What Is taeltx2_3.safetensors? The Complete Download & Install Guide for LTX 2.3

If you've tried to run LTX 2.3 in ComfyUI and hit an error about a missing VAE or TAE file, you've landed in the right place. taeltx2_3.safetensors is the one file that every LTX 2.3 workflow requires — no exceptions. This guide explains exactly what it is, where to get it, and how to install it correctly.

What Is taeltx2_3.safetensors?

taeltx2_3.safetensors is the Tiny AutoEncoder (TAE) for LTX 2.3. It's a lightweight encoder/decoder that converts between pixel space and the latent space used by the diffusion model during video generation.

Think of it this way: the main checkpoint (like ltx-2.3-22b-distilled.safetensors) handles the actual video generation in compressed latent space. The TAE is the translator — it encodes your input image or noise into latents at the start, and decodes the final latents back into visible video frames at the end.

Without taeltx2_3.safetensors, no LTX 2.3 workflow will run. It doesn't matter whether you're using the 16GB FP8 variant or the full 32GB official checkpoint — this file is always required.

TAE vs Standard VAE

LTX 2.3 introduced a new high-fidelity VAE architecture compared to earlier versions. The "TAE" (Tiny AutoEncoder) naming follows the convention from madebyollin's TAESD — a compact autoencoder that uses the same latent API as the full VAE but with a much smaller footprint.

The LTX 2.3 TAE specifically was trained on higher-quality data alongside the new 22B parameter model, which is why it produces noticeably sharper output than LTX 2.0/2.1:

• Hairlines and eyelashes hold together across frames instead of clumping
• Chrome edges and jewelry maintain tight highlights without ballooning
• Fabric textures (denim, linen) respect weave patterns without "texture drift"
• Gradients in skies and shadows show less banding

Where to Download taeltx2_3.safetensors

The file is hosted on HuggingFace by Kijai, who maintains the FP8 quantized variants for LTX 2.3:

Direct download page: https://huggingface.co/Kijai/LTX2.3_comfy

Look for taeltx2_3.safetensors in the file list. The file size is approximately 300–500 MB — much smaller than the main checkpoints (which range from 23 GB to 42 GB).

You can also find it linked from the official ComfyUI-LTXVideo repository by Lightricks.

Where to Place the File

Once downloaded, place taeltx2_3.safetensors in your ComfyUI VAE folder:

ComfyUI/
└── models/
    └── vae/
        └── taeltx2_3.safetensors   ← place it here

Do not place it in models/checkpoints/ or models/clip/. ComfyUI's VAE Loader node specifically looks in models/vae/.

Full File Checklist for LTX 2.3

Here's everything you need depending on your VRAM:

16GB VRAM (RTX 40xx+)

32GB VRAM

Note: LTX 2.3 also uses a Gemma 3 12B text encoder (gemma_3_12B_it_fp8_scaled.safetensors) placed in models/text_encoders/. This is separate from the TAE.

Step-by-Step Installation

1. Download the file

Go to Kijai/LTX2.3_comfy on HuggingFace and download taeltx2_3.safetensors.

2. Place it in the correct folder

# Move to ComfyUI VAE folder
mv taeltx2_3.safetensors /path/to/ComfyUI/models/vae/

3. Verify in ComfyUI

Open ComfyUI and add a VAE Loader node. Click the dropdown — you should see taeltx2_3 listed. If it doesn't appear, restart ComfyUI and check the folder path.

4. Connect it in your workflow

In any LTX 2.3 workflow, the VAE Loader node connects to both the LTXVModelLoader (for encoding) and the VAE Decode node (for decoding the final output).

Common Errors and Fixes

"VAE not found" or blank VAE dropdown → File is in the wrong folder. Move it to models/vae/, not models/checkpoints/.

"Unexpected keys in state dict" → You may have downloaded a file from an older LTX version. The taeltx2_3 file is specific to LTX 2.3 — don't reuse the VAE from LTX 2.0 or 2.1.

Out of memory during decode → The TAE is lightweight, but if you're running on 12GB VRAM with a large batch, enable sequential offloading in ComfyUI settings.

Video output looks blurry or washed out → Double-check you're using taeltx2_3.safetensors and not an older LTX VAE. The new TAE is specifically trained for LTX 2.3's latent space.

Why This File Matters for Video Quality

The LTX 2.3 TAE was rebuilt from scratch alongside the 22B parameter model. According to Lightricks' release notes, the new VAE produces:

• Sharper fine details across the full frame
• Better texture consistency between frames (less "texture drift")
• Cleaner gradients with less banding
• Improved facial detail retention during motion

This is a meaningful upgrade from LTX 2.0/2.1. If you're running an older VAE with the new checkpoint, you'll get degraded output — the latent spaces are not compatible.

Summary

Once this file is in place, you're ready to load any LTX 2.3 workflow. Check out our Models page for direct download links to all LTX 2.3 checkpoints, or the Setup Guide for the full installation walkthrough.

Sources

• LTX-2.3 in ComfyUI Workflow — AI Study Now
• LTX-2.3: What's New in Lightricks' 22B Video Model (2026) — WaveSpeedAI
• LTX-2.3 ComfyUI Setup: Two-Stage Pipeline, VRAM Fixes & Gemma Encoder — WaveSpeedAI
• Kijai/LTX2.3_comfy — HuggingFace
• Lightricks/ComfyUI-LTXVideo — GitHub
• madebyollin/taesd: Tiny AutoEncoder for Stable Diffusion — GitHub