Garage Force

Pipeline at Scale

A complete AI video production pipeline for a garage flooring company. Built a system that scrapes real job site photos from CompanyCam, isolates garage floors using SAM2 segmentation, and generates before/after transformation videos using Kling AI and ComfyUI. Tested 100+ creative transition effects across 115 real properties with a 3-act video structure: declutter (remove clutter via segment-inpaint), transform (apply the new floor digitally), beauty shot (final polished reveal). Multiple engines tested systematically — Kling V3, LTX Video, Wan 2.2, FramePack — to find the right look for each property.

Accuracy came from two custom LoRAs I trained for the job: a final-floor LoRA locked to Garage Force's actual installed floor textures (flake blends, color ratios, reflectivity) so every "after" render matched what the crew would actually lay down, and a previz LoRA trained on their real job-site staging so the declutter and transform passes read like the same garage, not a generic render. That is what made the before/after believable at scale instead of off-brand AI guesswork.