	Smoothing/simplifying textural detail	~20%
NB2	Treating collage as filter rather than physical media	~15%
NB2	Defaulting to more "illustrated" vs "mixed-media"	~10%

5. Recommended Hybrid Pipeline

Phase	Model	Purpose	Cost
Exploration	NB2	Rapid composition/layout testing	~$0.06/img
Refinement	NB2	Prompt tuning before committing to Pro	~$0.06/img
Finals	NBP	Hero images with full reference stack	~$0.134/img
Edits/Fixes	NBP	Targeted edits on near-final images	~$0.134/img

Three-model triangle: GPT-Image-2 (structural accuracy, text-heavy spreads) + NB2 (rapid iteration, prompt validation) + NBP (final quality, material fidelity, metallic/iridescent rendering).

6. Budget Estimate for Draft 3

Scenario	Images	Cost
NB2-only	80	~$4.80
NBP-only	80	~$10.72
Hybrid (NB2 explore + NBP finals)	80 NB2 + 40 NBP	~$10.16
Hybrid optimized	40 NB2 + 20 NBP finals + 20 NBP edits	~$7.76

Recommended total budget: $12-18 for full Draft 3 hybrid approach (covers exploration, refinement, finals, and edits with margin for iteration).

7. Generation-Rules Skill Update Needed

Current rule "USE ONLY NB2" should be updated to:

• Primary (NB2): All exploration, composition testing, prompt validation
• Secondary (NBP): Final hero images, character-critical close-ups, material-heavy renders
• Rule: Start every spread with NB2 exploration. Only escalate to NBP when composition is approved by creator. Never use NBP for "testing." Track NBP spend per session.

1. API Access

Leonardo AI has a fully public REST API at https://cloud.leonardo.ai/api/rest/v1/. Authentication via Bearer token. New API accounts receive $5 in free credit.

Operation	Method	Endpoint
Image Generation	POST	/generations
Upload Init Image	POST	/init-image
Universal Upscaler	POST	/variations/universal-upscaler
Create Dataset	POST	/datasets
Train Element (LoRA)	POST	/elements

2. Style Reference via ControlNet

Leonardo's direct equivalent to Midjourney --sref. Upload a reference image and include it as a controlnet with the Style Reference preprocessor ID. The system extracts visual style (not content/composition) and applies it to new generations.

Preprocessor IDs by Model

Model	Preprocessor ID
SDXL	67
Phoenix	166
Flux Dev	299
Flux Schnell	298
Lucid Realism	431

Strength Types: Low, Mid, High, Ultra, Max

Multiple Style References: Combine multiple style reference images with an influence parameter (0-1) to blend different style aspects (e.g., torn paper texture at 0.6 + gold foil gouache at 0.4).

3. Custom LoRA Training (Elements)

For deeper style consistency, train custom LoRA models ("Elements") on a set of reference images. The style becomes baked into a reusable model weight.

Process

• Prepare 10-50 reference images capturing the target style
• Create a dataset via API (POST /datasets)
• Upload images to the dataset
• Train an Element with a trigger word (POST /elements)
• Use the Element in all subsequent generations via userElements array

Training Specs

• Minimum 10 images, maximum 50 (Flux Dev) or 40 (SDXL)
• Training time: minutes to hours depending on dataset size
• Multiple Elements can be blended at inference with adjustable weights
• Trigger word activates the style

Limitation: Element training currently supports SDXL-based models. Compatibility with Phoenix and Flux models for custom Elements should be verified.

4. Comparison to Midjourney --sref

Feature	Midjourney --sref	Leonardo Style Ref	Leonardo Elements (LoRA)
Style from image	Yes	Yes (controlnet)	N/A (training-based)
API access	No (Discord only)	Yes (REST API)	Yes (REST API)
Strength control	--sw 0-1000	Low/Mid/High/Ultra/Max	weight 0-1.0
Multiple refs	Yes	Yes (multiple controlnets)	Yes (blend Elements)
Reusable style code	Yes (style code)	No persistent code	Yes (trained Element ID)
Training required	No	No	Yes (10-50 images)
Batch consistency	High	Moderate-High	High

Key advantage: Leonardo's LoRA training on 10-50 reference images can learn nuances more deeply than a single reference image can convey. This dual approach (Element + Style Reference) would likely outperform Midjourney --sref for batch style consistency.

5. Pricing

API Credits	Basic ($)	Standard ($)	Pro ($)
5,000	$15.00	$11.00	$8.00
10,000	$30.00	$22.00	$16.00
25,000	$75.00	$55.00	$40.00
50,000	$150.00	$110.00	$80.00

Per-Image Cost Estimates

Scenario	Credits/image	Cost at Pro	Cost at Basic
Standard gen (SDXL)	~5	~$0.008	~$0.015
Alchemy-enhanced	~12	~$0.019	~$0.036
Phoenix generation	~8-15	~$0.013-$0.024	~$0.024-$0.045
With style reference	+2-5 extra	+$0.003-$0.008	+$0.006-$0.015

Comparison to Other Models

Model	Cost/image	Notes
Leonardo (Pro, Phoenix + sref)	~$0.016-$0.032	Cheapest option
NB2 (Replicate)	$0.04-$0.08	Fixed per generation
NBP (Google AI Studio)	$0.08-$0.134	Fixed per generation
GPT-Image-2	~$0.02-$0.08	Varies by resolution

6. Universal Upscaler

Endpoint: POST /variations/universal-upscaler

Ultra Mode (Recommended for AI Art)

• ultraUpscaleStyle: ARTISTIC (for illustrations) or REALISTIC
• creativityStrength: How much the AI adds detail (1-10)
• detailContrast: HDR-like detail contrast (1-10)
• similarity: Structural similarity to original (1-10)
• upscaleMultiplier: Scale factor
• Maximum output: 20MP (Ultra mode can reach up to 8K)

AI-native training handles synthetic grain better than general-purpose upscalers like Real-ESRGAN.

7. Outpainting Gap

No dedicated API endpoint confirmed. Canvas outpainting appears to be a web UI feature only. For programmatic outpainting, workarounds include:

• Use the web Canvas tool manually
• Use image-to-image generation with a padded canvas as the init image
• Use a different tool for outpainting (Stability AI, RunwayML)

This is a gap in Leonardo's API for the Ollie & Dot pipeline, where 3:2 to 2:1 aspect ratio conversion is needed.

8. Phoenix 2.0 Model

• Model ID: de7d3faf-762f-48e0-b3b7-9d0ac3a3fcf3
• Built specifically for prompt adherence
• Phoenix 2.0: improved style transfer fidelity, better preservation of line weight, color palette, compositional habits
• Competitive with Midjourney v7 for photorealistic output
• Rewards precise, descriptive prompting over vague/artistic language
• Phoenix-specific params: contrast (3/3.5/4), alchemy: true, ultra: true
• Leans photorealistic by default — requires more prompting effort for stylized/mixed-media looks

9. Recommended Strategy

Phase 1: Quick Test (Free Credits)

• Upload 1-2 best existing Ollie & Dot illustrations as style reference images
• Generate 4-8 test images using Style Reference controlnet (preprocessorId 166) with Phoenix
• Test at different strength levels (Mid, High, Ultra)
• Evaluate whether zero-shot style reference captures the mixed-media complexity

Phase 2: Element Training (If Needed)

• Select 15-30 of the best existing illustrations that exemplify the target style
• Create a dataset and train an Element with trigger word
• Generate test images using the trained Element
• Compare quality to Phase 1 results

Phase 3: Production Pipeline

Style Reference Image + Trained Element + Precise Prompt → Leonardo Phoenix API → Raw Generation (1024x768) → Universal Upscaler (ARTISTIC mode) → Print-Ready Illustration (4K+)

10. Gaps and Limitations

• No outpainting API — would need a separate tool or manual work
• No persistent style code like Midjourney --sref generates
• Element training is SDXL-based — may not work with Phoenix/Flux directly
• Less "artistic default" than Midjourney — requires more precise prompting
• Phoenix 2.0 leans photorealistic, which may need extra effort for mixed-media

1. The Core Challenge

The Ollie and Dot style is fundamentally a compositing problem, not a single-generation problem. The style requires: photorealistic underwater environments (Hawaiian reef, caustics, light physics), torn paper collage elements with visible fiber texture, gold foil metallic stipple dots, gouache-painted cartoon characters (baby octopus + pearl), watercolor paper base texture, bioluminescent glow effects, and nacre/pearl iridescence. No single model excels at all of these simultaneously.

2. Model-by-Model Analysis

Flux 2 Pro (Black Forest Labs)

Best-in-class photorealism in the open-weight space. 4MP output, strongest underwater environments of any model evaluated. Available as Pro ($0.03–0.055/image) and Dev ($0.025/image). Full LoRA support on Dev/Klein for custom style training. Flux Kontext enables instruction-based editing with character identity preservation across scenes.

Ideogram 3.0

Style References (up to 3 images) + reusable Style Codes for batch consistency. 4.3 billion style presets. Character Reference for facial/trait consistency. Best-in-class text rendering. Turbo: $0.03/image, Quality: $0.09/image. No LoRA training — inference-time only. May not preserve photo-vs-illustration contrast.

Recraft V3

20B parameter design-focused model. Style Lock constrains line weight, color application, detail level, proportions, and rendering approach. Native SVG output. Character consistency still developing (feature request on roadmap). Raster: $0.04/image, Vector: $0.08/image. Design-grade illustration, not photorealism.

Stable Diffusion 3.5

Most mature LoRA ecosystem. IP-Adapter + ControlNet = most fine-grained control available. Canny edge, depth maps, blur ControlNets. Can combine Style IP-Adapter + Depth ControlNet + Character LoRA. Lower base quality than Flux 2, but unmatched control mechanisms. $0.065/image (API) or free self-hosted.

Leonardo Phoenix 2.0

Style/Character/Content Reference via ControlNet preprocessors. Custom LoRA training ("Elements") on 10–50 images. Very affordable ($0.007–0.02/image). Leans photorealistic by default. Will homogenize mixed-media layers. Best role: cheap iteration and concept testing.

InstantCharacter (Tencent/Hunyuan)

Tuning-free character consistency from a single reference image. Built on scalable diffusion transformer, trained on 10M+ character samples. Available on FAL.ai. Designed primarily for humanoid characters — performance with baby octopus cartoon character is untested.

3. Style Reference Mechanisms

Model	Mechanism	Type
Leonardo Phoenix	Style/Character/Content Reference (3 preprocessors)	Inference-time
Leonardo Elements	LoRA training on 10–50 images	Trained
Flux Kontext	Instruction-based editing	Inference-time
Flux LoRA	Custom fine-tuning	Trained
Ideogram 3.0	Style Reference images + Style Codes	Inference-time
Recraft V3	Style Lock	Inference-time
SD3.5	IP-Adapter + ControlNet	Inference-time
NB2/NBP	Subject consistency (up to 5 characters)	Inference-time
InstantCharacter	Single-image character preservation	Inference-time

4. Pricing Comparison

Model/Platform	Cost/Image	Best For
Leonardo (Apprentice)	~$0.007	Cheapest, token-based
Flux Kontext Dev (FAL)	$0.02	Style transfer, editing
Flux 2 Dev (FAL)	~$0.025	LoRA-based generation
Ideogram 3.0 Turbo	$0.03	Fast style-consistent
Flux 2 Pro (BFL)	$0.03–0.055	Top photorealism
Recraft V3 (raster)	$0.04	Design illustration
GPT Image 2	$0.04	Composition + text
NB2/NBP (Google AI Studio)	$0.134 (paid) / free	Current primary model
Ideogram 3.0 Quality	$0.09	Maximum quality

5. Recommended Pipeline Architecture

[Flux 2 Pro]                    [NB2/NBP]              [GPT-Image-2]
     |                              |                        |
Photorealistic                 Characters              Collage/Texture
Underwater BG                  (octopus, pearl)        Elements
     |                              |                        |
     +----------+------------------+------------------------+
                |
         [Composite Layer]
         (manual or GPT-Image-2 compose)
                |
         [Style Unification]
         (Flux Kontext OR trained LoRA)
                |
         [Upscale + Final QC]
         (Flux 2 Pro 4MP or Gemini review)
                |
         Final Spread

6. FAL.ai Platform

Unified API aggregator: one API key, one billing system, access to Flux 2 Pro/Dev/Schnell, Flux Kontext Pro/Dev, Ideogram 3.0, Recraft V3, SD 3.5, InstantCharacter, NB2, and community LoRAs. Pay-per-use with no subscription. Swap models by changing an endpoint string. Ideal platform layer regardless of which models are chosen.

7. The Verdict

No single model replaces NB2/NBP. But specific models do specific layers much better:

• Flux 2 Pro — better photorealistic underwater environments
• Flux Kontext — better character-across-scenes consistency
• Ideogram Style Codes — better batch-level style consistency
• NB2/NBP remains best for artistic character generation

The best path forward: enhance the current NB2/NBP + GPT-Image-2 pipeline with targeted additions, accessed via FAL.ai as a unified platform.

8. Recommended Next Steps

1. Test Flux Kontext via FAL.ai — character identity in new scenes
2. Test Flux 2 Pro for 3–5 photorealistic Hawaiian reef backgrounds
3. Manual compositing test: Flux background + NB2 character + GPT-Image-2 collage
4. If compositing works, build multi-layer pipeline
5. If too jarring, train Flux Dev LoRA on approved reference images
6. Test Ideogram Style Codes for batch consistency

9. What NOT to Do

• Don't switch entirely to Leonardo — cheaper but lower quality for this use case
• Don't train LoRA on small dataset — 20–30 images minimum
• Don't expect any single model to produce photorealistic water + painted characters
• Don't abandon NB2/NBP for characters — still the best for stylized artistic elements