| feat(models): add LTX-2.5 as an installable model family - Taxonomy, configs and loaders for the LTX-2.5 component folder, the single-file transformer (bf16 / int8-convrot / header-named nvfp4) and the Gemma-4 text tower; diffusers model classes with pinned configs and key maps verified against the release - Architecture definition with a fitted latent preview space, starter models from the ungated mirror, shared single-file quantization guards extracted from the H3 loader - Identification is structural (2.5 = no video feed-forward bias; four-stage VAE decoder); stock sharded Gemma-4 stays a generic text LLM | 16 天前 |