一个用于系统性测试多个提示和强度值的 LoRA 模型的工具包，使用本地 ComfyUI 服务器。包含批处理生成脚本和基于浏览器的图库，用于比较结果。

它通过内置的 comfy_local 技能构建，使用 Claude Code，但 Python 脚本和画廊也可以独立运行。

盒子里有什么

先修条件

• ComfyUI，本地运行于8188端口
• Z-Image Turbo 模型文件安装：
• z_image_turbo_bf16.safetensors（扩散模型）
• qwen_3_4b.safetensors（CLIP文本编码器）
• ae.safetensors（VAE）
• Python 3.8+（无 pip 依赖——仅使用 stdlib）

快速启动

1.配置你的测试

编辑lora_test.py顶部以设置提示和默认LoRA：

2.做测试

3.浏览结果

LoRA Tester

A toolkit for systematically testing LoRA models across multiple prompts and strength values using a local ComfyUI server. Includes a batch generation script and a browser-based gallery for comparing results.

Built to work with Claude Code via the included comfy_local skill, but the Python script and gallery work standalone too.

What’s in the box

Prerequisites

• ComfyUI running locally on port 8188
• Z-Image Turbo model files installed:
  • z_image_turbo_bf16.safetensors (diffusion model)
  • qwen_3_4b.safetensors (CLIP text encoder)
  • ae.safetensors (VAE)
• Python 3.8+ (no pip dependencies – uses only stdlib)

Quick Start

1. Configure your test

Edit the top of lora_test.py to set your prompts and default LoRA:

LORA = "z_image_turbo\\my-lora.safetensors"

PROMPTS = [
    "a wizard casting a spell",
    "a sunset over mountains",
    "a cat sitting on a windowsill",
]

STRENGTHS = [0.0, 0.25, 0.5, 0.75, 1.0]

2. Run the test

# Use defaults from the script config:
python lora_test.py

# Or override via CLI:
python lora_test.py --lora "z_image_turbo\\my-lora.safetensors" --strengths "0,0.5,1.0" --name "my lora test"

# List available LoRAs on your ComfyUI server:
python lora_test.py --list-loras

3. Browse results

Open gallery.html in your browser and click Open Project to select the generated project folder from projects/.

No server needed – it runs entirely in the browser using local file access.

CLI Options

python lora_test.py [options]

--lora FILENAME     LoRA file path (e.g. "z_image_turbo\\my-lora.safetensors")
--strengths LIST    Comma-separated strength values (e.g. "0,0.25,0.5,0.75,1.0")
--name NAME         Project name (defaults to LoRA filename slug)
--notes TEXT        Optional notes saved to the manifest
--list-loras        List all available LoRAs on the ComfyUI server and exit

How it works

The script submits one ComfyUI job per (prompt, strength) combination using the Z-Image Turbo pipeline. All jobs are queued at once and processed sequentially by ComfyUI (~2 seconds per image on an RTX 5090).

• Strength 0.00 generates a baseline image with no LoRA applied (the LoraLoader node is skipped entirely)
• The same seed is used per prompt across all strengths, so you get an apples-to-apples comparison of just the LoRA’s effect

Each run creates a self-contained project folder:

projects/