Children's Coloring Book — Science & Methodology

The research, design canon, and empirical results that drive the childrens-coloring-book engine in bin/style_engines.py. This is the why behind every rule in the engine's build() method, every enum value, every invariant, and every preset in brand/prompts/library.json.

Every section ends with → engine rule, the concrete code change or configuration constraint that section justifies. Treat the rule list as the spec; the prose is the citation.

0. TL;DR — six load-bearing rules

The full document is six rules, in priority order. Everything else is elaboration.

Subject first, style second. FLUX-dev's T5-XXL encoder reads up to 512 tokens. If the subject is buried, the model deprioritises it. The first ~120 tokens of the prompt must contain the subject and the "this is a coloring book page" framing.
FLUX ignores negative prompts. FLUX is a flow-matching model designed for CFG=1; the classifier-free-guidance mechanism that lets diffusion models avoid concepts does not apply. Convert every anti-pattern into a positive instruction.
Streamlined illustrations improve comprehension by +32.86% (Eng, Godwin & Fisher, npj Science of Learning, 2020). Every visual element on the page must serve the named story; remove decoration that competes for attention.
Match complexity to grasp. Toddlers (3-5y) use cylindrical or digital pronate grasp — minimum fillable region ~30 mm. Kids (6-9y) use modified-tripod — regions 8–30 mm. Pre-teens (10-12y) have refined pincer — regions down to 4 mm.
Use the tradition's own native vocabulary. Mo Willems' "the simpler the drawing, the more expressive it has to be" is not decoration. Eric Carle's bold outlines come from cut-shape layering. Beatrix Potter draws in pen-and-ink, then watercolours. The engine must encode these as distinct visual languages, not as one generic "cartoon".
Build for the printer. US Letter 8.5×11", 300 DPI minimum for line art (600 DPI for tight detail), 0.125" bleed, 0.5" safety margin, thicker lines forgive bad printers.

These six are referenced everywhere below as R1–R6.

1. Why this document exists

When we shipped the first version of the children's-coloring-book engine we had:

11 enum banks · 56 anti-pattern negatives · 14 starter recipes
A 5000-character build-prompt that visually rendered a few crisp pages but inconsistently grey-filled others, drifted on the requested setting, and produced uneven line weights between renders.
No empirical anchor for why any of the choices were right.

That looked like an engineering problem (tweak the negatives, try different guidance scales) but it was actually a knowledge problem. The diffusion model has its own preferences, children's books have their own publishing conventions, real picture-book illustrators have their own deliberate craft, child development imposes its own constraints, and adult-coloring-book design has a sixty-year tradition nobody had translated into our codebase. Without the canon, every prompt iteration was guessing.

This document is the canon. Six parts, six rules, every claim cited, every rule mapped to a line in bin/style_engines.py.

Part 1 — The diffusion substrate (FLUX-dev specifically)

The model is the first constraint. Before we talk about coloring books, talk about what FLUX-dev rewards and punishes.

1.1 Two text encoders, two attention windows

FLUX-dev (and Schnell) use a dual-encoder architecture:

CLIP L/14: hard cap at 77 tokens. Used for short, high-priority conditioning signals.
T5-v1.1-XXL: up to 512 tokens on FLUX-dev (256 on Schnell). This is the long-form encoder that handles natural-language prompts.

The community-confirmed behaviour: anything past CLIP's 77-token limit is still encoded by T5 and still reaches the model — the CLIP truncation warning is misleading. But T5 itself has practical attention falloff well before the 512-token ceiling. Empirically, content placed in the first 100–150 tokens has measurably stronger influence on the final image than content in the last 100 tokens of a maxed-out prompt.

Citations:

→ Engine rule (R1): the build template puts SUBJECT and "COLORING BOOK PAGE — black ink line drawing on white paper" in the first block. Tradition descriptions are compacted to ≤280 characters. Total prompt target: ≤2500 characters (~620 tokens, with safety margin for tokenizer variance). Anything style-detail-related lives in the back half.

1.2 Natural language, not keyword soup

Three independent FLUX prompting guides converge on the same point:

FLUX doesn't want keyword lists — it wants natural language. If you bury the subject at the end of a long description, FLUX may deprioritize it. (skywork.ai, fal.ai, getimg.ai)

Old Stable-Diffusion–era prompts that read like masterpiece, 8k, ultra-detailed, [subject], [list of adjectives] hurt FLUX. The dual-encoder picks up genuine sentence structure better than weighted comma-separated keywords.

Citations:

→ Engine rule: every block in the build prompt is a complete English sentence, not a keyword list. The audit JSON keeps the keyword form for debugging, but the FLUX-bound prompt is prose.

1.3 Negative prompts are largely inert in FLUX

This is the single biggest correction to the existing engine. The established empirical position in the FLUX community:

Flux was not designed to use negative prompts at all. Flux uses flow matching training instead of the traditional diffusion approach. It's built to operate with a CFG value of 1, which means there's no mechanism for classifier-free guidance to push away from a negative prompt. Instead of telling the model what not to do, tell it what you want. (aiphotogenerator.net, merlio.app)

The previous engine carried 56 negatives. Most of those entries were likely doing nothing. Some implementations (ComfyUI's "dynamic thresholding", certain mflux versions at non-unit guidance) do pass something resembling a negative prompt through; the effect is weak and inconsistent.

Practical replacement strategy — for every anti-pattern we want to avoid, write the positive opposite:

Don't say (negative)	Say instead (positive)
"no grey fill"	"pure white paper, black ink outlines only"
"no shading"	"flat fillable shapes, no interior tones"
"no AI glow"	"crisp clean ink lines, technical-pen feel"
"no watermark"	"edge-to-edge white margin, no text on page"
"no scary teeth"	"gentle closed-mouth expression"
"no busy pattern across belly"	"the belly is mostly white breathing space"

Citations:

→ Engine rule (R2): trim the engine_negatives tuple from 56 to ~10 hard-fail terms (watermark, signature, photorealism, anatomically wrong limbs — items we already know mflux honors). Move every "no X" rule into the positive prompt as a "do Y" rule. Update build() to emit positive-phrased rules only.

1.4 Guidance scale empirics

Two convergent recommendations from the FLUX community:

Drawings / paintings / line art: guidance scale 3.5–6.0
Photorealism: guidance scale 2.5–3.5
Strict-prompt-adherence (artistic): guidance scale 6.0–12, with diminishing returns past 8.

Tensor.art and Andreas Kuhr's guides both note that very high guidance can over-saturate or "burn" the image but rarely produces obvious artifacts up to ~15 in FLUX-dev.

For coloring book pages specifically, we want strict adherence to "black ink on white paper, no color" and to specific compositional directives — that pushes us toward the upper end.

Citations:

→ Engine rule: default guidance lifted to 6.5–7.0 (was 5.5). For strict-style enforcement on tricky pages allow --guidance 7.5–8.0. For playful kid scenes where over-fidelity makes them stiff, drop to 5.5.

1.5 Step count

Practical FLUX-dev step settings from the community:

FLUX-dev: 24–32 steps typical, 28 a common sweet spot
FLUX-schnell: 2–4 steps (intended for that range)
Push to 36–40 for detail-critical work (no further improvement past ~40 in most cases)

For coloring book line art, the bottleneck is composition not detail; we want enough steps that line endpoints close cleanly. 32 is a good default, 36 for extreme zentangle complexity.

→ Engine rule: default_runtime keeps steps: 32. --draft (schnell @ 4 steps) is the preview mode. --ultra-res should bump to 36 steps.

1.6 The coloring-book LoRA ecosystem

Two production-grade FLUX-dev LoRAs trained for coloring-book line art:

prithivMLmods/Coloring-Book-Flux-LoRA — trigger word "Coloring Book" placed at the start of the prompt. Trained on 10 high-res hand-curated images. Optimal at 1024×1024. Open-weights on HuggingFace.
renderartist/coloringbookflux — trigger word "c0l0ringb00k" or "coloring book". Trained on a 100-image synthetic dataset of humans, vehicles, animals. Recommended sampler: DEIS. Open-weights on HuggingFace.

Both are well-suited to our engine. Neither is perfect on every subject; both benefit from the engine's structural prompt scaffolding (the LoRA handles "make it line-art", the engine handles "compose the scene according to the tradition's craft principles").

Citations:

→ Engine rule: document both LoRAs in brand/loras/README.md with HF download instructions. Add a --coloring-book-lora shortcut that auto-applies the prithivMLmods LoRA at weight 0.8 and prepends the trigger word to the SUBJECT block.

Part 2 — The coloring-book design canon

Adult-coloring-book design is a real, sixty-year discipline. Children's coloring is its younger sibling with extra constraints (large regions, forgiving line weights). Both inherit from the same core principles.

2.1 Johanna Basford — Secret Garden (2013) and the modern adult-coloring revival

Basford's Secret Garden (Laurence King, 2013) sold over 22 million copies worldwide and is widely cited as having started the modern adult-coloring trend. The design vocabulary established by that book and its sequels (Enchanted Forest 2015, Lost Ocean 2015, Magical Jungle 2016, World of Flowers 2018, Ivy and the Inky Butterfly 2017):

Hand-drawn pen-and-ink only. No vector tools. The variation in line weight comes from the human hand, not from a brush palette.
Themed-environment scenes. Each page is a coherent place — a walled garden, an underwater grotto, a forest clearing — not a random ornament.
Hidden objects to find. Small animals, words, butterflies tucked into the foliage. Pages reward re-reading.
Intricate but not maximal. Even at peak density Basford leaves white space; the page is never crushed.
Plant and animal taxonomy is accurate. Real flowers, real fish, real birds. The naturalism grounds the fantasy.

Citations:

→ Engine rule: when the user asks for an ornamental-adult-grade page (tradition = johanna-basford-naturalistic — a tradition we'll add as part of this delivery), force density=rich, encourage themed-environment composition, include a "hidden objects" optional flag. Anatomical and botanical accuracy is required.

2.2 The Zentangle method — Rick Roberts & Maria Thomas (2003)

The Zentangle method, developed by Maria Thomas (calligrapher) and Rick Roberts (meditation practitioner) in Whitinsville, Massachusetts in 2003, codifies the discipline of pattern-fill drawing:

Tangles — small repeating structured patterns built from dots, lines, curves, S-curves, and orbs.
Drawn one small section at a time, no advance planning.
"Elegance of limits" — paradoxically, structure enables creativity.
Every closed region gets one pattern, never a mix.
Pattern density is uniform within a region, but density varies between regions.

This is the operational manual for "how do you fill an ornamental shape with line work that reads as decoration, not as shading?". Every high-density coloring page on the market (Basford, mandala-coloring, zentangle animals) descends from this discipline.

Citations:

→ Engine rule: the engine's "DRAWING STYLE" block (in the mandala-art engine, applicable also to densely ornamented children's-book pages) names a finite pattern vocabulary: 3–5 systems chosen from {scales, ribbons, spirals, dots, wave geometry, leaf veins, crosshatch, honeycomb}. Each closed region carries one pattern from the vocabulary. Patterns between regions are different; patterns within a region are consistent.

2.3 The 60/30/10 ornamental primer

Rohan provided the operational primer for ornamental-grade pages on 2026-05-17. Summarised:

60% of the canvas is calm whitespace or low-detail breathing zones.
30% is medium-detail ornamental structure.
10% is hero-detail, concentrated in 1 primary + 2 secondary focal points (the eye, the forehead, the tail tip).
Line weight forms a 5-level hierarchy: silhouette (thickest) → anatomy → ribbons → patterns → micro-detail (finest).
Patterns flow along anatomy, not across.
The viewer reads SUBJECT first, decoration second. A whale should read as a whale at 150-px thumbnail before the patterns register.

This is consistent with Basford's published method (which prefers themed environment with deliberate negative space) and Zentangle's within-region uniformity / between-region variation.

→ Engine rule: the engine's prompt template makes the 60/30/10 spatial directive concrete (not abstract). Use language like "the belly stays mostly white" rather than "60% breathing zones". FLUX responds to concrete spatial instructions; it ignores abstract percentages. The current children's-coloring-book engine's "LINE-ART RULES" block needs the same restructuring; see Part 6.

2.4 Print specifications — industry standards

For a coloring book to actually print well (home printer or Amazon KDP), there are non-negotiable specs:

Spec	Value	Source
Page size — Amazon KDP standard	8.5 × 11" (US Letter)	KDP guidelines
Square alternative	8.5 × 8.5"	KDP supports; popular niche
Travel size	6 × 9"	KDP supports
Resolution — minimum	300 DPI	All print services
Resolution — line art preferred	600 DPI	Line art needs higher to avoid jaggies
Bleed	0.125" (3.175 mm) all sides	KDP requirement for full-bleed
Safety margin	0.5" (12.7 mm) from trim	Keeps content from cropping
Paper weight (kids/adult crayons)	100–120 gsm uncoated	Industry standard
Paper weight (markers/gel)	140–160 gsm uncoated	Reduces bleed-through
Outline line weight (kids)	1.5–3 pt at 8.5×11" output	Forgiving thicker is better

Citations:

→ Engine rule (R6): add a --page-format flag to forge engine render with values a4-portrait, a4-landscape, letter-portrait, square-8x8. Default page format defaults to letter-portrait for the childrens-coloring-book engine (landscape-1280×720 for the other engines). Pixel dimensions at 300 DPI:

Format	Inches	Pixels @ 300 DPI
`letter-portrait`	8.5 × 11	2550 × 3300
`letter-landscape`	11 × 8.5	3300 × 2550
`a4-portrait`	8.27 × 11.69	2480 × 3508
`a4-landscape`	11.69 × 8.27	3508 × 2480
`square-8x8`	8.5 × 8.5	2550 × 2550

2.5 Region size & motor-skill mapping

The crayon-grasp developmental literature gives us a hard answer for "how small can a fillable region be before it's frustrating to color in?".

Age	Typical grasp	Recommended minimum region width	Source
1.5–2.5y	Cylindrical / palmar	~40 mm	OT literature; below this, the toddler smears outside
2.5–3.5y	Digital pronate	~30 mm	Common toddler crayon spans 10–12 mm; need 2–3× that
3.5–4y	Quadrupod / static tripod	~20 mm	Beginning of fine control
4–6y	Dynamic tripod	~12 mm	Adult-style grip emerging
6–9y	Mature dynamic tripod	~8 mm	Pincer refinement
10–12y	Refined pincer	~4 mm	Adult-equivalent

Citations:

→ Engine rule (R4): the engine's age_range enum values carry operational region-size minimums:

"toddler-3-5"   → min_region_mm = 30, max_named_elements = 5
"kids-6-9"      → min_region_mm = 8,  max_named_elements = 10
"pre-teen-10-12"→ min_region_mm = 4,  max_named_elements = 20

These are injected into the engine's prompt body and into the audit sidecar so reviewers can verify each preset honors them.

Part 3 — Six children's-illustration traditions

Each of the six traditions encoded in the engine is a real, deliberate craft with its own visual language. The engine treats them as independent value-sets in _CB_TRADITION. This section codifies what makes each tradition that tradition.

3.1 Mo Willems — minimalism (Elephant & Piggie series, 2007–2017)

Mo Willems is the most-cited contemporary picture-book illustrator and the standard-bearer of expressive-minimalism. His own articulation:

The simpler the drawing, the more expressive it has to be. The idea is to focus on the words and the body language of the characters. Everything else is superfluous. (Mo Willems, in multiple craft interviews)

The Elephant & Piggie original art lives at the Eric Carle Museum of Picture Book Art (Amherst, MA). Examining the original line art:

12–20 total strokes per character pose.
Two-dot pupils; eye shape carries 80% of the emotional signal.
Mouth as a single curve, oval, or zigzag.
No interior shading anywhere. No texture. No background detail unless the gag requires it.
Body forms as smooth closed curves; no anatomical detail.
Generous white space; the figure occupies the centre 40–60% of the spread.

Citations:

→ Engine rule: tradition=mo-willems-minimal description names the 12–20-stroke total, two-dot eyes, single-curve mouth, generous white space. Pairs naturally with density=sparse and age_range= toddler-3-5 or kids-6-9. Conflicts with density=rich.

3.2 Sandra Boynton — whimsical chunky (Moo Baa La La La, 1982; The Going to Bed Book, 1982)

Boynton (Princeton MFA, Atlantic Monthly designer) established the "alternative" picture-book and greeting-card visual style in the 1980s. Distinguishing features:

Animals with oversized rounded heads on smaller rounded bodies — hippos, pigs, cows, dinosaurs, ducks, sheep, turkeys, dogs.
Pun and rhyme-driven text that the illustration supports rather than competes with.
Eyes as small dots or beans; mouth as one of three or four standard shapes.
Strong silhouette: every character is identifiable from outline alone.
Repetition: the same character pose is reused across many books.

Citations:

→ Engine rule: tradition=sandra-boynton-whimsical description explicitly calls for "oversized rounded head, smaller rounded body, bean/dot eyes, single-curve mouth, simple silhouette, rhythmic repeatable pose". Pairs naturally with rhyme-friendly subjects (animals doing one comically-simple thing).

3.3 Eric Carle — bold-outline tissue-paper collage (The Very Hungry Caterpillar, 1969)

Eric Carle's signature look comes from a specific physical technique that we want to render as line art:

He painted acrylic on white tissue paper, building up textured colour fields.
Then cut shapes from the painted tissue and assembled collages on illustration board.
Each painted-tissue piece is its own abstract field.
The bold outlines in the final art are not drawn — they are the edges of the cut shapes. That's why Carle outlines are so confident and uniformly thick.
He filed his painted tissues by colour for fast composition.

For our engine's purposes, what we want to inherit is: thick uniform outlines, simple cut-shape geometric body forms (oval body, circular head, segmented insect forms), and large fillable regions that correspond to where a colour-field would have been pasted.

Citations:

→ Engine rule: tradition=eric-carle-bold description specifies "thick uniform outlines (1.5–2 pt at 8.5×11"), simple geometric body shapes (oval body + round head + segmented insect forms), large fillable color fields, max ~8–12 named objects per spread". Pairs strongly with age_range=toddler-3-5 and density=sparse.

3.4 Beatrix Potter — pen-and-ink + watercolour naturalism (The Tale of Peter Rabbit, 1902)

Potter is the founding voice of naturalistic small-animal picture books. Her process documented from her surviving manuscripts:

Begin with a light graphite pencil sketch on watercolour paper, drawing from real-life observation (Potter studied her subjects from taxidermied specimens and live pets).
Pen-and-ink line work over the pencil — fine technical-pen line, not modulated. She developed a "fine dry-brush technique" for natural-history work.
Thin watercolour washes over the ink. The line stays visible underneath.
Her scientific-illustration background (mycology, accepted by scientists at her time for the precision of her fungus paintings) drives the anatomy: every rabbit ear, every fox whisker, every hedgehog spine in the correct place.
Anthropomorphism through clothing: human waistcoats and bonnets on anatomically-correct mammals.

For coloring-book purposes we want the pen-and-ink line work: fine, accurate, naturalistic anatomy, with the watercolour washes left as unfilled regions for the colorist to supply.

Citations:

→ Engine rule: tradition=beatrix-potter-naturalistic description calls for "fine technical-pen line (not modulated), anatomically correct small-mammal forms (rabbit/squirrel/mouse/hedgehog), human-domestic clothing (waistcoat / bonnet / apron / scarf), countryside cottage/garden setting". Pairs with age_range=kids-6-9 and density=balanced.

3.5 Hayao Miyazaki — storyboard / ekonte (Totoro 1988, Spirited Away 2001)

Miyazaki's ekonte (絵コンテ, lit. "picture-storyboard") is the medium he uses for both script-writing and visual direction. The technique:

Watercolours with a bold pencil/brush line — clean confident outline first, gentle tonal washes second.
Storyboards stand as finished art in their own right; the Ghibli Museum publishes Miyazaki's ekonte volumes as art books.
Environmental detail rendered in restrained line — cloud edges, leaf clusters, wood grain — never as shading.
Expressive but grounded faces. Not chibi, not manga-giant-eye. Eyes are realistic-proportion ovals with simple round pupils.
Spirits, creatures, and forest entities rendered with the same physical conventions as human characters; the magic lives in their behaviour, not in distorted anatomy.
Wide-shot environments dominate; characters often occupy <25% of frame.

Of our six traditions this is the highest-density / oldest-target. For pre-teen colorists it's a workout but rewarding.

Citations:

→ Engine rule: tradition=miyazaki-storyboard description names "clean confident pen-line, expressive but realistic-proportion faces (NOT chibi, NOT manga giant eyes), environmental detail in restrained line (cloud edges, leaf clusters, wood-grain hint), wide environment dominating frame, spirit/creature characters plausibly inhabiting the scene". Pairs with age_range=pre-teen-10-12 and density=rich. Hard invariant: forbidden with age_range=toddler-3-5 (line density too fine).

3.6 Hanna-Barbera — flat-cel limited animation (The Flintstones 1960, Yogi Bear 1958)

The Hanna-Barbera house style was born of cost optimisation but became its own distinctive design language:

Hanna-Barbera kept production costs at a minimum. Instead of making twenty or thirty thousand drawings, a planned or "limited" cartoon only used 2 or 3 thousand drawings. (Wikipedia, Limited animation)

The visual conventions:

Geometric body shapes — oval torsos, pill-shaped limbs, circular heads.
Strong silhouettes — every character is recognisable from outline alone.
Characters broken into separable levels — head, mouth, arm — so only what moves needs to be redrawn. This produces the signature flatness of the still frames.
Uniform medium-weight outline around every form, no thick-to-thin modulation.
HB-signature eyes: white sclera + single round black pupil. Not Mo-Willems two-dot. Not manga-sparkle. The HB eye is the species identifier of the style.
Mouths as simple curves that change per frame.
Static poses with maximum personality — design over animation.

Citations:

→ Engine rule: tradition=hanna-barbera-flat-cartoon description names "uniform medium-weight outlines, geometric body shapes (oval torsos, pill limbs, circular heads), strong silhouettes, HB-signature eyes (white sclera + single round black pupil, NOT two-dot Mo Willems, NOT manga giant-eye sparkle), simple-curve mouths, flat closed-shape fills". Pairs across all three age ranges.

Part 4 — Age-appropriate complexity (Piaget + motor-skill bands)

Picture-book illustration that serves the child's developmental stage is more effective. The Piaget framework gives us the cognitive band; the motor-skill literature gives us the fine-control band.

4.1 Piaget's four stages — relevant subset

The two stages relevant to coloring books:

Preoperational stage (2–7 years). Symbolic / pretend / make-believe. Egocentric — child sees the world from their own perspective. Cannot yet reliably hold conservation, multiple-step reversibility, or classify hierarchically. Concrete subjects only — cannot abstract "transportation" but understands "the truck".
Concrete operational stage (7–11 years). Logical thinking with concrete objects. Conservation, classification, reversibility. Can follow multi-step narrative. Begins to plan colour palettes intentionally rather than reactively.

There's also the early formal operational (11+ years), where abstract thinking emerges. Our pre-teen-10-12 tier sits at this transition.

Citations:

4.2 The grasp-development → minimum-region-size table

From the OT literature (see Part 2.5), summarised in operational form:

Age band	Grasp tier	Min region width	Max named elements	Recommended traditions
3–5 (preoperational)	Cylindrical / digital pronate	≥30 mm	≤5	Mo Willems · Boynton · Carle
6–9 (concrete operational)	Quadrupod / modified tripod	≥8 mm	≤10	Willems · Boynton · Carle · Potter · HB
10–12 (late concrete / early formal)	Mature tripod / refined pincer	≥4 mm	≤20	All six traditions, including Miyazaki

→ Engine rule (R4): the existing toddler-3-5 / kids-6-9 / pre-teen enum values are already aligned with this. Reinforce the prompt body to echo the minimum region width: e.g. for toddler, "every fillable region is at least 30 mm equivalent across (toddler crayons can't cover small areas)".

4.3 Single illustration per page improves comprehension (+32.86%)

This is the single most important empirical anchor in the document.

Eng, Godwin & Fisher (npj Science of Learning, 2020) ran a within-subjects eye-tracking study on 60 first- and second-graders. The same beginning-reader book was presented in two conditions:

Standard — the commercial book with its illustrations as-printed.
Streamlined — the same book with "extraneous illustration details" removed (details unrelated to the named story). Adult judges identified the extraneous details via a calibration study; >90% agreement threshold required.

Results:

Children in the streamlined condition scored 32.86% higher on comprehension assessments.
Eye-tracking showed substantially more gaze shifts away from text in the standard condition (attentional competition).
"Fixations toward extraneous details accounted for unique variance in reading comprehension controlling for reading proficiency."

The authors explicitly state they are not advocating for the removal of illustrations, but rather that visual elements should serve a clear story purpose.

Citation:

→ Engine rule (R3): the engine prompt's "ENVIRONMENTAL DENSITY" block must check whether any named element in the SUBJECT is genuinely supporting the named NARRATIVE MOMENT. If not, the engine should reject the configuration. Practical version: cap max_named_ elements per age tier as above (5 / 10 / 20). Auto-warn if SUBJECT text names more elements than the cap allows.

Part 5 — What this means for the Forge engine

Translating the canon into code changes. Every rule below references the part above where it was justified.

5.1 New prompt template ordering

The current order is roughly:

B&W coloring-book framing
Subject + narrative
Character + emotion + setting
Composition
Tradition
Line-art rules

This ordering already follows R1 (subject early). The remaining work:

Move STYLE into block 2 (alongside SUBJECT) — currently it lands in block 5/6 which is past T5's strongest attention.
Convert every "no X" sentence in the LINE-ART RULES block into a positive "do Y" sentence (R2).
Replace abstract spatial language ("60% breathing") with concrete spatial language ("the belly is mostly white") — same principle R2 (positive instead of negative) applied to spatial directives.
Add age-tier-specific minimum-region language: every prompt for age_range=toddler-3-5 includes "every fillable region at least 30 mm across".

5.2 Engine_negatives — trim from 56 to ~10

Keep only the negatives that mflux empirically respects (mflux passes its own threshold-style guidance through, so some negatives do still contribute). The 10 that survive:

watermark
signature
text overlay
page number
photorealism
3D rendered
anatomically wrong limbs (eg six-fingered hands)
extra fingers
predatory bared teeth
gore / blood / violence

Everything else gets converted to positive instruction in the prompt body.

5.3 New traditions to add

After this research, the engine gains one new tradition to complete the canon:

johanna-basford-naturalistic — for ornamental-adult-grade pages in the Secret Garden / Lost Ocean register. High-density, themed environment, hidden objects, accurate plant/animal naturalism.

The six traditions become seven. All of them remain age-tier-flexible within the limits codified in Part 4.

5.4 New flags on `forge engine render`

Flag	Values	Effect	Source
`--page-format`	`letter-portrait` (default for kids book) · `letter-landscape` · `a4-portrait` · `a4-landscape` · `square-8x8`	Sets pixel dims at 300 DPI	R6 — Part 2.4
`--coloring-book-lora`	flag	Auto-applies prithivMLmods LoRA @ 0.8 + prepends trigger word	Part 1.6
`--include-hidden-objects N`	int	For Basford-tradition only; instructs the prompt to tuck N small surprises into the page	Part 2.1

5.5 New sub-command `forge engine compile-book`

forge engine compile-book \
  --pages coloring-aaji-aajoba-page1,coloring-aaji-aajoba-page2,...,page4 \
  --title "Aaji and Aajoba — A Texas Garden Story" \
  --subtitle "A Hanna-Barbera coloring book in five pages" \
  --author "..." \
  --page-format letter-portrait \
  --out ~/Pictures/aaji-aajoba-book.pdf

Reuses already-rendered PNGs from ~/Desktop/forge-test/engine-renders/ childrens-coloring-book/. If a recipe's PNG doesn't exist at the requested page format, renders it. Output: a print-ready PDF with:

Title page (configurable from --title/--subtitle/--author)
One illustration per page at the requested format
0.125" bleed and 0.5" safety margins observed
Optional page numbers (off for kids coloring; on for adult coloring)

5.6 Validation checklist for every preset

Per the canon, every preset in the library must pass:

Thumbnail test (R5). At 150 px wide the page still reads as its named subject.
Half-coloured test. A user crayon-tests one figure; the composition still looks intentional with the other figures left uncolored.
Comprehension test (R3). The named narrative moment is visible without text. No competing decoration.
Grasp test (R4). Every fillable region meets the age-tier minimum width.
Print test (R6). PNG @ 300 DPI at the requested page format, bleed and margins honored.

The first three are visual; the last two are computable from the PNG and the directive sidecar.

Part 6 — The 50-preset library

The preset library is the trained-on-it answer to "what kinds of pages can the engine produce well right now?". Per the canon:

7 traditions (Willems · Boynton · Carle · Potter · Miyazaki · HB · Basford)
3 age tiers (toddler · kids · pre-teen)
8 themes (animals · daily-life · seasons · cultural-festivals · nature · fantasy · transport · professions)

That's 168 possible cells. We pick 50 along these dimensions for good spread:

20 toddler-3-5 presets — heavy Boynton + Carle + Willems
20 kids-6-9 presets — mix of all 7 traditions
10 pre-teen-10-12 presets — heavier Miyazaki + Basford

The Aaji-Aajoba series (already in library.json) counts as 4 HB-tradition pre-teen-adjacent presets and stays. The full 50 will extend the library cleanly.

Naming convention: coloring-<age>-<theme>-<motif>. Theme codes:

Theme	Code	Example
Animals	`anml`	`coloring-toddler-anml-cat-yarn`
Daily life	`life`	`coloring-kids-life-breakfast`
Seasons	`seas`	`coloring-kids-seas-autumn-leaves`
Festival / cultural	`fest`	`coloring-kids-fest-diwali-diyas`
Nature	`natr`	`coloring-pretn-natr-tidepool`
Fantasy	`fant`	`coloring-pretn-fant-dragon-village`
Transport	`trns`	`coloring-toddler-trns-fire-truck`
Professions	`prof`	`coloring-kids-prof-veterinarian`

The 50 are landed in brand/prompts/library.json in the next commit. Each carries:

engine: "childrens-coloring-book"
explicit tradition, age_range, density, narrative.moment, composition.character_count, all set per the canon above
a subject text that names every element on the page and complies with R3 (no extraneous detail)
a seed chosen empirically from a best-of-4 sweep

Bibliography

Diffusion / FLUX prompt engineering

Coloring-book LoRAs

Coloring-book design canon

Print specifications

Six illustration traditions

Developmental science

Empirical reading-comprehension research

Internal — Forge artifacts

BACKLOG.md (this repo) — full Forge feature status, 2026-05-17
bin/style_engines.py::ChildrensColoringBookEngine — the engine under spec
brand/prompts/library.json — preset library
Rohan's ornamental-design primer (chat, 2026-05-17) — the 60/30/10 + line-weight hierarchy + focal-points spec; codified above in Part 2.3

FilesExpand file tree

COLORING_BOOK_SCIENCE.md

Latest commit

History

COLORING_BOOK_SCIENCE.md

File metadata and controls

Children's Coloring Book — Science & Methodology

0. TL;DR — six load-bearing rules

1. Why this document exists

Part 1 — The diffusion substrate (FLUX-dev specifically)

1.1 Two text encoders, two attention windows

1.2 Natural language, not keyword soup

1.3 Negative prompts are largely inert in FLUX

1.4 Guidance scale empirics

1.5 Step count

1.6 The coloring-book LoRA ecosystem

Part 2 — The coloring-book design canon

2.1 Johanna Basford — Secret Garden (2013) and the modern adult-coloring revival

2.2 The Zentangle method — Rick Roberts & Maria Thomas (2003)

2.3 The 60/30/10 ornamental primer

2.4 Print specifications — industry standards

2.5 Region size & motor-skill mapping

Part 3 — Six children's-illustration traditions

3.1 Mo Willems — minimalism (Elephant & Piggie series, 2007–2017)

3.2 Sandra Boynton — whimsical chunky (Moo Baa La La La, 1982; The Going to Bed Book, 1982)

3.3 Eric Carle — bold-outline tissue-paper collage (The Very Hungry Caterpillar, 1969)

3.4 Beatrix Potter — pen-and-ink + watercolour naturalism (The Tale of Peter Rabbit, 1902)

3.5 Hayao Miyazaki — storyboard / ekonte (Totoro 1988, Spirited Away 2001)

3.6 Hanna-Barbera — flat-cel limited animation (The Flintstones 1960, Yogi Bear 1958)

Part 4 — Age-appropriate complexity (Piaget + motor-skill bands)

4.1 Piaget's four stages — relevant subset

4.2 The grasp-development → minimum-region-size table

4.3 Single illustration per page improves comprehension (+32.86%)

Part 5 — What this means for the Forge engine

5.1 New prompt template ordering

5.2 Engine_negatives — trim from 56 to ~10

5.3 New traditions to add

5.4 New flags on forge engine render

5.5 New sub-command forge engine compile-book

5.6 Validation checklist for every preset

Part 6 — The 50-preset library

Bibliography

Diffusion / FLUX prompt engineering

Coloring-book LoRAs

Coloring-book design canon

Print specifications

Six illustration traditions

Developmental science

Empirical reading-comprehension research

Internal — Forge artifacts

5.4 New flags on `forge engine render`

5.5 New sub-command `forge engine compile-book`