Z-Image-Fun-Lora-Distill

Model Card

a. 2603 Models

Name	Description
Z-Image-Fun-Lora-Distill-2-Steps-2603.safetensors	A Distill LoRA for Z-Image that distills both steps and CFG. It requires only 2 steps instead of 8. Due to the random timesteps strategy, it is better adapted to sigmas below 0.500. The recommended sigma for the second step is between 0.800 and 0.500. A larger LoRA strength is recommended.
Z-Image-Fun-Lora-Distill-2-Steps-2603-ComfyUI.safetensors	ComfyUI version of Z-Image-Fun-Lora-Distill-2-Steps-2603.safetensors
Z-Image-Fun-Lora-Distill-4-Steps-2603.safetensors	A Distill LoRA for Z-Image that distills both steps and CFG. It requires only 4 steps instead of 8 steps. Due to the addition of a random timesteps strategy, it is better adapted to cases where sigmas are less than 0.500.
Z-Image-Fun-Lora-Distill-4-Steps-2603-ComfyUI.safetensors	ComfyUI version of Z-Image-Fun-Lora-Distill-4-Steps-2603.safetensors
Z-Image-Fun-Lora-Distill-8-Steps-2603.safetensors	A Distill LoRA for Z-Image that distills both steps and CFG. Compared to Z-Image-Fun-Lora-Distill-8-Steps-2602.safetensors, due to the addition of a random timesteps strategy, it is better adapted to cases where sigmas are less than 0.500.
Z-Image-Fun-Lora-Distill-8-Steps-2603-ComfyUI.safetensors	ComfyUI version of Z-Image-Fun-Lora-Distill-8-Steps-2603.safetensors

Model Features

• This is a Distill LoRA for Z-Image that distills both steps and CFG. It does not use any Z-Image-Turbo related weights and is trained from scratch. It is compatible with other Z-Image LoRAs and Controls.
• This model will slightly reduce the output quality and change the output composition of the model. For specific comparisons, please refer to the Results section.
• The purpose of this model is to provide fast generation compatibility for Z-Image derivative models, not to replace Z-Image-Turbo.

Results

The difference between the 2603 version model and the 2602 version model

The 2602 model tends to produce blurry images with sigmas below 0.500, as the distillation model was not trained on certain steps. The 2603 model introduces a random timesteps strategy, making it better adapted to sigmas below 0.500.

As shown below, when using kl_optimal, many sigmas fall below 0.500. The 2603 model handles these cases correctly, while the 2602 model does not. Note that although kl_optimal is used in the figure, we still recommend using the simple scheduler for inference.

Reprinted from ModelScope