Z-Image 6B pixel-space generation runs in ComfyUI at 30s per 1024x1024

A Reddit user (FitContribution2946) has published an experimental ComfyUI workflow for Z-Image 6B / L2P pixel-space generation. The workflow, built with help from Codex and adapted from an existing Hidream 01 setup, requires installing a custom node. On an NVIDIA 4090, it produces a 1024x1024 image at 30 steps in roughly 30 seconds. The creator emphasizes this is a proof of concept — not an optimized or production-ready implementation — but provides it as a working foundation for the community to explore. The GitHub repository (gjnave/ggf-ltp-zimage) includes basic instructions and the custom node code.

This release is significant because it brings Z-Image 6B, a pixel-space generation model, into the widely used ComfyUI ecosystem. Pixel-space models differ from latent diffusion models by generating images directly in pixel space, potentially offering different trade-offs in quality and behavior. While still experimental, this workflow lowers the barrier for AI artists and developers to experiment with Z-Image 6B without needing to build a pipeline from scratch. The open-source nature invites collaboration and optimization, which could accelerate the model's adoption in creative workflows. Future iterations may improve speed, stability, and integration with other ComfyUI nodes.