LTX 2.3 Examples

This example performs two-stage image-to-video generation with the LTX 2.3 22B checkpoint and writes one MP4 containing generated video and audio.

Model Source

Model	HuggingFace	ModelScope	Purpose
LTX 2.3 22B	Lightricks/LTX-2.3	Lightricks/LTX-2.3	Main transformer, spatial upsampler, and distilled stage-two LoRA
Gemma 3 12B	google/gemma-3-12b-it	N/A	Text encoder

Feature Support

Feature	Support	Notes
Image-to-video with audio	Supported	Produces an MP4 with an AAC audio track
Two-stage denoising	Supported	Base generation, latent upsampling, then refinement
Multi-GPU inference	Supported	Ulysses sequence parallelism and FSDP through --gpu_num
LoRA	Supported	Uses the stage-two distilled LoRA when present
VAE parallelism	Supported	Enabled with multi-GPU execution
Ray VAE actor	Unsupported	The checked-in script does not expose this mode
Server API	Unsupported	The script is an offline CLI entry point

Requirements

• GPU: two CUDA GPUs by default; use another supported count through --gpu_num
• Software: the standard TeleFuser installation and an ffmpeg executable on PATH
• Input assets: a PIL-readable source image supplied with --image_path

Install TeleFuser by following the development setup.

Model Directory

${TF_MODEL_ZOO_PATH}/
|-- LTX-2.3/
|   |-- ltx-2.3-22b-dev.safetensors
|   |-- ltx-2.3-spatial-upscaler-x2-1.0.safetensors
|   \-- ltx-2.3-22b-distilled-lora-384.safetensors
\-- gemma-3-12b-it-qat-q4_0-unquantized/
    |-- model-00001-of-00005.safetensors
    |-- ...
    \-- model-00005-of-00005.safetensors

export TF_MODEL_ZOO_PATH=/path/to/model_zoo

Quick Start

TELEAI_EXAMPLE_OUTPUT_DIR=work_dirs \
python examples/ltx_video/ltx23_22b_image_to_video_two_stage_h100.py \
  --model_root "$TF_MODEL_ZOO_PATH/LTX-2.3" \
  --image_path /path/to/input.png \
  --prompt "A slow camera move through a sunlit garden."

The command writes work_dirs/ltx23_22b_image_to_video_two_stage_h100_2gpu.mp4.

Examples

Two-Stage Image-To-Video

ltx23_22b_image_to_video_two_stage_h100.py

TELEAI_EXAMPLE_OUTPUT_DIR=work_dirs \
python examples/ltx_video/ltx23_22b_image_to_video_two_stage_h100.py \
  --gpu_num 4 \
  --model_root "$TF_MODEL_ZOO_PATH/LTX-2.3" \
  --image_path /path/to/input.png \
  --resolution 1080p \
  --prompt "A cinematic outdoor scene with natural motion."

Key options:

Option	Default	Description
--gpu_num	2	Number of inference workers
--image_path	Legacy local sample path	Pass an existing image explicitly
--model_root	${TF_MODEL_ZOO_PATH}/LTX-2.3	LTX checkpoint directory
--resolution	1080p	One of 720p, 1080p, 2k, or 4k
--num_inference_steps	30	Steps in each denoising stage
--num_frames	121	Generated frame count
--seed	42	Random seed

Notes

• The default input asset is not packaged, so pass --image_path explicitly.
• The stage-two LoRA is optional at load time but recommended for the intended output quality.