Alibaba-pai Z-Image-Fun-Lora-Distill for Fast Images
Z-Image-Fun-Lora-Distill is a new LoRA adapter that speeds up image generation for the Z-Image model. It reduces the number of inference steps required while also handling CFG internally, making the generation process much simpler and faster.
Developed by Alibaba-pai, this tool helps users generate images more quickly without needing separate CFG calculations. The adapter is trained from scratch and works alongside other Z-Image LoRAs and control tools, giving users flexibility in their existing workflows.
Model Size: 0.57GB & VRAM GPU: requirements vary
Speed and compatibility features
- Reduces inference steps from 8 to as few as 2 steps.
- Eliminates the need for separate CFG calculations during generation.
- Compatible with ComfyUI through dedicated model versions.
- Works alongside ControlNet adapters for pose, depth, and edge detection
- Available in multiple versions optimized for different step counts.
The reduced step count means less waiting time per image, which can be useful for batch processing or testing multiple prompts. Users who rely on ControlNet workflows can continue using those tools with this adapter installed.
Technical details and recommended settings
The developers note that this model will slightly reduce output quality and may change the composition of generated images compared to the original Z-Image model. The latest 2603 versions introduce a random timesteps strategy that improves results when using sigmas below 0.500, addressing blurriness issues found in earlier 2602 releases.
According to the model card,
'The purpose of this model is to provide fast generation compatibility for Z-Image derivative models, not to replace Z-Image-Turbo.'
Users should set their LoRA weight between 0.7 and 0.9 for best results, with 0.8 being the recommended default. The model requires setting CFG to 1.0 since guidance is handled internally by the distillation process.
