This is a Fine-tuned Stable Diffusion model on a custom dataset of {image, caption} pairs. This model has been built on top of the fine-tuning script provided by Hugging Face. This model uses the KerasCV implementation of stability.ai's text-to-image model. Unlike other open-source alternatives like Hugging Face's Diffusers, KerasCV offers advantages such as XLA compilation and mixed precision support, resulting in state-of-the-art generation speed.
This model serves as a fine-tuned version of stability.ai's v1-4 Stable Diffusion model for generating high-quality Renaissance-style portraits.
It is finetuned from KerasCV's implementation of Stable Diffusion.
Keras CV is a deep learning library that is built on top of TensorFlow and Keras.
It provides a number of pre-trained models for image classification, object detection, and segmentation.
The Keras CV implementation of Stable Diffusion is a simple and easy-to-use way to generate images from text.
To use the model, you simply need to provide a text prompt and the model will generate an image that matches the prompt.
In the specific case of this fine-tuned model, upon any prompt input the model is capable of generating an image resembling that of a Renaissance era portrait.
Developed by: Martin Gasparyan, Tatev Kyosababyan
Shared by: Martin Gasparyan, Tatev Kyosababyan
Model type: Computer Vision Model
Language(s) (NLP): eng
License: creativeml-openrail-m
Parent Model: CompVis/stable-diffusion-v1-4
Resources for more information: More information needed
The model is intended for research purposes only. Possible research areas and tasks include
Safe deployment of models which have the potential to generate harmful content.
Probing and understanding the limitations and biases of generative models.
Generation of artworks and use in design and other artistic processes.
Applications in educational or creative tools.
Research on generative models.
Excluded uses are described below.
Out-of-Scope Use
The model was not trained to be factual or true representations of people or events, and therefore using the model to generate such content is out-of-scope for the abilities of this model.
Misuse and Malicious Use
Using the model to generate content that is cruel to individuals is a misuse of this model. This includes, but is not limited to:
Generating demeaning, dehumanizing, or otherwise harmful representations of people or their environments, cultures, religions, etc.
Intentionally promoting or propagating discriminatory content or harmful stereotypes.
Impersonating individuals without their consent.
Sexual content without consent of the people who might see it.
Mis- and disinformation
Representations of egregious violence and gore
Sharing of copyrighted or licensed material in violation of its terms of use.
Sharing content that is an alteration of copyrighted or licensed material in violation of its terms of use.
Bias, Risks, and Limitations
Limitations
The model does not achieve perfect photorealism
The model cannot render legible text
The model does not perform well on more difficult tasks which involve compositionality, such as rendering an image corresponding to “A red cube on top of a blue sphere”
Faces and people in general may not be generated properly.
The model was trained mainly with English captions and will not work as well in other languages.
The autoencoding part of the model is lossy
The model was trained on a large-scale dataset LAION-5B which contains adult material and is not fit for product use without additional safety mechanisms and considerations.
No additional measures were used to deduplicate the dataset. As a result, we observe some degree of memorization for images that are duplicated in the training data. The training data can be searched at https://rom1504.github.io/clip-retrieval/ to possibly assist in the detection of memorized images.
Bias
While the capabilities of image generation models are impressive, they can also reinforce or exacerbate social biases. Stable Diffusion v1 was trained on subsets of LAION-2B(en), which consists of images that are primarily limited to English descriptions. Texts and images from communities and cultures that use other languages are likely to be insufficiently accounted for. This affects the overall output of the model, as white and western cultures are often set as the default. Further, the ability of the model to generate content with non-English prompts is significantly worse than with English-language prompts.
Training Details
Stable Diffusion v1-4 is a latent diffusion model which combines an autoencoder with a diffusion model that is trained in the latent space of the autoencoder. During training,
Images are encoded through an encoder, which turns images into latent representations. The autoencoder uses a relative downsampling factor of 8 and maps images of shape H x W x 3 to latents of shape H/f x W/f x 4
Text prompts are encoded through a ViT-L/14 text-encoder.
The non-pooled output of the text encoder is fed into the UNet backbone of the latent diffusion model via cross-attention.
The loss is a reconstruction objective between the noise that was added to the latent and the prediction made by the UNet.
Training Data
We used 11 Renaissance portraits to train the model and created a .csv file with two columns, one for image path and the other for textual description. Dataset can be found at https://huggingface.co/datasets/morj/renaissance_portraits and can be downloaded using
curl -X GET \
"https://datasets-server.huggingface.co/splits?dataset=morj%2Frenaissance_portraits"
Training Procedure
Note: Only the diffusion model is fine-tuned. The VAE and the text encoder are kept frozen.
Training details: The fine-tuning process involves adapting the Stable Diffusion model to the specific task of generating Renaissance-style portraits from textual descriptions.
The dataset we trained our model on can be found here. We used 11 Renaissance portraits to train the model and created a .csv file with two columns, one for image path and the other for textual description.
When launching training, a diffusion model checkpoint is generated epoch-wise only if the current loss is lower than the previous one. To avoid OOM and faster training, we used an A100 GPU in Google Colab.
We fine-tuned the model on two different resolutions: 256x256 and 512x512. We only varied the batch size and number of epochs for fine-tuning with these two different resolutions. The best results were obtained with 512 x 512 pixels, 72 epochs, batch size of 1 and mixed precision set to True.
Hardware: A100 GPU
Optimizer: AdamW
Batch: 1
Learning rate: warmup to 0.0001 for 10,000 steps and then kept constant
Experiments were conducted using Google Cloud Platform in region us-west1, which has a carbon efficiency of 0.3 kgCO$_2$eq/kWh. A cumulative of 50 hours of computation was performed on hardware of type A100 PCIe 40/80GB (TDP of 250W).
Total emissions are estimated to be 3.75 kgCO$_2$eq of which 100 percents were directly offset by the cloud provider.
%Uncomment if you bought additional offsets:
%XX kg CO2eq were manually offset through \href{link}{Offset Provider}.
@article{lacoste2019quantifying,
title={Quantifying the Carbon Emissions of Machine Learning},
author={Lacoste, Alexandre and Luccioni, Alexandra and Schmidt, Victor and Dandres, Thomas},
journal={arXiv preprint arXiv:1910.09700},
year={2019}
}
5. Create a variable to hold the values of the to-be-generated image such as prompt, batch size, iterations, and seed
python
1img = my_base_model.text_to_image(2 prompt='A woman with an enigmatic smile against a dark background',3 batch_size=1,# How many images to generate at once4 num_steps=25,# Number of iterations (controls image quality)5 seed=123,# Set this to always get the same image from the same prompt6)
