Z-Image Base-to-Turbo two-stage workflow balances variance and polish

A two-stage ComfyUI workflow chains Z-Image Base and Z-Image Turbo to combine the former's compositional variety with the latter's visual polish. The approach treats Z-Image Base as a high-variance seed generator, then feeds its output into Z-Image Turbo at reduced denoise strength to smooth anatomy and surface detail without flattening the composition.

Z-Image Base and Z-Image Turbo occupy opposite ends of the speed-quality spectrum. Base produces diverse layouts and poses but frequently renders malformed hands, asymmetric faces, and anatomical drift. Turbo converges faster and delivers cleaner surfaces, but its low step count and distilled sampling path leave little room for compositional exploration. The two-stage pattern attempts to exploit both: let Base explore the latent space freely, then hand the result to Turbo for a detail pass.

Denoise tuning

The workflow starts Z-Image Base at full denoise (1.0), generating a complete image from noise. The second stage runs Z-Image Turbo at 0.35 denoise, meaning it treats 35 percent of the latent as fresh noise and retains 65 percent of the Base output's structure. Users report grain and texture inconsistency when denoise climbs above 0.4 or drops below 0.15. The narrow working range suggests Turbo's distilled sampler expects near-converged input; too much noise and it reverts to its own low-variance priors, too little and it can't correct Base's anatomical errors.

The workflow uses the default Z-Image Turbo prompt template from ComfyUI, which includes quality tags and negative embeddings tuned for Turbo's training distribution. Swapping in a Base-optimized prompt for the first stage, or splitting the prompt into separate composition and detail clauses, may reduce grain. The two-model pattern is common in other distilled-turbo ecosystems—SDXL Base to SDXL Turbo, Pony to Pony Turbo—but denoise sweet spots vary by checkpoint pair.