LTX-2.3 Cinematic VAE
A custom fine-tuned VAE for LTX-Video 2.3 that bakes a subtle, darker greenish-teal "Matrix-style" cinematic color grade directly into decoder weights—eliminating post-processing color correction.
🎬 Side-by-Side Video Demonstration
Left half: Default LTX VAE Output | Right half: Cinematic VAE Output (Teal/Matrix Grade)
✨ Key Model Features
🎨 Baked-in Color Grade
Directly replaces standard decoding to output rich, moody teal/green tones without requiring LUTs or post-processing nodes.
⚡ Zero Latency Overhead
Identical layer architecture to standard LTX-Video VAE, adding zero performance or VRAM overhead during decoding.
🧩 ComfyUI & Diffusers Native
Drop-in replacement safetensors file for ComfyUI VAE loaders and Hugging Face Diffusers AutoencoderKLLTXVideo.
💻 Quick Integration Guide
1. ComfyUI Setup
Save LTX23_video_vae_bf16_cinematic.safetensors inside your ComfyUI models/vae folder. Use the KJ VAE Loader for monolithic checkpoints.
2. Hugging Face Diffusers Code snippet
from diffusers import AutoencoderKLLTXVideo, LTXPipeline
from huggingface_hub import hf_hub_download
import safetensors.torch
# 1. Load standard LTX VAE
vae = AutoencoderKLLTXVideo.from_pretrained("Lightricks/LTX-Video", subfolder="vae")
# 2. Download and load Cinematic VAE decoder weights
vae_path = hf_hub_download(repo_id="rzgar/LTX-2.3-Cinematic-VAE", filename="LTX23_video_vae_bf16_cinematic.safetensors")
state_dict = safetensors.torch.load_file(vae_path)
# Map decoder weights
decoder_dict = {k[len("decoder."):]: v for k, v in state_dict.items() if k.startswith("decoder.")}
vae.decoder.load_state_dict(decoder_dict, strict=False)
# 3. Initialize LTX Pipeline with Cinematic VAE
pipe = LTXPipeline.from_pretrained("Lightricks/LTX-Video", vae=vae, torch_dtype=torch.bfloat16).to("cuda")
video = pipe(
prompt="Cyberpunk detective walking through rainy neon streets, matrix teal cinematic tone",
num_inference_steps=25,
).frames[0]📦 Download Model Weights
File: LTX23_video_vae_bf16_cinematic.safetensors | Format: Safetensors (bf16)