--- license: apache-2.0 prior: - kandinsky-community/kandinsky-2-2-prior tags: - text-to-image - kandinsky --- # Kandinsky 2.2 Kandinsky inherits best practices from Dall-E 2 and Latent diffusion while introducing some new ideas. It uses the CLIP model as a text and image encoder, 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. The Kandinsky model is created by [Arseniy Shakhmatov](https://github.com/cene555), [Anton Razzhigaev](https://github.com/razzant), [Aleksandr Nikolich](https://github.com/AlexWortega), [Igor Pavlov](https://github.com/boomb0om), [Andrey Kuznetsov](https://github.com/kuznetsoffandrey) and [Denis Dimitrov](https://github.com/denndimitrov) ## Usage Kandinsky 2.2 is available in diffusers! ```python pip install diffusers transformers accelerate ``` ### Text-to-Image Generation with ControlNet Conditioning ```python import torch import numpy as np from transformers import pipeline from diffusers.utils import load_image from diffusers import KandinskyV22PriorPipeline, KandinskyV22ControlnetPipeline # let's take an image and extract its depth map. def make_hint(image, depth_estimator): image = depth_estimator(image)["depth"] image = np.array(image) image = image[:, :, None] image = np.concatenate([image, image, image], axis=2) detected_map = torch.from_numpy(image).float() / 255.0 hint = detected_map.permute(2, 0, 1) return hint img = load_image( "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/kandinskyv22/cat.png" ).resize((768, 768)) # We can use the `depth-estimation` pipeline from transformers to process the image and retrieve its depth map. depth_estimator = pipeline("depth-estimation") hint = make_hint(img, depth_estimator).unsqueeze(0).half().to("cuda") # Now, we load the prior pipeline and the text-to-image controlnet pipeline pipe_prior = KandinskyV22PriorPipeline.from_pretrained( "kandinsky-community/kandinsky-2-2-prior", torch_dtype=torch.float16 ) pipe_prior = pipe_prior.to("cuda") pipe = KandinskyV22ControlnetPipeline.from_pretrained( "kandinsky-community/kandinsky-2-2-controlnet-depth", torch_dtype=torch.float16 ) pipe = pipe.to("cuda") # We pass the prompt and negative prompt through the prior to generate image embeddings prompt = "A robot, 4k photo" negative_prior_prompt = "lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, out of frame, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, username, watermark, signature" generator = torch.Generator(device="cuda").manual_seed(43) image_emb, zero_image_emb = pipe_prior( prompt=prompt, negative_prompt=negative_prior_prompt, generator=generator ).to_tuple() # Now we can pass the image embeddings and the depth image we extracted to the controlnet pipeline. With Kandinsky 2.2, only prior pipelines accept `prompt` input. You do not need to pass the prompt to the controlnet pipeline. images = pipe( image_embeds=image_emb, negative_image_embeds=zero_image_emb, hint=hint, num_inference_steps=50, generator=generator, height=768, width=768, ).images images[0].save("robot_cat.png") ``` ![img](https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/kandinskyv22/cat.png) ![img](https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/kandinskyv22/robot_cat_text2img.png) ### Image-to-Image Generation with ControlNet Conditioning ```python import torch import numpy as np from diffusers import KandinskyV22PriorEmb2EmbPipeline, KandinskyV22ControlnetImg2ImgPipeline from diffusers.utils import load_image from transformers import pipeline img = load_image( "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" "/kandinskyv22/cat.png" ).resize((768, 768)) def make_hint(image, depth_estimator): image = depth_estimator(image)["depth"] image = np.array(image) image = image[:, :, None] image = np.concatenate([image, image, image], axis=2) detected_map = torch.from_numpy(image).float() / 255.0 hint = detected_map.permute(2, 0, 1) return hint depth_estimator = pipeline("depth-estimation") hint = make_hint(img, depth_estimator).unsqueeze(0).half().to("cuda") pipe_prior = KandinskyV22PriorEmb2EmbPipeline.from_pretrained( "kandinsky-community/kandinsky-2-2-prior", torch_dtype=torch.float16 ) pipe_prior = pipe_prior.to("cuda") pipe = KandinskyV22ControlnetImg2ImgPipeline.from_pretrained( "kandinsky-community/kandinsky-2-2-controlnet-depth", torch_dtype=torch.float16 ) pipe = pipe.to("cuda") prompt = "A robot, 4k photo" negative_prior_prompt = "lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, out of frame, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, username, watermark, signature" generator = torch.Generator(device="cuda").manual_seed(43) # run prior pipeline img_emb = pipe_prior(prompt=prompt, image=img, strength=0.85, generator=generator) negative_emb = pipe_prior(prompt=negative_prior_prompt, image=img, strength=1, generator=generator) # run controlnet img2img pipeline images = pipe( image=img, strength=0.5, image_embeds=img_emb.image_embeds, negative_image_embeds=negative_emb.image_embeds, hint=hint, num_inference_steps=50, generator=generator, height=768, width=768, ).images images[0].save("robot_cat.png") ``` Here is the output. Compared with the output from our text-to-image controlnet example, it kept a lot more cat facial details from the original image and worked into the robot style we asked for. ![img](https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/kandinskyv22/robot_cat.png) ## Model Architecture ### Overview Kandinsky 2.2 is a text-conditional diffusion model based on unCLIP and latent diffusion, composed of a transformer-based image prior model, a unet diffusion model, and a decoder. The model architectures are illustrated in the figure below - the chart on the left describes the process to train the image prior model, the figure in the center is the text-to-image generation process, and the figure on the right is image interpolation.

Specifically, the image prior model was trained on CLIP text and image embeddings generated with a pre-trained [CLIP-ViT-G model](https://huggingface.co/laion/CLIP-ViT-g-14-laion2B-s12B-b42K). The trained image prior model is then used to generate CLIP image embeddings for input text prompts. Both the input text prompts and its CLIP image embeddings are used in the diffusion process. A [MoVQGAN](https://openreview.net/forum?id=Qb-AoSw4Jnm) model acts as the final block of the model, which decodes the latent representation into an actual image. ### Details The image prior training of the model was performed on the [LAION Improved Aesthetics dataset](https://huggingface.co/datasets/bhargavsdesai/laion_improved_aesthetics_6.5plus_with_images), and then fine-tuning was performed on the [LAION HighRes data](https://huggingface.co/datasets/laion/laion-high-resolution). The main Text2Image diffusion model was trained on [LAION HighRes dataset](https://huggingface.co/datasets/laion/laion-high-resolution) and then fine-tuned with a dataset of 2M very high-quality high-resolution images with descriptions (COYO, anime, landmarks_russia, and a number of others) was used separately collected from open sources. The main change in Kandinsky 2.2 is the replacement of CLIP-ViT-G. Its image encoder significantly increases the model's capability to generate more aesthetic pictures and better understand text, thus enhancing its overall performance. Due to the switch CLIP model, the image prior model was retrained, and the Text2Image diffusion model was fine-tuned for 2000 iterations. Kandinsky 2.2 was trained on data of various resolutions, from 512 x 512 to 1536 x 1536, and also as different aspect ratios. As a result, Kandinsky 2.2 can generate 1024 x 1024 outputs with any aspect ratio. ### Evaluation We quantitatively measure the performance of Kandinsky 2.1 on the COCO_30k dataset, in zero-shot mode. The table below presents FID. FID metric values ​​for generative models on COCO_30k | | FID (30k)| |:------|----:| | eDiff-I (2022) | 6.95 | | Image (2022) | 7.27 | | Kandinsky 2.1 (2023) | 8.21| | Stable Diffusion 2.1 (2022) | 8.59 | | GigaGAN, 512x512 (2023) | 9.09 | | DALL-E 2 (2022) | 10.39 | | GLIDE (2022) | 12.24 | | Kandinsky 1.0 (2022) | 15.40 | | DALL-E (2021) | 17.89 | | Kandinsky 2.0 (2022) | 20.00 | | GLIGEN (2022) | 21.04 | For more information, please refer to the upcoming technical report. ## BibTex If you find this repository useful in your research, please cite: ``` @misc{kandinsky 2.2, title = {kandinsky 2.2}, author = {Arseniy Shakhmatov, Anton Razzhigaev, Aleksandr Nikolich, Vladimir Arkhipkin, Igor Pavlov, Andrey Kuznetsov, Denis Dimitrov}, year = {2023}, howpublished = {}, } ```