Z-Image Turbo Finally Gets More Variety | Diversity LoRA + ComfyUI Workflow
A new workflow tackles Z-Image Turbo's biggest flaw: generating the same pose and angle repeatedly.
A developer has released a custom workflow for the popular ComfyUI platform that directly addresses a key limitation of the Z-Image Turbo image generation model. Z-Image Turbo is known for its speed but often produces images with frustratingly similar poses, camera angles, and compositions. The new solution integrates a specialized 'Diversity LoRA'—a small, fine-tuned adapter file—into the generation pipeline. This LoRA modifies the model's behavior to encourage greater variation without sacrificing the core quality or speed that makes Z-Image Turbo appealing.
The creator provides the complete ComfyUI workflow file for download, along with links to the required Z-Image Turbo model, the Diversity LoRA, and a VAE (Variational Autoencoder) for proper decoding. They demonstrate its effectiveness with two highly detailed, multi-sentence prompts designed to generate specific, cinematic portraits—one of a woman on a Mediterranean coast and another in a private jet cabin. The workflow allows users to test these prompts and see significantly more diverse outputs from the same setup. For those without a local GPU, the creator also suggests trying it via online services, making the tool more accessible.
- Integrates a 'Diversity LoRA' to combat Z-Image Turbo's repetitive output for poses and angles.
- Provides a full, downloadable ComfyUI workflow and all necessary model files (GGUF, LoRA, VAE) for immediate use.
- Demonstrates capability with complex, multi-clause prompts aimed at achieving specific, cinematic portrait styles.
Why It Matters
It provides a practical fix for a major flaw in a fast, open-source model, enhancing its utility for professional creative work.