Character LoRA (Qwen)
Video Lecture
| Section | Video Links |
|---|---|
| Character LoRA (Qwen) | ![]() |
Description
We will create our own character LoRA using Qwen models.
DataSet
When creating a LoRA we will need a selection of good quality images of our character or object to train from.
You can use photos of yourself, somebody else who you have permission from, images of your pets, other objects around you, or even AI generated images, provided that the license of the model checkpoint that you used allows reusing your images for training.
| Initial Image | Workflow |
|---|---|
![]() |
![]() |
I am using the same models from the Qwen Edit 2509 lesson. Although I am using a newer version of Qwen-Image-Lightning-4steps-V2.0-bf16.safetensors
📂 ComfyUI/
├── 📂 models/
│ ├── 📂 loras/
│ │ └── Qwen-Image-Lightning-4steps-V2.0-bf16.safetensors
Generating the LoRA
Register at https://modelscope.ai/civision/home and use invitation code qwenday.
We will generate a LoRA using the Qwen Image 2512 model.
Using the LoRA to Create Images
When using the LoRA, I am also using the Qwen Image 2512 models.
- qwen_image_2512_fp8_e4m3fn.safetensors (20.4 GB)
- Qwen-Image-Lightning-4steps-V2.0-bf16.safetensors (850 mb)
- qwen_2.5_vl_7b_fp8_scaled.safetensors (9.38 gb)
- qwen_image_vae.safetensors (254 mb)
📂 ComfyUI/
├── 📂 models/
│ ├── 📂 diffusion_models/
│ │ └── qwen_image_2512_fp8_e4m3fn.safetensors
│ ├── 📂 loras/
│ │ └── Qwen-Image-Lightning-4steps-V2.0-bf16.safetensors
│ ├── 📂 text_encoders/
│ │ └── qwen_2.5_vl_7b_fp8_scaled.safetensors
│ └── 📂 vae/
│ └── qwen_image_vae.safetensors
Final Workflow
-
SBCODE_Pria (ckpt-20) (236 MB)
-
SBCODE_CosmoCat (ckpt-20) (236 MB)
📂 ComfyUI/
├── 📂 models/
│ ├── 📂 loras/
│ │ ├── SBCODE_Pria_20.safetensors
│ │ └── SBCODE_CosmoCat_20.safetensors
Drag this workflow into ComfyUI.




































