LTX Video 2.3 — I2V Two-Pass with Audio

Model: LTX Video 2.3 — 22B Distilled (FP8) Type: Image-to-Video

Two-pass Image-to-Video workflow using LTX Video 2.3 with audio generation support. Takes two guide images (start and end frame) and generates a video in a 720p first pass, then upscales to 1080p in a second pass. Audio is encoded via the LTX 2.3 audio VAE and composited into the final output.

Preview

Click to open on YouTube

Requirements

ComfyUI: recent stable build
Model: ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors — Kijai/LTX2.3_comfy
Video VAE: LTX23_video_vae_bf16.safetensors — Kijai/LTX2.3_comfy
Audio VAE: LTX23_audio_vae_bf16.safetensors — Kijai/LTX2.3_comfy
Preview VAE: taeltx2_3.safetensors — Kijai/LTX2.3_comfy
Text encoder: gemma_3_12B_it_fp8_scaled.safetensors — Kijai/LTX2.3_comfy
Text projection: ltx-2.3_text_projection_bf16.safetensors — Kijai/LTX2.3_comfy
Custom nodes:

Notes

The workflow uses two guide images (start and end frame) via LTXVAddGuideMulti. Image 1 is placed at frame 0, Image 2 at frame 161 by default.
Pass 1: 1280x720, 241 frames @ 24fps, 8 steps, lcm sampler, linear_quadratic scheduler, CFG 1. NAG guidance is enabled (scale 11, rescale 0.25).
Pass 2: The output is upscaled to 1920x1080 with nvidia_rtx_vsr via ImageResizeKJv2, then re-encoded and sampled again with euler_ancestral sampler for 8 steps.
An audio file can be supplied via VHS_LoadAudioUpload — the audio is encoded with the LTX 2.3 audio VAE and blended into the latent space during both passes.
A LoRA (ltx-2.3-22b-distilled-lora-dynamic) is present in the graph but disabled — enable it if you want dynamic motion tuning.
LTX2MemoryEfficientSageAttentionPatch, LTXVChunkFeedForward, and ModelPatchTorchSettings are used for VRAM optimization. Requires SageAttention installed.

Changelog

2026-03-30 — Initial upload

Support

Producing and sharing this kind of open-source work requires renting cloud GPUs, which gets expensive quickly. If you find it useful and would like me to keep contributing, your support is very much appreciated: