---
title: Z-Image-Fun-Distill-ComfyUI
canonical_url: "https://www.modelscope.cn/models/AI-ModelScope/Z-Image-Fun-Distill-ComfyUI"
md_url: "https://www.modelscope.cn/models/AI-ModelScope/Z-Image-Fun-Distill-ComfyUI.md"
repository: AI-ModelScope/Z-Image-Fun-Distill-ComfyUI
last_updated: 2026-02-11
license: apache-2.0
base_model:
  - Tongyi-MAI/Z-Image
base_model_relation: adapter
parameters: 260.0M
tensor_type:
  - BF16
library_name:
  - lora
  - diffusers
  - safetensors
  - pytorch
frameworks:
  - pytorch
supports_inference: img2img
downloads: 215
stars: 2
tags:
  - text-to-image
  - lora
  - diffusers
  - "template:diffusion-lora"
---

# Z-Image-Fun-Distill-ComfyUI

> Z-Image-Fun-Distill-ComfyUI - AI-ModelScope 在 ModelScope 开源的模型。Z-Image-Fun-Lora-Distill-8-Steps ComfyUI

AI-ModelScope/Z-Image-Fun-Distill-ComfyUI 是 ModelScope 魔搭社区上的 260.0M 参数机器学习模型，采用 apache-2.0 许可，基于 Tongyi-MAI/Z-Image 构建，并支持在线推理（img2img）。

- **Repository**: AI-ModelScope/Z-Image-Fun-Distill-ComfyUI
- **License**: apache-2.0
- **Parameters**: 260.0M
- **Base model**: Tongyi-MAI/Z-Image
- **Online inference**: img2img
- **Tags**: text-to-image, lora, diffusers, template:diffusion-lora
- **Downloads**: 215
- **Stars**: 2
- **Last updated**: 2026-02-11

Source: https://www.modelscope.cn/models/AI-ModelScope/Z-Image-Fun-Distill-ComfyUI

---

# Z-Image-Fun-Lora-Distill-8-Steps ComfyUI

<Gallery />

## Model description 

Credit to alibaba-pai for more fantastic work!

**New update 2/10/26**
The new Fun-distill-2602 version by alibaba-pai is correctly converted for ComfyUI. Unfortunately there are new feed forward keys in this version.
These keys affect inference, while they help to improve contrast and brightness, they also prevent any step inference from converging when sigmas are <0.500.
This means that many schedulers that have steps below this, like exponential, bong_tangent, will not work.

I have vigorously tested this, and initially removed all feed forward keys but the result, although otherwise clean, would appear somewhat washed out.
I have selectively removed the feed forward keys from the weights, keeping only main blocks 0, 20-25 intact.
All refiner blocks have also been kept intact.

This allows for all schedulers and samplers to work properly, while also maintaining the benefits of the new version.

Refer to the "Tests" folder for comparisons.

---

This is the lora converted for ComfyUI compatibility. Meant to be used with the Z-Image base model and its derivatives.

Best run at 8 steps with CFG 1. Avoid sageattention. Samplers that add noise per step such as LCM seem to work well.

[Original Lora by alibaba-pai](/alibaba-pai/Z-Image-Fun-Lora-Distill)


## Download model


[Download](/UDCAI/Z-Image-Fun-Distill-ComfyUI/tree/main) them in the Files & versions tab.