---
license: other
license_name: ltx-2-community-license-agreement
license_link: https://github.com/Lightricks/LTX-2/blob/main/LICENSE.md
tags:
  - image-to-video
  - ltx-2.5
---

# LTX-2.5 serving bundle

A **mirror**, not a new model. Every weight file here is copied byte-for-byte
(server-side, via `huggingface_hub.copy_files`) from
[`Lightricks/LTX-2.5`](https://huggingface.co/Lightricks/LTX-2.5) and is
governed by the **LTX-2.x Community License** linked above. All credit to
Lightricks.

It exists because RunPod Serverless' model-caching feature allows exactly one
cached repo per endpoint and cannot filter which files it pulls — naming the
upstream repo directly would pre-provision all ~299 GB of it, including
ComfyUI-only int8 variants this deployment never loads. This repo is the
subset `ltx-pipelines` actually needs, and nothing else.

## Contents (140.75 GB)

| File | GB | Used by |
|---|---|---|
| `diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors` | 42.02 | `DistilledPipeline` (image-to-video, fast path) |
| `diffusion_models/ltx-2.5-22b-distilled-transformer-nvfp4.safetensors` | 18.72 | same, `--quantization nvfp4-prequant` (Blackwell SM>=10) |
| `diffusion_models/ltx-2.5-22b-dev-transformer-bf16.safetensors` | 42.02 | `KeyframeInterpolationPipeline`, `TI2VidTwoStages*Pipeline` |
| `loras/ltx-2.5-22b-distilled-lora-450-bf16.safetensors` | 8.90 | required by both dev-transformer pipelines above |
| `text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors` | 26.26 | all pipelines |
| `vae/ltx-2.5-video-vae-bf16.safetensors` | 1.47 | all pipelines |
| `vae/ltx-2.5-audio-vae-bf16.safetensors` | 0.36 | all pipelines (LTX generates audio natively) |
| `latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors` | 1.00 | stage-2 spatial upsampler |

Deliberately excluded: `*-comfy-int8-convrot.safetensors` (ComfyUI-only,
`ltx-pipelines` cannot read them), `vae/ltx-2.5-video-vae-conv-bf16` (faster,
lower-quality alternative), `latent_upscale_models/*temporal*` and
`model_patches/ltx-2.5-duration-head-bf16` (auto-duration is unwanted where
billing is per requested duration).

## Notes

- Resolution must be divisible by **64** for two-stage pipelines
  (`assert_resolution(..., is_two_stage=True)`) — the model card's "divisible
  by 32" is the one-stage rule.
- Frame count must satisfy `num_frames % 8 == 1`.
- LTX-2.3 files are **not** interchangeable with these, and a LoRA only works
  with the model it was trained on.