ad-astra-video / LTX-Loras

A curated collection of LTX LoRAs, pre-packaged for the video-creator runner.

Every adapter here is ready to load out of the box: keys are in the ltx-pipelines / Lightricks format that the video-creator LTX worker fuses (.lora_A.weight / .lora_B.weight under a diffusion_model. prefix), so no code changes or conversion are needed — just point the runner's LoRA catalog at this repo.

Layout

catalog.json          # machine-readable registry (lora_catalog.json schema v1)
<ltx-version>/
  <lora-name>/
    lora.safetensors  # runnable adapter
    README.md         # per-LoRA notes (base, upstream, usage)

The catalog.json in the repo root is the machine-readable index: the video-creator runner (LTX-2.3) can consume it directly as LORA_CATALOG_SOURCE, or you can merge entries into the app's own catalog.

Included LoRAs

id	base	what it does	upstream
vbvr-video-reasoning	LTX-2.3	video reasoning (motion/interaction/causality)	Video-Reason/VBVR-LTX2.3-diffsynth

More adapters are added under their own 2.x/<name>/ folders as they're on-boarded.

Using in video-creator

Point the LTX worker at this repo's catalog:

LORA_CATALOG_SOURCE=https://huggingface.co/ad-astra-video/LTX-Loras/resolve/main/catalog.json

Then reference a LoRA by its catalog id (e.g. vbvr-video-reasoning) in the frontend / runner request. All files are public — no HF login required.