LTX-2.5 distilled or dev transformer: which file to download

For generating clips, download the distilled bf16 transformer from the LTX-2.5 repository; the card describes it as a fixed 8-step schedule with CFG 1. Use the dev transformer when you want to train or fine-tune, because the card calls it the full, trainable DiT. The int8 files are for ComfyUI only, and the NVFP4 file needs Blackwell hardware or ComfyUI. Sume's catalog has no LTX id, so this choice only matters if you self-host.

Everything about the files comes from the LTX-2.5 model card, read 2026-10-02. Sume's side is from Video generation.

What is in the LTX-2.5 checkpoint pack?

The card says LTX-2.5 ships as a split pack with one safetensors file per component, not one monolith, and that each CLI flag or loader points at one file. That is why the download command lists a transformer, a text encoder, two VAEs, a duration head and a spatial upscaler separately.

The text encoder is gemma4-12b-with-proj-ltx-2.5-bf16.safetensors, a Gemma 4 12B encoder plus projections. There are two video VAEs: the card calls the DiffVAE higher quality and heavier, and the Conv VAE faster and lighter.

What do I download to run the distilled model?

The card's distilled split pack is below. It needs a Hugging Face login first, and the repository is gated, as this post on the gate explains. If you run short of VRAM, the card suggests --quantization fp8-cast and --offload cpu, and it says to use the bf16 files with ltx-pipelines because the int8 files do not load there.

hf auth login
hf download Lightricks/LTX-2.5 \
  diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors \
  text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors \
  vae/ltx-2.5-video-vae-bf16.safetensors \
  vae/ltx-2.5-audio-vae-bf16.safetensors \
  model_patches/ltx-2.5-duration-head-bf16.safetensors \
  latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors \
  --local-dir models/ltx-2.5

Does any of this apply on Sume?

No checkpoint choice exists on Sume. A Video Router job names a catalog id such as seedance-2.5 or wan-3.0, and the request schema reserves model_params for per-model knobs with an empty allowlist in v1. The OpenAPI model enum has no LTX row, so run LTX-2.5 yourself or call a model Sume lists. Read GET /v1/video-router/models for the current catalog.