Tencent's Z-Image 6B brings pixel-space generation to SD-WebUI

Tencent recently dropped Z-Image 6B, an image generation model that operates directly in pixel space—bypassing the traditional VAE (variational autoencoder) used by most diffusion models. It supports native 1K resolution without latent compression, aiming for sharper details and simpler architecture. Developer sangoi-exe (Reddit user isnaiter) quickly implemented support in his SD-WebUI-Codex extension, a custom webui for Stable Diffusion workflows. He also uploaded GGUF-quantized versions of Z-Image 6B and its Tencent-optimized variant to Hugging Face, making the model accessible on consumer hardware.

Early user feedback is mixed. While the concept is novel—generating pixels directly rather than decoding from latent space—the gen quality doesn't yet match leading models like SDXL or Flux. Commenters noted artifacts and inconsistent composition, especially at higher resolutions. Still, the integration is valuable for researchers and tinkerers: it provides a platform to test pixel-space approaches, compare with latent diffusion, and explore potential optimizations. The model's open-weight release and community tools lower the barrier for experimentation in this emerging subfield.