| # Modification notice | |
| The files in this repository are quantized derivatives of the official | |
| Lightricks LTX-2.5 weights at revision | |
| `28dac7acdc1f78a70e98687db261a949754f8941`. | |
| Changes made by DmitryDB / DDB: | |
| - rebuilt Dev and Distilled into two transformer profiles with exact-source DDB searched scaling; | |
| - published the transformer profiles as `int8-convrot` and `nvfp4`; | |
| - converted the Gemma-4 LTX text encoder to `int8`, | |
| `int8_lean_convrot`, and `nvfp4` profiles; | |
| - converted the rank-450 Distilled LoRA to three DDB descriptor-based profiles; | |
| - retained selected sensitive tensors in BF16 according to deterministic | |
| architecture-aware policies generated from the BF16 source inventory; | |
| - added reproducibility recipes, layer policies, integrity hashes, and analysis | |
| summaries. | |
| No claim is made that Lightricks authored, approved, or endorsed these modified | |
| weights. | |