Created my own 6 step sigma values for ltx 2.3 that go with my custom workflow that produce fairly cinematic results, gen times for 30s upscaled to 1080p about 5 mins.
After 40 hours of iteration, a custom sigma chain produces cinematic AI video in record time.
An independent AI video creator has publicly shared a highly optimized workflow for the LTX 2.3 model, the result of an intensive 40-hour development sprint. The core innovation is a custom chain of six sigma values—0.9, 0.7, 0.5, 0.3, 0.1, and 0—that guide the model's denoising process during video generation. This specific progression, described as "spinning the sigma wheel," was iterated upon to find the optimal balance between rendering speed and visual fidelity, culminating in a process that can produce a 30-second video upscaled to 1080p resolution in approximately five minutes.
The creator's workflow, detailed on Pastebin, represents a significant practical advancement for the AI video community. By fine-tuning the often-opaque sigma parameters, which control the step-by-step noise removal in diffusion models, they've demonstrated a path to faster, more controllable generation without sacrificing the desired "cinematic" quality. This move from abstract experimentation to a shareable, reproducible pipeline lowers the barrier to entry for high-quality output. It provides a concrete starting point for other creators using LTX 2.3, enabling them to achieve professional results more efficiently and sparking further community-driven optimization.
- Custom 6-step sigma chain (0.9, 0.7, 0.5, 0.3, 0.1, 0) optimized for LTX 2.3 model.
- Generates a 30-second, 1080p upscaled AI video in just 5 minutes, a major speed boost.
- Workflow is the result of 40 hours of iterative testing to balance speed and cinematic quality.
Why It Matters
Democratizes high-quality AI video creation by providing a fast, reproducible pipeline for the LTX 2.3 community.