README.md

# Step1X-Anime-Edit-Lora

This repository provides LoRA (Low-Rank Adaptation) support for the Step1X-Edit anime image editing model. It allows for fine-tuned control over image editing outputs.

## Installation

Refer to the main Step1X-Edit installation instructions at:  
https://github.com/stepfun-ai/Step1X-Edit

```bash
wget https://huggingface.co/stepfun-ai/Step1X-Edit/resolve/main/step1x-edit-i1258.safetensors
wget https://huggingface.co/stepfun-ai/Step1X-Edit/resolve/main/vae.safetensors
huggingface-cli download Qwen/Qwen2.5-VL-7B-Instruct --local-dir Qwen2.5-VL-7B-Instruct
```

## Usage Examples

### Basic Setup
```python
from inference import *

image_edit = ImageGenerator(
    ae_path="vae.safetensors",
    dit_path="step1x-edit-i1258.safetensors",
    qwen2vl_model_path='Qwen2.5-VL-7B-Instruct',
    max_length=640,
    quantized=True,
    offload=True,
    lora="change_output/step1x-edit_change-step00003000.safetensors",
    mode="flash"
)
```

### Example 1: Changing Background and Adding Elements

![image/png](https://cdn-uploads.huggingface.co/production/uploads/634dffc49b777beec3bc6448/Fhj-nsp6i7zLj_KIPLSmn.png)

```python
image_path = "万叶.png"
prompt = '''
将背景改成公园，添加一些小松鼠
'''

num_steps = 28
cfg_guidance = 4.5
seed = 42
size_level = 512  # Can also be 768 or 1024

image = image_edit.generate_image(
    prompt,
    negative_prompt="",
    ref_images=Image.open(image_path).convert("RGB"),
    num_samples=1,
    num_steps=num_steps,
    cfg_guidance=cfg_guidance,
    seed=seed,
    show_progress=True,
    size_level=size_level,
)[0]
image.save("万叶在公园.png")
```

- original output

![image/png](https://cdn-uploads.huggingface.co/production/uploads/634dffc49b777beec3bc6448/oniiy9n070-a7dlqcjxRk.png)

- lora output


![image/png](https://cdn-uploads.huggingface.co/production/uploads/634dffc49b777beec3bc6448/8xLtoa9aN2dLwjmBFGglj.png)

### Example 2: Advanced Scene Modification


![image/png](https://cdn-uploads.huggingface.co/production/uploads/634dffc49b777beec3bc6448/UqFgDkgrvivl3wzTAUPwA.png)

```python
image_path = "万叶.png"
prompt = '''
将背景改成公园，添加一些小松鼠，天气为黄昏，调整为橙色光照，让男孩微笑
'''

# Same parameters as above
image = image_edit.generate_image(...)
image.save("万叶在黄昏.png")
```

- original output


![image/png](https://cdn-uploads.huggingface.co/production/uploads/634dffc49b777beec3bc6448/L3L1Gow9_BUrZQaC-cOKa.png)

- lora output

![image/png](https://cdn-uploads.huggingface.co/production/uploads/634dffc49b777beec3bc6448/duxI2L919IGIIFmxuhGMD.png)


### Example 3: Character Modification



![image/jpeg](https://cdn-uploads.huggingface.co/production/uploads/634dffc49b777beec3bc6448/kJUn3IboVQqN8xVQRSQrc.jpeg)

```python
image_path = "塔利亚.jpg"
prompt = '''
将图片背景变成海边，手里拿着一个冰淇凌
'''

num_steps = 28
cfg_guidance = 6  # Higher guidance for more complex changes
seed = 42
size_level = 512

image = image_edit.generate_image(...)
image.save("塔利亚在海边.jpg")
```

- original output


![image/jpeg](https://cdn-uploads.huggingface.co/production/uploads/634dffc49b777beec3bc6448/qld6MfI8ZUOKgoNeMihtI.jpeg)


- lora output


![image/png](https://cdn-uploads.huggingface.co/production/uploads/634dffc49b777beec3bc6448/3g0Pjrku6jqdBfMGG1p1E.png)


### Example 4: Object Replacement and Style Change


![image/jpeg](https://cdn-uploads.huggingface.co/production/uploads/634dffc49b777beec3bc6448/TZDq4f6wRHf9Y075zQGmw.jpeg)

```python
image_path = "星铁海报.jpg"
prompt = '''
将桌子上的鞋替换成一个汉堡，背景换成星光咖啡厅，帽子换成小熊帽
'''

num_steps = 28
cfg_guidance = 4.5
seed = 42
size_level = 512

image = image_edit.generate_image(...)
image.save("星铁小猫在咖啡厅.png")
```

- original output


![image/png](https://cdn-uploads.huggingface.co/production/uploads/634dffc49b777beec3bc6448/3RLwHrYgmibIxbYMRkqVW.png)

- lora output


![image/png](https://cdn-uploads.huggingface.co/production/uploads/634dffc49b777beec3bc6448/oANb6mpmM8LMmfTzCTnw2.png)


## Parameters
- `num_steps`: Number of diffusion steps (typically 28)
- `cfg_guidance`: Guidance scale (4.5-6 recommended)
- `seed`: Random seed for reproducibility
- `size_level`: Output resolution (512)

## Output Comparison
Each example shows the original output vs. LoRA-enhanced output for comparison.