---
license: apache-2.0
tags:
- text-to-image
- kandinsky
inference: false
---
# Kandinsky 3.0 IP Adapter

## Usage
```python
pip install git+https://github.com/ai-forever/kandinsky3-diffusers.git
```
### Image variations

```python
from diffusers.models.attention_processor import Kandi3AttnProcessorIpAdapter, Kandi3AttnProcessor
from diffusers.pipelines.kandinsky3.kandinsky3_pipeline_ip_adapter import KandinskyV3PipelineIpAdapter
from PIL import Image
import torch
pipe = KandinskyV3PipelineIpAdapter.from_pretrained('ai-forever/kandinsky3_ip_adapter', torch_dtype=torch.float16, low_cpu_mem_usage=False, device_map=None)
pipe = pipe.to('cuda')
img = Image.open('path_to_img.jpg')
out_img = pipe('4k caption', img=[img], weights=[1], negative_prompt='', height=1024, width=1024, guidance_scale=7.5, num_inference_steps=50, cut_context=1)[0][0]
```

### Image + Image mixing

```python
from diffusers.models.attention_processor import Kandi3AttnProcessorIpAdapter, Kandi3AttnProcessor
from diffusers.pipelines.kandinsky3.kandinsky3_pipeline_ip_adapter import KandinskyV3PipelineIpAdapter
from PIL import Image
import torch
pipe = KandinskyV3PipelineIpAdapter.from_pretrained('ai-forever/kandinsky3_ip_adapter', torch_dtype=torch.float16, low_cpu_mem_usage=False, device_map=None)
pipe = pipe.to('cuda')
img1 = Image.open('path_to_img1.jpg')
img2 = Image.open('path_to_img2.jpg')

out_img = pipe('4k photo', img=[img1, img2], weights=[0.5, 0.5], negative_prompt='', height=1024, width=1024, guidance_scale=7.5, num_inference_steps=50, cut_context=1)[0][0]
```

### Text + Image mixing

```python
from diffusers.models.attention_processor import Kandi3AttnProcessorIpAdapter, Kandi3AttnProcessor
from diffusers.pipelines.kandinsky3.kandinsky3_pipeline_ip_adapter import KandinskyV3PipelineIpAdapter
from PIL import Image
import torch
pipe = KandinskyV3PipelineIpAdapter.from_pretrained('ai-forever/kandinsky3_ip_adapter', torch_dtype=torch.float16, low_cpu_mem_usage=False, device_map=None)
pipe = pipe.to('cuda')
img = Image.open('path_to_img.jpg')
caption = 'cat, 4k photo'
out_img = pipe(caption, img=[img], weights=[1], negative_prompt='', height=1024, width=1024, guidance_scale=7.5, num_inference_steps=50, cut_context=1, img_weight=0.5)[0][0]
```