---
license: apache-2.0
base_model: Qwen/Qwen-Image
library_name: diffusers
pipeline_tag: text-to-image
tags:
  - lora
  - qwen-image
  - text-to-image
  - diffusers
  - safetensors
inference: false
---

# Azis1 LoRA for Qwen-Image

A LoRA (Low-Rank Adaptation) model trained on the Qwen-Image base model for generating images of a specific person.

## Model Details

| Property | Value |
|----------|-------|
| **Base Model** | [Qwen/Qwen-Image](https://huggingface.co/Qwen/Qwen-Image) |
| **LoRA Rank** | 64 |
| **LoRA Alpha** | 64 |
| **Training Resolution** | 512x512 |
| **Training Epochs** | 5 |
| **Training Images** | ~2400 |
| **Trigger Word** | `azis1` |

## Usage

### Trigger Word

Always include `azis1` at the beginning of your prompt for best results.

### With Diffusers (Python)

```python
import torch
from diffusers import DiffusionPipeline

# Load base model
pipe = DiffusionPipeline.from_pretrained(
    "Qwen/Qwen-Image",
    torch_dtype=torch.bfloat16,
)
pipe.enable_model_cpu_offload()

# Load LoRA
pipe.load_lora_weights("YOUR_USERNAME/azis1-qwen-lora")

# Generate
image = pipe(
    prompt="azis1, portrait of a man, professional photography, studio lighting",
    num_inference_steps=50,
    guidance_scale=4.0,
    height=1024,
    width=1024,
).images[0]

image.save("output.png")
```

### With ComfyUI

1. Download the `.safetensors` file to `ComfyUI/models/loras/`
2. Use this node setup:
   - **UNETLoader** → `qwen_image_fp8_e4m3fn.safetensors`
   - **LoraLoaderModelOnly** → `azis1.safetensors` (strength: 0.85)
   - **CLIPLoader** → `qwen_2.5_vl_7b_fp8_scaled.safetensors`
   - **VAELoader** → `qwen_image_vae.safetensors`

3. Recommended settings:
   - Steps: 50
   - CFG Scale: 4.0
   - Sampler: Euler

## Example Prompts

```
azis1, portrait of a man, professional photography, studio lighting, high quality, 4K
azis1, man in casual clothes, outdoor, natural lighting
azis1, close-up portrait, dramatic lighting, artistic
azis1, man with beard, formal attire, corporate headshot
```

## Training Details

- **Framework**: OneTrainer
- **Optimizer**: AdamW 8-bit
- **Learning Rate**: 0.0001 (Cosine schedule)
- **Batch Size**: 2 (with gradient accumulation of 2)
- **Precision**: bfloat16
- **Hardware**: NVIDIA RTX 5090 (32GB VRAM)

## Limitations

- This LoRA only works with **Qwen-Image** base model
- It will NOT work with Stable Diffusion 1.5, SDXL, or Flux models
- Best results at 512x512 to 1024x1024 resolution

## License

Apache 2.0 - Same as the base Qwen-Image model.

## Acknowledgments

- [Qwen-Image](https://huggingface.co/Qwen/Qwen-Image) by Alibaba
- [OneTrainer](https://github.com/Nerogar/OneTrainer) for training