M Hudson Design × aiChemist · production-tech briefing

The pipeline is built and measured.
Six decisions set what's next.

FROM Jonathan (aiChemist)TO Matt Hudson DATE 2026-07-04NEXT 7/7 show-and-tell
Status pulse: your full 509-image Connors archive has been processed under quality gates · two public showcase pages are live in your website's own brand · your visual signature is now trained into the model itself. Everything below is measured, not estimated.
Action needed · ~5 minutes

Six decisions, each with a recommended default

These are the only inputs the build needs from you before the mid-August demo track locks in. The form takes about five minutes; every question has a recommendation you can simply accept.

1 · Does the trained brand style read as "M Hudson"?
Recommended: approve direction + commission v2
We trained your visual signature into the model from 96 of your own images. Samples are in the show-and-tell. v2 retrains on your real built-work photography — sharper realism, and it drops the watermark artifact v1 picked up from the Lumion stills.
2 · Client project pages: dark gallery look, or light like your website?
Recommended: dark sage gallery
Public pages now match mhudsondesign.com's light botanical brand. A dark-sage interior for the client exhibits keeps the "studio gallery" drama while staying on-palette.
3 · Keep the client project pages public, or gate them again?
Your call — both take minutes
The Connors / B. Davis / Wiener exhibit pages are currently reachable by anyone with the link. Public = easy sharing at 7/7. Gated = password protected as before.
4 · Keep the current cost posture?
Recommended: yes — keep it
69% of your archive's finished variations cost $0 (in-house GPU). The paid model is only used where measurement proved it necessary — roughly 13–24¢ per image at those points.
5 · 7/7 agenda: run the proposed sequence?
Recommended: proposed sequence
Your plans rendered → any drawing from anywhere → your 509-still archive, restyled and browsable → walk inside a design → the pick-a-mood interaction your clients will drive themselves.
6 · Green-light the first capture-grade walkthrough property?
Recommended: yes — pick the property at 7/7
Today's 360 walks are AI teasers. The real product is a scanned, navigable model of an actual property — Carlos's capture feeds it. Choosing the first property starts that clock.
Open the decision form →
Live now · public

Two showcase phases, in your brand

Both pages carry your website's own palette and typography — extracted from mhudsondesign.com — so everything the studio produces already looks like you.

Phase I — Your plans, rendered

Five of your own plan sheets: each becomes a photoreal render, then every mood, three planting design languages, and a 360 walk. Geometry held to the drawing throughout.

Phase II — Any drawing, any source

Five arbitrary drawings pulled from the public web — four at thumbnail size, including raw CAD linework and an architect's section sheet. The intake router classified each one and ran it end-to-end, unassisted.

Your archive, working

509 stills in. 273 finished variations out. Every one quality-gated.

The Connors PP-SI export folder — fifteen design rounds of Lumion stills — went through the packaged pipeline. Nothing was hand-tuned; outputs that failed the quality bar were quarantined, not published.

509
Stills classified
273
Gated variations published
69%
Of passes at $0 marginal
100%
Outputs quality-scored
Every output was scored; only passes were published. "Quarantined" means the quality gate caught it — that's the system working, not failing.
How it works

One pipeline, any drawing

Every input takes the same path: classify → realize → vary → walk → gate. The gate is an independent scorer — geometry fidelity, mood truth, realism — and nothing reaches a page without passing it.

The paid model appears only where measurement proved the in-house lane can't do the job; everything else runs on the studio GPU at $0 per image.
Dates

From Monday's show-and-tell to the November launch

Your decisions this week set the LoRA v2, exhibit styling, and capture tracks in motion — all three land well before the mid-August demo.
Brand lock

Your signature, in the weights

Version 1 of the brand model trained on 96 curated images of your work. The output samples carry your material palette, planting language and lighting temperament — including night scenes, which the base model couldn't produce at all.

What this unlocks

Decisions → build tracks → demo

Decision #1 starts the sharper v2 brand model. Decision #2 finishes the exhibit styling. Decision #6 starts the real-property walkthrough with Carlos. All three converge on the mid-August demo.

Your 6 decisions LoRA v2 + dark-sage exhibits First property capture Mid-Aug demo Nov 1 studio launch
Who does what once the decisions land.
Make the six calls →