Kandinsky 2.2

— Inference example

— Fine-tuning with LoRA

• Habr post
• Demo fusionbrain.ai
• Telegram-bot

Description:

Kandinsky 2.2 brings substantial improvements upon its predecessor, Kandinsky 2.1, by introducing a new, more powerful image encoder - CLIP-ViT-G and the ControlNet support.

The switch to CLIP-ViT-G as the image encoder significantly increases the model's capability to generate more aesthetic pictures and better understand text, thus enhancing the model's overall performance.

The addition of the ControlNet mechanism allows the model to effectively control the process of generating images. This leads to more accurate and visually appealing outputs and opens new possibilities for text-guided image manipulation.

Architecture details:

• Text encoder (XLM-Roberta-Large-Vit-L-14) - 560M
• Diffusion Image Prior — 1B
• CLIP image encoder (ViT-bigG-14-laion2B-39B-b160k) - 1.8B
• Latent Diffusion U-Net - 1.22B
• MoVQ encoder/decoder - 67M

Сheckpoints:

• Prior: A prior diffusion model mapping text embeddings to image embeddings
• Text-to-Image / Image-to-Image: A decoding diffusion model mapping image embeddings to images
• Inpainting: A decoding diffusion model mapping image embeddings and masked images to images
• ControlNet-depth: A decoding diffusion model mapping image embedding and additional depth condition to images

Inference regimes

How to use:

Check our jupyter notebooks with examples in ./notebooks folder

1. text2image

python from kandinsky2 import get_kandinsky2 model = get_kandinsky2('cuda', task_type='text2img', model_version='2.2') images = model.generate_text2img( "red cat, 4k photo", decoder_steps=50, batch_size=1, h=1024, w=768, )

Kandinsky 2.1

Habr post

Demo

pip install "git+https://github.com/ai-forever/Kandinsky-2.git"

Model architecture:

Kandinsky 2.1 inherits best practicies from Dall-E 2 and Latent diffusion, while introducing some new ideas.

As text and image encoder it uses CLIP model and diffusion image prior (mapping) between latent spaces of CLIP modalities. This approach increases the visual performance of the model and unveils new horizons in blending images and text-guided image manipulation.

For diffusion mapping of latent spaces we use transformer with numlayers=20, numheads=32 and hidden_size=2048.

Other architecture parts:

• Text encoder (XLM-Roberta-Large-Vit-L-14) - 560M
• Diffusion Image Prior — 1B
• CLIP image encoder (ViT-L/14) - 427M
• Latent Diffusion U-Net - 1.22B
• MoVQ encoder/decoder - 67M

Kandinsky 2.1 was trained on a large-scale image-text dataset LAION HighRes and fine-tuned on our internal datasets.

How to use:

Check our jupyter notebooks with examples in ./notebooks folder

1. text2image

python from kandinsky2 import get_kandinsky2 model = get_kandinsky2('cuda', task_type='text2img', model_version='2.1', use_flash_attention=False) images = model.generate_text2img( "red cat, 4k photo", num_steps=100, batch_size=1, guidance_scale=4, h=768, w=768, sampler='p_sampler', prior_cf_scale=4, prior_steps="5" )

prompt: "Einstein in space around the logarithm scheme"

2. image fuse

python from kandinsky2 import get_kandinsky2 from PIL import Image model = get_kandinsky2('cuda', task_type='text2img', model_version='2.1', use_flash_attention=False) images_texts = ['red cat', Image.open('img1.jpg'), Image.open('img2.jpg'), 'a man'] weights = [0.25, 0.25, 0.25, 0.25] images = model.mix_images( images_texts, weights, num_steps=150, batch_size=1, guidance_scale=5, h=768, w=768, sampler='p_sampler', prior_cf_scale=4, prior_steps="5" )

3. inpainting

```python from kandinsky2 import get_kandinsky2 from PIL import Image import numpy as np

model = getkandinsky2('cuda', tasktype='inpainting', modelversion='2.1', useflashattention=False) initimage = Image.open('img.jpg') mask = np.ones((768, 768), dtype=np.float32) mask[:,:550] = 0 images = model.generateinpainting( 'man 4k photo', initimage, mask, numsteps=150, batchsize=1, guidancescale=5, h=768, w=768, sampler='psampler', priorcfscale=4, prior_steps="5" ) ```

Kandinsky 2.0

Habr post

Demo

pip install "git+https://github.com/ai-forever/Kandinsky-2.git"

Model architecture:

It is a latent diffusion model with two multilingual text encoders: * mCLIP-XLMR 560M parameters * mT5-encoder-small 146M parameters

These encoders and multilingual training datasets unveil the real multilingual text-to-image generation experience!

Kandinsky 2.0 was trained on a large 1B multilingual set, including samples that we used to train Kandinsky.

In terms of diffusion architecture Kandinsky 2.0 implements UNet with 1.2B parameters.

Kandinsky 2.0 architecture overview:

How to use:

Check our jupyter notebooks with examples in ./notebooks folder

1. text2img

```python from kandinsky2 import get_kandinsky2

model = getkandinsky2('cuda', tasktype='text2img') images = model.generatetext2img('A teddy bear на красной площади', batchsize=4, h=512, w=512, numsteps=75, denoisedtype='dynamicthreshold', dynamicthresholdv=99.5, sampler='ddimsampler', ddimeta=0.05, guidancescale=10) ```

prompt: "A teddy bear на красной площади"

2. inpainting

```python from kandinsky2 import get_kandinsky2 from PIL import Image import numpy as np

model = getkandinsky2('cuda', tasktype='inpainting') initimage = Image.open('image.jpg') mask = np.ones((512, 512), dtype=np.float32) mask[100:] = 0 images = model.generateinpainting('Девушка в красном платье', initimage, mask, numsteps=50, denoisedtype='dynamicthreshold', dynamicthresholdv=99.5, sampler='ddimsampler', ddimeta=0.05, guidance_scale=10) ```

prompt: "Девушка в красном платье"

3. img2img

```python from kandinsky2 import get_kandinsky2 from PIL import Image

model = getkandinsky2('cuda', tasktype='img2img') initimage = Image.open('image.jpg') images = model.generateimg2img('кошка', initimage, strength=0.8, numsteps=50, denoisedtype='dynamicthreshold', dynamicthresholdv=99.5, sampler='ddimsampler', ddimeta=0.05, guidance_scale=10) ```

Authors

• Arseniy Shakhmatov: Github, Blog
• Anton Razzhigaev: Github, Blog
• Aleksandr Nikolich: Github, Blog
• Vladimir Arkhipkin: Github
• Igor Pavlov: Github
• Andrey Kuznetsov: Github
• Denis Dimitrov: Github