# LTX-2.5 WanGP runtime files

## Main transformers (repository root)

- `ltx-2.5-22b-dev_diffusion_model_bf16.safetensors`
- `ltx-2.5-22b-dev_diffusion_model_int8_convrot.safetensors`
- `ltx-2.5-22b-distilled_diffusion_model_bf16.safetensors`
- `ltx-2.5-22b-distilled_diffusion_model_int8_convrot.safetensors`
- `ltx-2.5-22b-distilled_diffusion_model_nvfp4.safetensors`

## Shared/offloadable components (repository root)

- `ltx-2.5-22b_video_embeddings_connector_bf16.safetensors`
- `ltx-2.5-22b_audio_embeddings_connector_bf16.safetensors`
- `ltx-2.5-22b_video_embeddings_connector_int8_convrot.safetensors`
- `ltx-2.5-22b_audio_embeddings_connector_int8_convrot.safetensors`
- `ltx-2.5-22b_video_embeddings_connector_nvfp4_bf16.safetensors`
- `ltx-2.5-22b_audio_embeddings_connector_nvfp4_bf16.safetensors`
- `ltx-2.5-22b_text_embedding_projection_bf16.safetensors`
- `ltx-2.5-22b_video_vae_bf16.safetensors`
- `ltx-2.5-22b_audio_vae_bf16.safetensors`
- `ltx-2.5-22b_vocoder_bf16.safetensors`
- `ltx-2.5-spatial-upscaler-x2-1.0_bf16.safetensors`
- `ltx-2.5-temporal-upscaler-x2-1.0_bf16.safetensors`

Dev and Distilled connector tensors are shared only after equality verification. The official NVFP4 transformer has different connector values, so it uses its own BF16 connector pair.

## Optional LTX-2.5 LoRA

- `ltx-2.5-22b-distilled-lora-450_bf16.safetensors`

This upstream LTX-2.5 LoRA is retained as an optional asset but is not automatically enabled by the WanGP LTX-2.5 defaults. LTX-2.0/2.3 LoRAs and their feature workflows remain isolated from LTX-2.5.

## Gemma 4 (`gemma4-12b-ltx-v1/`)

- `gemma4-12b-ltx-v1_bf16.safetensors`
- `gemma4-12b-ltx-v1_int8_convrot.safetensors`
- `config.json`
- `tokenizer.json`
- `tokenizer_config.json`
- `chat_template.jinja`

These files are derivatives of the official LTX-2.5 release and are distributed under `LICENSE-LTX-2.5.txt`.