LTX 2.3 FLF2V

The LTX 2.3 KJ Stripped 1.1 workflow is a specialized audio-visual (AV) generation pipeline designed to interpolate between two static images. It utilizes a "Guided" architecture to ensure the generated video adheres strictly to a designated start and end point.

Core Architecture & Processing

• Model Loading: The workflow employs the LTX-Video 2.3 22B Distilled transformer model via a specialized KJ loader node.

• Dual-Modality Processing: It initializes both an EmptyLTXVLatentVideo and an LTXVEmptyLatentAudio space.

• AV Concatenation: Visual and audio latents are merged into a single latent stream using the LTXVConcatAVLatent node, allowing the sampler to process both simultaneously.

Frame Guidance System

• Start Frame Anchor: An image is loaded and preprocessed to serve as the reference for frame_idx: 0.

• End Frame Anchor: A second image is loaded and preprocessed to serve as the reference for the final frame (frame_idx: -1).

• Guide Application: The LTXVAddGuide nodes apply these images to the latent space with a configurable strength (set to 0.7 in this version) to dictate the video’s trajectory.

Sampling & Decoding

• KSampler Configuration: The workflow uses a high-speed 8-step Euler sampler with a simple scheduler.

• Separation & Decoding: After sampling, the LTXVSeparateAVLatent node splits the results back into distinct video and audio streams.

• Tiled Decoding: Visuals are decoded using VAEDecodeTiled to manage high-resolution output (1024x1024) efficiently.

Enhancement & Output

• NVIDIA RTX Integration: The workflow includes an RTXVideoSuperResolution node that provides a 2x hardware-accelerated upscale on the final image sequence.

• Final Assembly: The VHS_VideoCombine nodes produce two versions of the video—a base generation and a super-resolved version—both featuring synchronized, AI-generated audio.