🏠 HomeVision

AI Real Estate Photo Editor — 4-model pipeline running entirely on iPhone. No cloud. No per-image costs.

Current Phase

Phase 1 — All Models Training Overnight

Last Updated

April 17, 2026 — 8:15 AM

Day 2 of 60 — Target: June 15, 2026 10%

Live Training Dashboard

4 / 4

Models Training

137

Total Epochs

0.89

Best Sky IoU

0.011

Editor Best Loss

🎭 Scene Mask Model

Epoch 46 / 500

0.89

Best Sky IoU

0.98

Best Grass IoU

0.71

Best Object IoU

98.6%

Best Ext. Accuracy

Segmentation IoU — higher is better (dashed = 90% target)

Epochs Done

0.83

Best Score

~8 min/ep

Epoch Time

MacBook Pro

MPS GPU

Training & validation loss — lower is better

✏️ Editor Diffusion Model

Epoch 42 / 500

88M

Parameters

0.011

Best Val Loss

0.050

V-Loss (latest)

0.134

Detail Loss

Total loss + V-prediction loss — converging fast toward target

Epochs Done

~9 min/ep

Avg Epoch Time

Self-Cond

Enabled

MacBook Pro

MPS GPU

🎨 Color Diffusion Model

Epoch 29 / 1000

~10M

Parameters

0.046

Best Val Loss

0.189

V-Loss (latest)

0.012

Chroma Loss

Residual diffusion — predicts color delta for fp16-safe on-device inference

Epochs Done

~14 min/ep

Avg Epoch Time

Residual

NEW Prediction

Intel Mac

CPU (8 threads)

🔍 4x Upscaler Model

Epoch 7 / 500

~1.8M

Parameters

0.071

Best Val Loss

0.029

Pixel Loss

0.094

SSIM Loss

Pixel + SSIM + Edge loss — guided 4x super-resolution

Epochs Done

~54 min/ep

Avg Epoch Time

Gated Fusion

BEFORE Guidance

Intel Mac

CPU (8 threads)

Architecture Overview

📸 Photo → Scene Analysis → Color Correction → Editor → 4x Upscaler → ✨

Model	Task	Params	Key Innovation
Scene Mask	Segmentation + classification	~2M	Multi-head: masks + exterior + difficulty
Color	Color correction	~10M	Residual diffusion — fp16-safe
Editor	Sky/object/grass editing	~88M	Edit-mask-weighted + self-conditioning
Upscaler	4x super-resolution	~1.8M	Gated guidance fusion from BEFORE

Phase 1 Changes (Night 1)

✅ Complete code rewrite — 11 files across all 4 models
✅ Color model → residual prediction (fp16 fix)
✅ Self-conditioning on color + editor
✅ Perceptual losses: detail, color, SSIM, saturation
✅ Reflection padding + bilinear upsampling
✅ EMA + log files for persistent monitoring
✅ All training data utilized — zero waste

8-Week Roadmap

Weeks 1–2 — April 16–30

Phase 1: Model Rewrite & Training

Code review all 4 models
Rewrite all notebooks
Phase 1 training — all 4 running
Review results, tune hyperparams
Continue creating masks
Share Xcode project

Weeks 3–4 — May 1–14

Phase 2: High-Res + CoreML

Full resolution training
CoreML export
On-device testing
Begin app UI

Weeks 5–6 — May 15–28

Phase 3: App Development

Complete features + StoreKit 2
Integration testing

Weeks 7–8 — May 29 – June 15

Phase 4: Polish & Launch 🚀

App Store assets + TestFlight + Submit

Competitive Edge

Feature	HomeVision	Autoenhance	BoxBrownie	Lightroom
On-device AI	✅	❌	❌	❌
Works offline	✅	❌	❌	✅
Sky replacement	✅	✅	✅	❌
4x upscaling	✅	❌	❌	❌
Per-image cost	$0	$$$	$1.60–4	$0
Speed	~30s	Minutes	Hours	Manual

Pricing

Free

5 images/month

Monthly

$7.99

per month

Annual ⭐

$49.99

per year

$4.17/mo — 48% savings

Target: 10,000 subscribers × $49.99/yr = $500K ARR

Launch Checklist

CoreML models (all aspects, fp16)
On-device testing
App feature-complete
StoreKit 2
App Store screenshots + video
TestFlight beta
Submit + Launch 🚀