Kandinsky2.1 / app.py
ai-forever's picture
Update app.py
import os
import gradio as gr
import torch
from torch import autocast
from kandinsky2 import get_kandinsky2
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
from kandinsky2 import get_kandinsky2
model = get_kandinsky2('cuda', task_type='text2img', model_version='2.1', use_flash_attention=False)
sample r='ddim_sampler',
def infer(prompt, negative='low quality, bad quality'):
images = model.generate_text2img(prompt,
h=768, w=768,
return images
css = """
.gradio-container {
font-family: 'IBM Plex Sans', sans-serif;
.gr-button {
color: white;
border-color: black;
background: black;
input[type='range'] {
accent-color: black;
.dark input[type='range'] {
accent-color: #dfdfdf;
.container {
max-width: 730px;
margin: auto;
padding-top: 1.5rem;
#gallery {
min-height: 22rem;
margin-bottom: 15px;
margin-left: auto;
margin-right: auto;
border-bottom-right-radius: .5rem !important;
border-bottom-left-radius: .5rem !important;
#gallery>div>.h-full {
min-height: 20rem;
.details:hover {
text-decoration: underline;
.gr-button {
white-space: nowrap;
.gr-button:focus {
border-color: rgb(147 197 253 / var(--tw-border-opacity));
outline: none;
box-shadow: var(--tw-ring-offset-shadow), var(--tw-ring-shadow), var(--tw-shadow, 0 0 #0000);
--tw-border-opacity: 1;
--tw-ring-offset-shadow: var(--tw-ring-inset) 0 0 0 var(--tw-ring-offset-width) var(--tw-ring-offset-color);
--tw-ring-shadow: var(--tw-ring-inset) 0 0 0 calc(3px var(--tw-ring-offset-width)) var(--tw-ring-color);
--tw-ring-color: rgb(191 219 254 / var(--tw-ring-opacity));
--tw-ring-opacity: .5;
#advanced-btn {
font-size: .7rem !important;
line-height: 19px;
margin-top: 12px;
margin-bottom: 12px;
padding: 2px 8px;
border-radius: 14px !important;
#advanced-options {
display: none;
margin-bottom: 20px;
.footer {
margin-bottom: 45px;
margin-top: 35px;
text-align: center;
border-bottom: 1px solid #e5e5e5;
.footer>p {
font-size: .8rem;
display: inline-block;
padding: 0 10px;
transform: translateY(10px);
background: white;
.dark .footer {
border-color: #303030;
.dark .footer>p {
background: #0b0f19;
.acknowledgments h4{
margin: 1.25em 0 .25em 0;
font-weight: bold;
font-size: 115%;
display: flex;
flex-wrap: wrap;
justify-content: space-between;
align-items: center;
.animate-spin {
animation: spin 1s linear infinite;
@keyframes spin {
from {
transform: rotate(0deg);
to {
transform: rotate(360deg);
#share-btn-container {
display: flex; padding-left: 0.5rem !important; padding-right: 0.5rem !important; background-color: #000000; justify-content: center; align-items: center; border-radius: 9999px !important; width: 13rem;
#share-btn {
all: initial; color: #ffffff;font-weight: 600; cursor:pointer; font-family: 'IBM Plex Sans', sans-serif; margin-left: 0.5rem !important; padding-top: 0.25rem !important; padding-bottom: 0.25rem !important;
#share-btn * {
all: unset;
flex: 1 1 50%; border-top-right-radius: 0; border-bottom-right-radius: 0;
gap: 0;
min-height: 700px
block = gr.Blocks(css=css)
examples = [
'Thinking man in anime style'
SPACE_ID = os.getenv('SPACE_ID')
with block as demo:
[![Framework: PyTorch](https://img.shields.io/badge/Framework-PyTorch-orange.svg)](https://pytorch.org/) [![Huggingface space](https://img.shields.io/badge/πŸ€—-Huggingface-yello.svg)](https://huggingface.co/sberbank-ai/Kandinsky_2.0)
<p>For faster inference without waiting in queue, you may duplicate the space and upgrade to GPU in settings. <a href="https://huggingface.co/spaces/{SPACE_ID}?duplicate=true"><img style="display: inline; margin-top: 0em; margin-bottom: 0em" src="https://bit.ly/3gLdBN6" alt="Duplicate Space" /></a></p>
[Offical BlogPost](https://habr.com/ru/company/sberbank/blog/725282/)
[Offical Telegram Bot](https://t.me/kandinsky21_bot)
[Offical site](https://fusionbrain.ai/diffusion)
## 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 num_layers=20, num_heads=32 and hidden_size=2048.
Other architecture parts:
- Text encoder (XLM-Roberta-Large-Vit-L-14) - 560M
- Diffusion Image Prior β€” 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.
**Kandinsky 2.1** architecture overview:
with gr.Group():
with gr.Box():
with gr.Row().style(mobile_collapse=False, equal_height=True):
text = gr.Textbox(
label="Enter your prompt", show_label=True, max_lines=2
border=(True, False, True, True),
rounded=(True, False, False, True),
negative = gr.Textbox(
label="Enter your negative prompt", show_label=True, max_lines=2
border=(True, False, True, True),
rounded=(True, False, False, True),
btn = gr.Button("Run").style(
rounded=(False, True, True, False),
gallery = gr.Gallery(label="Generated images", show_label=False, elem_id="generated_id").style(
grid=[2], height="auto"
ex = gr.Examples(examples=examples, fn=infer, inputs=[text, negative], outputs=gallery, cache_examples=True)
ex.dataset.headers = [""]
text.submit(infer, inputs=[text, negative], outputs=gallery)
btn.click(infer, inputs=[text, negative], outputs=gallery)
# Authors
+ Arseniy Shakhmatov: [Github](https://github.com/cene555), [Blog](https://t.me/gradientdip)
+ Anton Razzhigaev: [Github](https://github.com/razzant), [Blog](https://t.me/abstractDL)
+ Aleksandr Nikolich: [Github](https://github.com/AlexWortega), [Blog](https://t.me/lovedeathtransformers)
+ Vladimir Arkhipkin: [Github](https://github.com/oriBetelgeuse)
+ Igor Pavlov: [Github](https://github.com/boomb0om)
+ Andrey Kuznetsov: [Github](https://github.com/kuznetsoffandrey)
+ Denis Dimitrov: [Github](https://github.com/denndimitrov)