docs: add LTX-2.5 bundle card and licenses

cecfdc1 verified about 2 months ago

3.76 kB

license: other
license_name: ltx-2-community-license-agreement
license_link: https://github.com/Lightricks/LTX-2/blob/v1.2.0/LICENSE.md
base_model:
  - Lightricks/LTX-2.5
  - google/gemma-4-12B-it
library_name: mlx
pipeline_tag: text-to-video
tags:
  - mlx
  - candle
  - apple-silicon
  - cuda
  - ltx-2
  - ltx-video
  - text-to-video
  - image-to-video
  - audio-video-generation

LTX-2.5 — turnkey SceneWorks MLX and Candle bundle

This is the pinned, token-free model bundle used by SceneWorks for native LTX-2.5 generation. It contains the complete split LTX-2.5 pipeline at q4, q8, and bf16, both distilled and guided dev transformers, the shipped distilled rank-450 adapter, and the separately staged stock Gemma 4 prompt enhancer. A fresh SceneWorks install downloads these ready-to-run files; no in-app conversion or access to the gated source repositories is required.

Repository layout

path	contents
`distilled/q4/`, `distilled/q8/`, `distilled/bf16/`	Complete distilled LTX-2.5 tier: transformer, self-contained Gemma 4 text encoder, connector, Conv VAE, DiffVAE, spatial and temporal upsamplers, duration head, audio VAE, vocoder, and split manifests.
`dev/q4/`, `dev/q8/`, `dev/bf16/`	The same complete tier with the guided 30-step dev transformer. Distilled remains the SceneWorks default.
`distilled_lora/`	Upstream LTX-2.5 distilled LoRA, with rank and alpha read from its safetensors metadata.
`enhancer/`	Stock `google/gemma-4-12B-it` snapshot used only for opt-in T2V/I2V prompt enhancement. The tier-local self-contained text encoder remains the generation encoder.
`licenses/`	LTX-2.x Community License Agreement (August 11, 2026), Apache License 2.0, the Gemma Prohibited Use Policy, and provenance notice.

Each tier is an indivisible pipeline contract: q4/q8/bf16 selects the matching representation for every quantizable segment. The tier's split_model.json records component identities and quantization; SceneWorks rejects incomplete or mismatched layouts.

Provenance and modifications

LTX-2.5 components and the distilled LoRA derive from Lightricks/LTX-2.5 and the LTX-2 v1.2.0 reference. SceneWorks converted and, where selected, quantized the components into the native split MLX/Candle layout; generative model behavior was not fine-tuned.
The dev transformer derives from the LTX-2.5 dev checkpoint and is packed through the same converter into q4/q8/bf16.
enhancer/ mirrors the stock google/gemma-4-12B-it snapshot for offline-only prompt enhancement.
Hugging Face extra_gated_* repository metadata is intentionally absent. Removing the hosting-service click-through does not alter any license or use restriction included here.

Licenses

This repository contains components governed by separate terms:

LTX-2.5 weights, derivatives, and the distilled adapter: LTX-2.x Community License Agreement, license date August 11, 2026. The agreement includes use restrictions and commercial licensing conditions and must travel with redistributed derivatives.
Gemma 4 text-encoder/enhancer weights: Apache License 2.0, with Google's Gemma Prohibited Use Policy also reproduced here.
NOTICE records the exact sourcing and the material difference from the Gemma 1–3 Terms of Use used by the older LTX-2.3 bundle.

Downloading, using, modifying, or redistributing a component requires compliance with all terms applicable to that component.