LTX 2.3 Download: Every Official File Explained

The hardest part of running LTX 2.3 locally is not the GPU. It is figuring out which files you actually need.

Open the Hugging Face repository and you are looking at three 46 GB checkpoints, four upscalers, two distilled LoRAs, and a separate text encoder that lives in a completely different repo. Download the wrong combination and you either waste 150 GB of disk or get a pipeline that refuses to start because a "spatial upsampler" you never heard of is missing.

This guide maps the entire LTX 2.3 download surface: every official file, its real size, what it does, and the exact folder it belongs in. Every file name and size here was read directly off the official Lightricks/LTX-2.3 repository and the LTX-2 GitHub README, checked in July 2026. Nothing is estimated.

What this guide fixes

Most download posts give you one git clone line and move on. That works until the first pipeline error, because LTX 2.3 is not a single-file model. It is a set of components — transformer checkpoint, text encoder, upscaler, and (for most pipelines) a distilled LoRA — and different pipelines need different subsets. The download list that makes ComfyUI work is not the same list that makes the Python CLI work.

So this page is organized by what you plan to run, not by what happens to be in the repo.

The official LTX 2.3 file list (Hugging Face)

Everything below lives in Lightricks/LTX-2.3. The full repository is about 157 GB, which is why you should never clone it blindly.

Main checkpoints — pick one

They are the same size because distillation changes the weights, not the parameter count — all three are 22B models.

Distilled LoRAs

This is the file people most often miss. The official README states the distilled LoRA is required for current two-stage pipeline implementations — with the exception of DistilledPipeline, ICLoraPipeline, and LipDubPipeline.

Upscalers

If you search for ltx 2.3 spatial upscaler x2 1.1 safetensors and land here confused about why a 1 GB file matters more than the 46 GB one — that is the right instinct. Without it, the recommended two-stage pipelines will not run at all.

Text encoder — a different repository

LTX 2.3 does not ship its own text encoder. It uses Gemma 3 12B, pulled from Google's repo:

hf download google/gemma-3-12b-it-qat-q4_0-unquantized --local-dir models/gemma-3-12b

This is the single most common reason a first run fails. The --gemma-root argument is not optional in the CLI pipelines.

Download commands that actually work

The official quickstart from the LTX-2 repository is the minimum viable set — distilled model plus spatial upscaler:

git clone https://github.com/Lightricks/LTX-2.git
cd LTX-2

hf auth login
hf download Lightricks/LTX-2.3 \
    ltx-2.3-22b-distilled-1.1.safetensors ltx-2.3-spatial-upscaler-x2-1.1.safetensors --local-dir models/ltx-2.3
hf download google/gemma-3-12b-it-qat-q4_0-unquantized --local-dir models/gemma-3-12b

If you get a 401 or 403, the repo is gated: accept the model terms on Hugging Face and log in with a Read token. Fine-grained tokens need the "read gated repos" scope enabled — that detail is called out explicitly in the official README, and it trips up a lot of people who assume their token is fine because it works elsewhere.

Then generate:

uv run python -m ltx_pipelines.distilled \
    --distilled-checkpoint-path models/ltx-2.3/ltx-2.3-22b-distilled-1.1.safetensors \
    --spatial-upsampler-path    models/ltx-2.3/ltx-2.3-spatial-upscaler-x2-1.1.safetensors \
    --gemma-root models/gemma-3-12b \
    --seed 42 \
    --output-path output.mp4 \
    --prompt "..."

Want the higher-quality path instead? Add the dev checkpoint and the distilled LoRA, and run ltx_pipelines.ti2vid_two_stages. That is roughly 100 GB of downloads instead of 47 GB, which is the real trade-off nobody mentions up front.

Just want the output, not the download? You can generate with LTX 2.3 straight from the browser at ltx23.app — no 46 GB checkpoints, no gated-token errors, no Gemma directory. It is the fastest way to check whether the model does what your project needs before you commit disk space to it.

Where each file goes in ComfyUI

ComfyUI uses a different naming and folder convention than the Python package, and it expects the FP8 builds rather than the full bf16 safetensors. According to the official ComfyUI LTX-2.3 tutorial:

ComfyUI/models/checkpoints/
├── ltx-2.3-22b-dev-fp8.safetensors            # T2V, I2V, IA2V, ID-LoRA
└── ltx-2.3-22b-distilled-fp8.safetensors      # FLF2V, IC-LoRA

ComfyUI/models/text_encoders/
└── gemma_3_12B_it_fp4_mixed.safetensors

ComfyUI/models/latent_upscale_models/
└── ltx-2.3-spatial-upscaler-x2-1.1.safetensors

ComfyUI/models/loras/
├── ltx_2.3_22b_distilled_1.1_lora_dynamic_fro09_avg_rank_111_bf16.safetensors
├── gemma-3-12b-it-abliterated_lora_rank64_bf16.safetensors
├── ltx-2.3-22b-ic-lora-union-control-ref0.5.safetensors   # IC-LoRA workflow only
└── ltx-2.3-id-lora-talkvid-3k.safetensors                 # ID-LoRA workflow only

ComfyUI/models/geometry_estimation/
└── moge_2_vitl_normal_fp16.safetensors                    # IC-LoRA workflow only

Two things worth internalizing:

1. The upscaler goes in latent_upscale_models, not upscale_models. Different folder, easy mistake, silent failure.
2. No custom nodes are required. ComfyUI's own documentation states all LTX-2.3 workflows are built in — you just need a current build. If a tutorial tells you to install a custom node pack first, it is out of date.

Optional LoRAs worth downloading

Beyond the base set, Lightricks publishes task-specific LoRAs in separate repositories. These are small relative to the checkpoints and each unlocks a capability:

• LTX-2.3-22b-IC-LoRA-Union-Control — structural control from depth, pose, or edges
• LTX-2.3-22b-IC-LoRA-LipDub — lip dubbing against a reference clip
• LTX-2.3-22b-IC-LoRA-HDR — linear HDR output for EXR export
• LTX-2-19b-LoRA-Camera-Control-* — one LoRA per camera move (dolly in/out/left/right, jib up/down, static)

Note the version prefix: the camera-control and pose-control LoRAs are 19b LTX-2 era files, while union-control, lipdub, and HDR are 22b LTX-2.3 files. Mixing them is the second-most-common source of "it loaded but the output is garbage."

Disk planning, honestly

The practical answer for most people: start with the distilled quickstart. If the quality holds up for your use case, you never need the dev checkpoint at all.

FAQ

Where is the official LTX 2.3 download? huggingface.co/Lightricks/LTX-2.3 for weights, github.com/Lightricks/LTX-2 for the inference and training code. Anything else is a mirror.

Do I need a Hugging Face account? Yes. The repo is gated — you must accept the terms and use a read token with gated-repo access.

Is there an LTX 2.3 GGUF version? Community quantizations exist and are popular for low-VRAM setups, but they are not part of the official Lightricks repository. The official memory-reduction path is FP8 (--quantization fp8-cast) plus --offload cpu.

What about Civitai? Community LoRAs and workflow files show up there. Base checkpoints should come from Hugging Face — that is where the signed official files live.

Can I skip the spatial upscaler? Only if you use a single-stage pipeline. The recommended two-stage pipelines require it.

Bottom line

LTX 2.3 is not one download, it is four decisions: which checkpoint, which upscaler, distilled LoRA or not, and where the text encoder lives. Get those right and setup is uneventful. Get them wrong and you will spend an evening debugging errors that have nothing to do with your GPU.

If you would rather skip the 47 GB and see the model's actual output first, run a prompt through ltx23.app and download only after you know it fits your work.

Sources

All file names, sizes, and commands in this guide were verified against primary sources in July 2026:

1. Lightricks/LTX-2.3 — Hugging Face model repository — file list, sizes, model card
2. Lightricks/LTX-2 — official GitHub repository — quickstart, full model list, token requirements
3. LTX-2 pipelines: installation & CLI flags — required components, flag reference
4. LTX-2 pipelines: available pipelines — which pipeline needs which file
5. LTX-2 optimization guide — FP8 quantization policies
6. ComfyUI official LTX-2.3 tutorial — ComfyUI file names and folder paths
7. google/gemma-3-12b-it-qat-q4_0-unquantized — text encoder source
8. LTX-2.3-22b-IC-LoRA-Union-Control — structural control LoRA
9. LTX-2.3-22b-IC-LoRA-LipDub — lip dubbing LoRA
10. ComfyUI-LTXVideo — Lightricks' ComfyUI integration repo — official integration notes
11. LTX-2 technical paper (arXiv 2601.03233) — model architecture

Sizes are as published on Hugging Face at the time of writing; Lightricks ships new point releases regularly, so re-check the repository file list before a large download.