Use Chroma to set the composition of Z-Image with the split sigma technique

A new AI image generation technique called the split sigma workflow solves persistent composition problems in popular models. The method specifically addresses Z-Image Turbo's weakness in creating varied, dynamic compositions while maintaining its superior realism. By using Chroma (or any Flux VAE-compatible model) for the initial 50-step composition phase, then switching to Z-Image Turbo for detailing, creators can achieve results that combine the best of both models.

The technique works because any model using the Flux VAE is latent compatible, meaning Z-Image Turbo can finish latents started by Chroma, Flux.1 Dev, or other variants. This three-stage process includes a composition sampler, refinement sampler, and detailing stage that gives creators precise control over how detail is added. The same approach works with Flux.2 VAE models, allowing Flux.2 Dev to set composition while Flux.2 Klein 9B acts as detailer.

This represents a significant advancement in hybrid model workflows, substantially increasing the world knowledge available during image sampling. The technique enables compositions that wouldn't be possible with any single model alone, whether using LoRAs or not. Early tests show marked improvements in dynamic range, contrast, prompt adherence, and text rendering compared to the "ZIT look" that previously characterized Z-Image Turbo outputs.