Karlo v1 alpha

Karlo is a text-conditional image generation model based on OpenAI's unCLIP architecture with the improvement over the standard super-resolution model from 64px to 256px, recovering high-frequency details only in the small number of denoising steps.

• Original codebase

Usage

Karlo is available in diffusers!

pip install diffusers transformers accelerate safetensors

Text to image

from diffusers import UnCLIPPipeline
import torch

pipe = UnCLIPPipeline.from_pretrained("kakaobrain/karlo-v1-alpha", torch_dtype=torch.float16)
pipe = pipe.to('cuda')

prompt = "a high-resolution photograph of a big red frog on a green leaf."

image = pipe([prompt]).images[0]

image.save("./frog.png")

Image variation

from diffusers import UnCLIPImageVariationPipeline
import torch
from PIL import Image

pipe = UnCLIPImageVariationPipeline.from_pretrained("kakaobrain/karlo-v1-alpha-image-variations", torch_dtype=torch.float16)
pipe = pipe.to('cuda')

image = Image.open("./frog.png")

image = pipe(image).images[0]

image.save("./frog-variation.png")

Model Architecture

Overview

Karlo is a text-conditional diffusion model based on unCLIP, composed of prior, decoder, and super-resolution modules. In this repository, we include the improved version of the standard super-resolution module for upscaling 64px to 256px only in 7 reverse steps, as illustrated in the figure below:

In specific, the standard SR module trained by DDPM objective upscales 64px to 256px in the first 6 denoising steps based on the respacing technique. Then, the additional fine-tuned SR module trained by VQ-GAN-style loss performs the final reverse step to recover high-frequency details. We observe that this approach is very effective to upscale the low-resolution in a small number of reverse steps.

Details

We train all components from scratch on 115M image-text pairs including COYO-100M, CC3M, and CC12M. In the case of Prior and Decoder, we use ViT-L/14 provided by OpenAI’s CLIP repository. Unlike the original implementation of unCLIP, we replace the trainable transformer in the decoder into the text encoder in ViT-L/14 for efficiency. In the case of the SR module, we first train the model using the DDPM objective in 1M steps, followed by additional 234K steps to fine-tune the additional component. The table below summarizes the important statistics of our components:

In the checkpoint links, ViT-L-14 is equivalent to the original version, but we include it for convenience. We also remark that ViT-L-14-stats is required to normalize the outputs of the prior module.

Evaluation

We quantitatively measure the performance of Karlo-v1.0.alpha in the validation split of CC3M and MS-COCO. The table below presents CLIP-score and FID. To measure FID, we resize the image of the shorter side to 256px, followed by cropping it at the center. We set classifier-free guidance scales for prior and decoder to 4 and 8 in all cases. We observe that our model achieves reasonable performance even with 25 sampling steps of decoder.

CC3M

MS-COCO

For more information, please refer to the upcoming technical report.

Training Details

This alpha version of Karlo is trained on 115M image-text pairs, including COYO-100M high-quality subset, CC3M, and CC12M. For those who are interested in a better version of Karlo trained on more large-scale high-quality datasets, please visit the landing page of our application B^DISCOVER.

BibTex

If you find this repository useful in your research, please cite:

@misc{kakaobrain2022karlo-v1-alpha,
  title         = {Karlo-v1.0.alpha on COYO-100M and CC15M},
  author        = {Donghoon Lee, Jiseob Kim, Jisu Choi, Jongmin Kim, Minwoo Byeon, Woonhyuk Baek and Saehoon Kim},
  year          = {2022},
  howpublished  = {\url{https://github.com/kakaobrain/karlo}},
}