🎯 Overview
Recommended approach: We are not using AI to redraw the whole page. Instead, the artist keeps the page and background exactly as approved, and AI only generates the personalized child character. Our system then composites the character back onto the artist's page automatically.
Why this works for FoxeTales: the artist's illustration stays pixel-perfect on every order. AI only touches the one element that needs to change per customer — the child. This gives us consistent brand quality, predictable results, and an easy review process.
Status: Personalization technology validated across 4 ethnicities on all 34 pages of T16. Ready to move to the recommended production setup with artist-prepared layers.
📋 Recommended Production Plan
Three clear stages. Each one has a clear owner and a clear output.
Keep the approved page fixed
The artist's finished page — background, props, layout, typography — stays exactly as approved. Nothing AI does will alter it. This is the "stage" the personalized character will stand on.
Generate only the personalized child
AI receives the customer's photo and generates only the child character — in FoxeTales watercolor style, matching the original pose. Output is a clean character image with a transparent background.
Auto-composite back together
Our system places the new character onto the approved page using the artist's placement, mask, and shadow metadata. Every order comes out aligned the same way — no per-order manual fix needed.
Net result: the brand-critical artwork is locked, the per-child element is personalized, and assembly is fully automatic.
⚙️ How It Works (Simple Pipeline)
One customer order, one page — step by step:
The background and page layout never go through AI. Only the child character does. That keeps the artist's design intact across every order.
📊 Results So Far
We tested the character personalization across the full book (34 pages) and across multiple ethnicities to make sure it holds up at production scale.
Sheet 1: 5 ethnicities × 2 templates
Sheet 2: Comparison of two personalization approaches
Sheet 3: Skin tone challenge (and fix)
Sheet 4: Recommended setting (winner)
Sheet 5: Skin tone post-process
Sheet 6: Full book validation
We ran the personalization across all 34 pages of T16, for 4 different ethnicities — a total of 136 personalized renders. Below: 6 representative pages × 4 ethnicities.
🎯 Tradeoffs & Why We Recommend This Path
We compared a few production approaches. Here is what we recommend, and why.
| Approach | What it means | Recommendation |
|---|---|---|
| Character-layer generation + deterministic compositing | Artist locks the page and background. AI generates only the personalized child as a transparent layer. Our system composites it back onto the approved page using artist-supplied placement, mask, and shadow data. | ✓ Recommended production path |
| Full-page / full-character AI inpainting | AI redraws a large portion of the page (character + surrounding pixels) in one shot. Faster to prototype, but every order touches the artist's artwork. | Fallback only — for pages where the layered approach is not yet ready |
| Pre-drawn template variants per gender / hair / ethnicity | Artist hand-draws several variants per page. No AI involvement. | ✗ Not scalable — multiplies artist workload |
Why we recommend the layered approach: the artist's page stays untouched on every order, results are consistent and reviewable, and the system is deterministic — same inputs always give the same output. The full-page approach remains available as a fallback for edge cases.
💰 Cost Analysis
Detailed cost model: 📄 Full cost analysis
Per-order economics
| Item | Cost | Note |
|---|---|---|
| AI compute per order (~35 pages + previews) | ~$1.61 | Serverless GPU, roughly 51 min of compute |
| + Safety buffer (cold start, retries) | $0.40 | Production margin |
| + Storage, database, CDN | $0.06 | File storage + delivery |
| Total per order | ~$2.07 (~53,000 VND) | Stable across scale |
Monthly cost by scale
| Orders / month | 1,000 | 2,500 | 5,000 | 7,500 | 10,000 |
|---|---|---|---|---|---|
| GPU hours (estimate) | 854 | 2,134 | 4,268 | 6,402 | 8,536 |
| GPU cost (raw) | $1,614 | $4,034 | $8,067 | $12,101 | $16,134 |
| + buffer + storage + monitoring | $503 | $1,185 | $2,291 | $3,426 | $4,552 |
| Total / month | ~$2,117 | ~$5,219 | ~$10,358 | ~$15,527 | ~$20,686 |
| Cost / order | $2.12 | $2.09 | $2.07 | $2.07 | $2.07 |
Revenue context
Industry benchmark: Wonderbly sells personalized books at $40–60. AI personalization at ~$2/order is ~5% of revenue — a very healthy margin.
Optimization headroom: roughly 40% cost reduction is achievable with about a week of optimization work after launch.
🚀 Next Steps
A focused plan to move from validated technology to the recommended production setup.
🙏 What We Need From the Artist Team
To move into the recommended production setup, we need the following per pilot page:
-
background_without_character.png— the approved page with the original child character removed, so we have a clean stage to place the personalized character onto. -
original_character_layer.png— the original character on a transparent background. (Alternatively: a full-canvas transparent character layer at the same dimensions as the page.) This is our reference for pose, scale, and style. -
character_mask.png— a clean mask defining exactly where the character sits on the page. Used to guide compositing and ensure crisp edges. - Placement metadata — position, scale, and rotation of the character on the page. A short data file (or simple notes) so the composite is pixel-accurate every time.
- Shadow / occlusion notes — any shadows the character casts, or props that overlap the character. Helps us preserve depth and realism in the composite.
- 5–10 consented customer photos — real customer photos (with consent) so we can validate the pipeline on real input, not just stock imagery.