Your upscale sigma ladder shipped verbatim (credited), and your audio-layer notes steered a full-weight merge
Three of your things ended up load-bearing in the JoyAI-Echo multishot pack, so you should know:
The 4-step upscale schedule you published (0.92 / 0.725 / 0.421875 / 0) now ships as the "strong (tenstrip 4-step)" mode in the pack's hires refine, credited in the README: https://huggingface.co/joeygambino/joyai-echo-multishot-workflow
Your reshaped r256 DMD LoRA powered a dev+EchoDMD experiment - and measuring its deltas against Echo's full weights confirmed your design notes exactly: video layers directionally toward Echo, audio layers deliberately pointing at the 384-distilled components instead ("JoyAI totally screwed audio up" — measured and confirmed). That measurement is what convinced me a full-weight surgical merge was worth building: Echo's video branch + stock LTX audio, exact instead of approximated: https://huggingface.co/joeygambino/joyai-echo-ltx23-echoVid-ltxAud-surgical
Worth passing back: the ~10s lip-sync drift people blame on models/LoRAs turned out to be a hardcoded VIDEO_FPS=24 in JoyAI-Echo's release inference code - at 25fps it accumulates ~40ms/s of AV rope skew. If your users report sync drift on long shots, that's the actual cause, not your LoRAs.
Thanks for publishing the details most people keep private — the sigma schedules and layer notes saved me weeks.