Community merge family built on LTX-2.3's audio branch + a working recipe for the x1.5 spatial upscaler's sequence-end corruption
Two things — a showcase and a workaround your users may want.
The showcase: LTX-2.3-distilled-1.1's audio branch is carrying an entire family of community merges — its voice quality scored 9-10/10 ("indistinguishable from a real actress") in blind line-by-line reviews, and we've paired it with JoyAI-Echo's multi-shot video branch in a per-tensor surgical merge, plus INT8 ConvRot exports using your official layer recipe (blocks 2-45 quantized, 0/1/46/47 bf16 — it transferred perfectly to the merged weights): https://huggingface.co/joeygambino/joyai-echo-ltx23-echoVid-ltxAud-surgical
The workaround, re: the confirmed x1.5-1.0 spatial upscaler issue (discussion #13, first/last-frame corruption on long sequences) — two practical rules that make it production-usable until a fixed x1.5 ships:
Reflect-pad the temporal axis ~8 latents per side before the upsampler and crop after — the boundary damage lands entirely in disposable mirrored latents. (Pad by mirroring [1..8] and [-9..-2], not edge-repeat.)
Keep spatial latent dims EVEN (canvas /32 must be even — 768-height, not 736): at x1.5 an odd latent count makes the rational resampler fabricate a half-latent band on one edge, which renders as a full-length smear.
With both rules applied the x1.5 upscaler runs clean on 300+ frame sequences. Happy to share the wrapper implementation if useful.
ooooh thank you!