LTX 2.3 three-stage sampling: correct as you upscale, and why every pass re-injects the image

The LTX 2.3 three-stage sampling method generates a high-quality, upscaled video while preserving the identity of the subject throughout the entire process. Unlike a normal upscale, this method enlarges the image while simultaneously correcting it. After each latent upscale, the image conditioning is reapplied through an image-to-video node, ensuring the face, composition, and reference style remain consistent at the new resolution.

This allows a distillation-only flow to maintain identity across three render passes instead of drifting.

The LTX 2.3 pipeline consists of 145 nodes, including three near-identical sampling assemblies. It requires a low-resolution latent canvas, positive prompt, audio-latent separation, and image-to-video node for each pass. The first pass builds motion on a low-resolution latent canvas, while subsequent passes double the latent and re-render with the upscale model's sigma. Each pass takes about three actual steps per refinement. The output resolution typically reaches 1080p, with the final video being 2048x1024.

Two crucial knobs in the process are the boolean in the image-to-video node, which turns the flow into text-to-video when flipped, and the 18-value blur, which blurs the reference image to prevent the conditioning from being too literal. The caveat is that if the flow is reduced to one or two passes, the 18 value can make the result more static, as the blur is not corrected by later refinement passes. Therefore, if the pass count is changed, the 18 value should be re-tuned.

Compared to two-stage sampling and single-stage-plus-spatial-upscale, this method re-conditions the image after each enlargement, rather than merely enlarging the image. The current upscaler in the official LTX 2.3 release has been updated to 1.1, which is recommended for local use, particularly for long clips. The shipped JSON still references the older 1.0 version. The graph was adapted from a Reddit user's workflow with the author's own modifications.

However, there are some limitations to this method. The shipped JSON uses an outdated upscaler version (1.0), which is behind the recommended 1.1. Additionally, there is no negative-prompt path yet at CFG 1. The 145-node architecture means three duplicated sampling blocks, making it more complex to maintain than a single-pass graph. Lastly, the 18-value blur requires re-tuning if the pass count is altered.