Skip to content

⬡ LTX Loader

Loads LTX-2.3 video+audio models with Gemma 3 dual CLIP, video/audio VAEs, and optional spatial upscaler.

Inputs

Name Type Required Description
diff_model COMBO LTX-2.3 diffusion model (GGUF or safetensors)
clip_gemma COMBO Gemma 3 text encoder (e.g. gemma-3-12b-it-IQ4_XS.gguf)
clip_ltx COMBO LTX text projection / embeddings connector
video_vae COMBO LTX Video VAE (e.g. LTX2_video_vae_bf16.safetensors)
audio_vae COMBO LTX Audio VAE (e.g. LTX2_audio_vae_bf16.safetensors)
latent_upscale_model COMBO Spatial upscaler 2x for dual-pass pipeline

Outputs

Name Type Description
model_bundle UME_BUNDLE Complete LTX-2.3 bundle with model, dual CLIP, video VAE, audio VAE, and optional upscaler

How It Works

  1. Diffusion Model: Loads the LTX-2.3 transformer (supports GGUF quantized and safetensors)
  2. Dual CLIP: Loads Gemma 3 + LTX text projection as a combined text encoder (supports GGUF for both)
  3. Video VAE: Loads the LTX video VAE (bf16)
  4. Audio VAE: Loads the LTX audio VAE with prefix remapping (audio_vae.autoencoder.)
  5. Upscaler (optional): Loads the spatial 2x latent upscaler for the dual-pass pipeline

GGUF Support

Both the diffusion model and text encoders support GGUF quantized formats for reduced VRAM usage.

Dual-Pass Pipeline

When a latent upscale model is connected, the LTX Video Generator automatically uses a dual-pass pipeline: Pass 1 at half resolution → upscale 2x → Pass 2 at full resolution.

Alternative: Bundle Auto-Loader

If LTX-2.3 models are registered in model_manifest.json, you can use the ⬡ 📦 Bundle Auto-Loader instead for automatic download and loading.