LTX 2.5 Download: The Gated Repo, Split Files, and Where They Go

LTX-2.5 does not download like LTX 2.3. 2.3 shipped as single bundled checkpoints — transformer, VAEs, and text projection in one file. LTX-2.5 is a split, Comfy-aligned pack: one .safetensors per component, and you only download the parts your pipeline needs. The official Quick Start set is roughly 66 GiB; the full repo carries more, including separate dev and distilled transformers.

There is also a new gate: the Lightricks/LTX-2.5 repository is gated on Hugging Face. Downloads fail with a 401/403 until you accept the license and your access request is approved. That gate, plus the split layout, is why "LTX 2.5 download" is its own skill rather than a repeat of 2.3.

This guide covers access, the complete official file list, what the Quick Start set contains, and where each file goes. Sourced from the LTX-2.5 model card, the LTX-2 GitHub README, and the ComfyUI LTX-2.5 tutorial. Checked August 13, 2026.

Step 1: Request access to the gated repo

The repository requires an access request before any file downloads:

1. Open Lightricks/LTX-2.5 on Hugging Face
2. Agree to the license terms (the LTX-2.x Community License)
3. Wait for approval

For the CLI path, the GitHub README adds two specifics: log in with hf auth login, and use a Read token — fine-grained tokens need the "read gated repos" scope enabled. If you get a 401/403, the README's diagnosis is that you either haven't accepted the terms or your token lacks the gated-repo scope.

There are actually three LTX-2.5 repositories on Hugging Face:

Most people only need the first one. The Pre-Trained repo exists for fine-tuning foundations, and the Diffusers repo for Hugging Face diffusers users (support is not in a released diffusers version yet — the official instruction is to install from main).

Step 2: Download the Quick Start set (~66 GiB)

The official Quick Start downloads five files into a models/ltx-2.5 folder:

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 \
  latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors \
  --local-dir models/ltx-2.5

That set runs the DistilledPipeline text-to-video example from the README. Note that the distilled pipeline also expects the spatial upscaler — the model card's download example additionally pulls the upscaler from the LTX-2.3 repo (ltx-2.3-spatial-upscaler-x2-1.1.safetensors) for the distilled two-stage path, while the newer ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0 lives in the 2.5 repo. Both upscaler files exist; check which one your pipeline expects.

Step 3: The complete official file map

The model card lists the full pack by component:

Transformers (DiT) — choose one per use case:

• ltx-2.5-22b-distilled-transformer-bf16.safetensors — distilled DiT, fixed 8-step schedule, CFG=1 (what DistilledPipeline, ICLoraPipeline, and DubItPipeline expect)
• ltx-2.5-22b-dev-transformer-bf16.safetensors — full trainable DiT (used by the guided two-stage pipelines)
• ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors — ComfyUI only int8 build of the distilled DiT
• ltx-2.5-22b-dev-transformer-comfy-int8-convrot.safetensors — ComfyUI only int8 build of the dev DiT
• ltx-2.5-22b-distilled-transformer-nvfp4.safetensors — NVFP4 build for Blackwell GPUs (ComfyUI, or ltx-pipelines with --quantization nvfp4-prequant + ltx-kernels)

Text encoder (required by every pipeline):

• gemma4-12b-with-proj-ltx-2.5-bf16.safetensors — the LTX-tuned Gemma 4 12B with text projection bundled
• gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors — ComfyUI int8 variant

VAEs:

• ltx-2.5-video-vae-bf16.safetensors — diffusion decoder (DiffVAE): higher quality, heavier
• ltx-2.5-video-vae-conv-bf16.safetensors — convolutional decoder: faster, lighter
• ltx-2.5-audio-vae-bf16.safetensors — audio VAE + vocoder

Upscalers:

• ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors — x2 spatial upscaler (required by multi-stage pipelines)
• ltx-2.5-latent-temporal-upscaler-x2-bf16-1.0.safetensors — x2 temporal upscaler (for DFR temporal refine rounds)

Adapters and heads:

• ltx-2.5-22b-distilled-lora-450-bf16.safetensors — distilled LoRA, required by two-stage pipelines that run the full model in stage 1
• ltx-2.5-duration-head-bf16.safetensors — optional duration head; omit --num-frames to let it predict clip length
• ltx-2.5-22b-ic-lora-pixel-spatial-upscaler-x2-1.0.safetensors — optional detailing IC-LoRA for DFR refinement, hosted in its own separate repo

Step 4: Where files go (ltx-pipelines vs ComfyUI)

ltx-pipelines path — load by flag, no fixed folder convention:

uv run python -m ltx_pipelines.distilled \
  --transformer-path     models/ltx-2.5/diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors \
  --text-encoder-path    models/ltx-2.5/text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors \
  --video-vae-path       models/ltx-2.5/vae/ltx-2.5-video-vae-bf16.safetensors \
  --audio-vae-path       models/ltx-2.5/vae/ltx-2.5-audio-vae-bf16.safetensors \
  --duration-head-path   models/ltx-2.5/model_patches/ltx-2.5-duration-head-bf16.safetensors \
  --spatial-upsampler-path models/ltx-2.3/ltx-2.3-spatial-upscaler-x2-1.1.safetensors \
  --prompt "A golden retriever running through a sunny meadow, cinematic lighting" \
  --seed 42 \
  --output-path output_distilled.mp4

Use the bf16 checkpoints here. The *-comfy-int8-convrot files are explicitly ComfyUI-only and are not loaded by the PyTorch path.

ComfyUI path — fixed folder layout from the official tutorial:

ComfyUI/
└── models/
    ├── diffusion_models/   → int8-convrot distilled transformer
    ├── text_encoders/      → gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot
    │                        + gemma4_e2b_it_bf16 (from Comfy-Org/gemma-4)
    ├── vae/                → ltx-2.5-video-vae-bf16 + ltx-2.5-audio-vae-bf16
    └── latent_upscale_models/ → ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0

What not to download

• 2.3 checkpoints won't work with 2.5 pipelines — the README is explicit that files are not interchangeable between the two models, and the encoder is version-checked (gemma4-12b-ltx-v1). Don't mix.
• Stock Gemma 4 is not a substitute for the LTX-tuned encoder — the README warns the stock release will fail the version check.
• The Pre-Trained bundle (43 GiB single file) is for research and fine-tuning, not for running the released 2.5 generation — it predates supervised fine-tuning.

Don't want to wait for gated access and 66 GiB of downloads? Generate with LTX 2.5 right now in your browser at ltx23.app — no approval queue, no disk space.

FAQ

How big is the LTX 2.5 download? The official Quick Start set is roughly 66 GiB. Full packs with dev + distilled + int8 + NVFP4 variants are larger; you only download the components your pipeline needs.

Why do my downloads return 401/403? The repo is gated. Accept the license on the HF page and wait for approval; for CLI use a Read token with the "read gated repos" scope.

Which transformer should I download? Distilled for fast generation (8-step, CFG=1) — used by DistilledPipeline, ICLoraPipeline, and DubItPipeline. Dev (full trainable) for the guided two-stage pipelines and fine-tuning. ComfyUI workflows use the int8-convrot builds.

Do I need both video VAEs? No. The diffusion decoder (DiffVAE) is higher quality and heavier; the conv VAE is faster and lighter. The official Quick Start and ComfyUI workflows use the diffusion decoder.

Where do the LTX-2.5 files go in ComfyUI? diffusion_models/, text_encoders/ (two encoders), vae/ (video + audio), and latent_upscale_models/. Exact layout in the official tutorial.

Bottom line

LTX 2.5 download = request gated access first, then pull only the components you need from the split pack (~66 GiB Quick Start set), and place them per your path: flag-based for ltx-pipelines with bf16 files, fixed folders for ComfyUI with the int8-convrot builds. Nothing from 2.3 carries over. In a hurry? Skip the queue entirely and generate at ltx23.app.

Sources

Verified August 13, 2026 against primary documentation:

1. Lightricks/LTX-2.5 — Hugging Face model card — full split file map, gated repo, download commands, ComfyUI-only int8 files
2. Lightricks/LTX-2 — GitHub README — Quick Start (~66 GiB), 401/403 causes, token scopes, file-interchange warning, pipeline file expectations
3. ComfyUI official LTX-2.5 tutorial — ComfyUI file storage layout, second encoder (gemma4_e2b_it_bf16), gated-repo warning
4. Lightricks/LTX-2.5-Pre-Trained — Hugging Face — 43 GiB pre-training checkpoint, intended use
5. Lightricks/LTX-2.5-Diffusers — Hugging Face — Diffusers packaging, install-from-main requirement

Licensing summaries are not legal advice. Read the LTX-2.x Community License in the repository before downloading.