Optimized Text-to-Image Synthesis Using Stable

Diffusion and the Diffusers Library

B Ashreetha
Electronics and Communication Thirugudu Shiva Kumar Yadav S Sreedhar
Engineering Electronics and Communication Electronics and Communication
Mohan Babu University (Erst.while Sree Engineering Engineering
Vidyanikethan Engineering College) Mohan Babu University(Erst.while Sree Mohan Babu University(Erst.while Sree
Tirupati, India Vidyanikethan Engineering College) Vidyanikethan Engineering College)
[email protected] Tirupati, India Tirupati, India
[email protected] [email protected]
V M Siddeswara
Electronics and Communication Ogeti Sudharani
Engineering Electronics and Communication
Mohan Babu University(Erst.while Sree Engineering
Vidyanikethan Engineering College) Mohan Babu University(Erst.while Sree
Tirupati, India Vidyanikethan Engineering College)
[email protected] Tirupati, India
[email protected]
Abstract— Text-to-image generation makes pictures from beautiful images. Because generative models can generate
content, making a difference in areas like media, promoting, high-quality synthetic data, including text, images, and even
and instruction. In any case, to get precise pictures from the three-dimensional objects, they have revolutionized the area
given content accurately had been a challenge, as numerous of machine learning. Among these, diffusion models are
models make non aliased or wrong comes about. A few
notable for their superior quality and robustness in task
strategies exist for text-to-image generation, counting GAN-
based models, which utilize ill-disposed preparing to refine
generation. This method has been further refined using
picture quality, and VAE-based strategies, which encode stable diffusion and latent diffusion models, enabling high-
content into a inactive space but frequently struggle with high- resolution and computationally effective image generation.
resolution points of interest. Ordinary models face challenges Diffusion models offer a stable and comprehensible
in semantic arrangement and regularly generate low-fidelity or framework for generative modeling, in contrast to
artifact-prone pictures. Stable Diffusion mitigates these Generative Adversarial Networks (GANs) and Variational
confinements through progressive denoising, parameterized Autoencoders (VAEs), which iteratively modify outputs
control, and computational effectiveness, enabling high- through a stochastic process. However, these are still in
resolution and semantically coherent outputs. This study
existence. Mode collapse and instability during training
employments Stable Diffusion with the Diffusers package to
produce high-quality images by turning content into visual limit the diversity and reliability of GANs, even if they can
patterns and refining them with user controls. The pictures are produce high-quality images. On the other hand, VAEs
tried for clarity, exactness, and productivity, appearing that usually result in unclear outputs by compromising
Stable Diffusion can make important visuals. interpretability. Early diffusion models addressed some of
these issues, but they were prohibitively expensive and
Keywords — Stable Diffusion, Text-to-Image Generation, necessitated a significant amount of resources for thorough
Diffusers Library, Visual Content Creation, Generative AI, Media inference and training.
Applications etc.
The necessity to address these issues by utilizing the
diffusion models is what inspired this effort.
I. INTRODUCTION
Stable Diffusion's public release has made high-quality
One of the most remarkable recent advances in computer generative models more accessible to a wider audience and
vision is image synthesis, which also has one of the highest demonstrated their promise for a variety of uses, including
processing needs. At the moment, scaling up likelihood- scientific visualization, advertising, and the arts. Finding a
based models—which may include several parameters in balance between user control, output quality, and computing
autoregressive (AR) transformers—dominate the synthesis efficiency is a challenge, though. By leveraging recent
of intricate, natural landscapes at exceptionally high advancements in diffusion modeling, including textual
resolutions. On the other hand, it has been shown that inversion methods and latent space optimizations, this study
because GANs' adversarial learning approach is not well aims to close these gaps and provide a useful, adaptable, and
suited to modeling intricate, multi-modal distributions, their scalable framework for text-to-image generation.
encouraging outcomes are primarily restricted to data with This paper's primary goals are to:
comparable low levels of variability. Multimodal learning 1. Use the Diffusers library to implement and optimize
has gained prominence in recent years, with image-text stable diffusion.
contrastive learning and text-to-image synthesis at the 2. To investigate diffusion models' scalability and
forefront. These models' state-of-the-art image generation computational effectiveness.
and editing capabilities have revolutionized science and 3. To offer more control over produced outputs by
sparked intense public interest. To advance this field of allowing for customization and picture restrictions.
study, a steady diffusion model from text to image was 4. To examine possible uses of stable diffusion in a
introduced. It achieves previously unheard-of levels of range of fields, including advertising, media, and
linguistic comprehension and photorealism by fusing the education.
strength of transformer language models (LMs) with high- The following contributions are made by this paper:
fidelity diffusion models for text-to-image synthesis. 1. A thorough implementation of stable diffusion that
Imagen's primary discovery is that, in contrast to earlier makes use of the Diffusers library to increase
studies that only employed image-text data for model flexibility and efficiency.
training, text embeddings from large LMs that were trained 2. An empirical evaluation of the model's functionality,
with text alone are remarkably effective for text-to-image taking into account output quality and computing
synthesis. To transform input text into a series of efficiency.
embeddings, Imagen uses a frozen T5-XXL encoder, two 3. Including user-specified limitations to enhance
Super Resolution diffusion models, and a 64x64 image control and customization of image outputs.
diffusion model. Choose from a range of text inputs using 4. Proof of the suggested framework's usefulness in
1024 x 1024 Imagen samples. Each diffusion model practical situations, such as the production of
employs classifier-free steering and is reliant on the text instructional materials and artistic content.
embedding sequence. Imagen employs unique sampling The structure of the paper is as follows: The literature
methods to produce images with higher fidelity and better survey is presented in Section II, which examines earlier
text-image alignment than was previously feasible. This research and significant advancements in text-to-image
makes it possible to employ big guide weights without synthesis, diffusion models, and generative models,
suffering from the decline in sample quality noted in more stressing their advantages and disadvantages. The
recent research. These are primarily used to turn text into
 significance of Stable Diffusion in the context of customized outcomes by giving the user more control over
generative modeling is covered in Section III, which also the generative process [7].
highlights its benefits over more conventional models Hosni (2022) described the setup, application
such as GANs and VAEs, especially with regard to cases, and optimization strategies for Stable Diffusion,
stability and quality. The suggested technique for offering useful insights into its implementation.
implementing text-to-image generation using Stable Practitioners who want to efficiently use Stable Diffusion
Diffusion is presented in Section IV, which also covers will find this material useful [8].
model setup, text encoding, and the incorporation of user- Ledig et al. (CVPR 2017) concentrated on GAN-
defined restrictions. The implementation procedure is based super-resolution challenges. Their methodology
explained in Section V, which also covers the workflow, informs methods for improving image quality, a crucial
technical components of the model setup, and efficiency- component of generative models, even though they are
boosting optimization strategies used. The results and different from diffusion models [9].
debates are presented in Section VI, which also compares In their discussion of utilizing GANs to learn latent
the suggested system's performance with current models spaces for 3D shapes, Wu et al. (NeurIPS 2019) illustrated
and analyzes the effects of the limitations in producing the wider use of generative models. The design of latent
high-quality photographs. Section VII, which wraps up diffusion techniques is also influenced by these ideas [10].
the study, summarizes the main conclusions, talks about A method for learning robust deep representations
the contributions made, and looks at potential future by optimizing mutual information between input data and
developments and uses for the subject of generative. high-level features was presented by Hjelm et al. (ICLR
2019). By maintaining significant data structures, this
II. RELATED WORKS approach provides fundamental ideas that can improve
Chen et al. (ICLR 2021) presented a new method for representation learning in diffusion models, even though it
creating images that makes use of stable diffusion processes. is not exclusive to generative models [11].
This study offered fundamental understandings of how An adaptive significance sampling method was
diffusion models may produce high-quality images and presented by McDermott and Mahoney (2021) to increase
stabilize training. In order to improve stability and diversity, the effectiveness of diffusion-based generative models. This
the methodology focused on using diffusion processes to method is very useful for optimizing large-scale diffusion
iteratively adjust images [1]. systems since it reduces computational costs without
Diffusion models and GANs were thoroughly sacrificing picture creation fidelity [12].
compared by Grathwohl et al. (ICLR 2021). They showed In order to achieve high-quality image synthesis,
that diffusion models perform better than GANs in terms of Brock et al. (ICLR 2018) investigated techniques for
robustness, diversity, and image quality. Diffusion-based training GANs at scale. In high-fidelity generation tasks,
models were established as a better option for picture their methods—such as memory optimization and the
synthesis tasks thanks in large part to this study [2]. incorporation of larger datasets—offer a standard by which
A significant advancement in accessible generative to compare GANs and diffusion models [13].
AI was made when Stability AI released Stable Diffusion to The application of self-supervised learning to
the public. Stable Diffusion's scalability and versatility for image synthesis and modification tasks was highlighted by
producing high-quality, configurable visual material were Zhang et al. (CVPR 2020). Improved comprehension of
highlighted in the blog article, which also covered its latent spaces is made possible by the suggested methods,
installation and applications [3]. and this can be used to improve the control and
StyleGAN inversion with hypernetworks was interpretability of diffusion-based models [14].
investigated by Alaluf et al. (2022), enabling fine-grained GANs were used for super-resolution tasks by
actual picture manipulation. This study adds to the area by Ledig et al. (CVPR 2017), producing high-quality images
presenting techniques for improving and modifying from low-resolution inputs. Despite having their roots in
generated images, demonstrating hybrid approaches, even if GANs, these techniques provide inspiration for diffusion
it is not explicitly diffusion-based [4]. models to tackle related problems in improving image
Advances in GAN training were demonstrated by quality [15].
Brock et al. (ICLR 2019), allowing for the large-scale In order to overcome the drawbacks of
synthesis of high-fidelity pictures. The methodologies autoencoder-based interpolation, Brock et al. (ICLR 2019)
served as a standard by which to evaluate more recent introduced an adversarial regularizer. Diffusion models that
approaches, such diffusion models [5]. use latent space traversals benefit from this method since it
By working in latent space as opposed to pixel enhances the consistency and smoothness of interpolated
space, latent diffusion models—which were first presented outputs, which informs generative modeling [16].
by Rombach et al. in 2022—optimize stable diffusion. This In order to bridge the gap between generative and
method established its usefulness in text-to-image discriminative tasks, Brock et al. (2016) studied voxel-based
generation by drastically lowering computational costs modeling utilizing CNNs. Although voxel modeling is
while preserving high-resolution outputs [6]. primarily concerned with 3D data, its principles can also be
By fine-tuning diffusion models with less data, Gal used to 2D generating processes in diffusion models [17].
et al. (2022) suggested textual inversion strategies to A thorough tutorial on using Stable Diffusion to
customize text-to-image production. This provided produce visually striking photographs was given by Nick
Babich in 2023. By highlighting best practices and use
scenarios, this useful tool helps practitioners maximize
diffusion model outputs for both commercial and creative A compressed version of the input image is intended for
purposes [18]. learning by an autoencoder. An encoder and a decoder are
Textual inversion approaches were presented by the two primary components of a variational autoencoder.
Gal et al. (2022) as a way to customize image production The image must be compressed into a latent form by the
with little training data. By enabling customized outputs and encoder, and the original image must then be recreated by
matching produced graphics with particular user needs, our the decoder using this latent representation.
work improves diffusion models [19]. b) U-NeT
Yossef Hosni (2022) provided an easy-to-read U-NeT A type of convolutional neural network
overview of stable diffusion, outlining its setup and uses. (CNN) called U-Net is used to clean up an image's latent
Both academics and developers will find the tutorial representation. It consists of a sequence of encoder-decoder
appealing as it emphasizes how simple it is to create blocks that gradually improve the quality of the image. After
diffusion models for a variety of image synthesis tasks [20]. the network's encoder lowers the image's resolution, the
In order to increase output quality and computational decoder attempts to restore the compressed image backup to
efficiency, Rombach et al. (2022) improved Stable its original, higher resolution while also removing any noise.
Diffusion using latent diffusion models, which function in c)Encoder for Text
compressed latent spaces. Diffusion models were Transforming text prompts into a latent form is the
established as the cutting-edge method for text-to-image task of the text encoder. Usually, a transformer-based model
generation by this ground-breaking work [21]. is used to accomplish this, like CLIP's Text Encoder, which
. converts a string of input tokens into a string of latent text
embeddings. To create an image from a word prompt, we
III. IMPORTANCE OF STABLE DIFFUSION must write a lot of code using the massive Stable Diffusion
A paradigm for creating visuals from text was presented, model framework. HuggingFace has addressed this issue by
called Stable Diffusion. This creative method builds pictures introducing diffusers. With just a few lines of Python code
from textual descriptions by using diffusion techniques. and Diffusers, we can quickly create a large number of
Apart from creating images, it may also be used for other images without having to bother about the underlying
activities like painting in and out and translating images to architecture. In this instance, the Diffusers library's state-of-
another image when given a text input. the-art StableDiffusionPipeline will be used.
One kind of generative model is a diffusion model, which is
trained to extract a sample of interest by de-noising an IV. PROPOSED METHOD
object, like an image. Until a sample is collected, the model The suggested approach makes use of stable diffusion's
is trained to de-noise the image a little at a time. In order to properties to create a framework for text-to-image
produce a final image that complies with the request, it first generation that is both effective and adaptable. The method
paints the image with noise and random pixels. It then starts by using the Diffusers package to make the Stable
attempts to eliminate the noise by modifying each phase. Diffusion pipeline deployment process more efficient. The
pre-trained model, which serves as the basis for producing
high-quality photos, is loaded first. The pipeline
incorporates a text encoder, which enables the system to
translate textual prompts into latent representations. By
directing the diffusion process, these encoded
representations make sure that the produced visuals closely
match the input text. The technique includes an optional
stage for specifying picture restrictions to provide users
more control over the process. This step allows users to
Fig. 1. Architectrue of stable Diffusion
select aspects like aspect ratio, style, and resolution. This
stage increases the framework's usefulness across a range of
The foundation of Stable Diffusion is a diffusion
applications by guaranteeing that the outputs satisfy
model called Latent Diffusion, which is renowned for its
particular requirements. High-resolution, coherent images
sophisticated capabilities in picture synthesis, particularly in
are then produced in the latent space by the Stable Diffusion
tasks like text-to-image production, image painting, and
model's repeated refinement process. By focusing on
style transfer. Latent diffusion incorporates cross-attention
computational efficiency, the framework lowers resource
layers into its architecture, in contrast to other diffusion
requirements without sacrificing output quality. The
models that only concentrate on pixel modification. The
suggested approach overcomes current constraints in text-to-
model may absorb data from multiple sources, such as text
image synthesis by incorporating scalable features and
and other inputs, thanks to these layers.
customizing possibilities, making it appropriate for a variety
Latent diffusion consists of three primary parts:
of real-world applications, including media content
a) An autoencoder generation, advertising, and instructional tools.
b) U-Net
c) Encoder for Text
a)An autoencoder
 Fig. 2. Proposed Block Diagram customizable images, addressing the limitations of previous
generative approaches.
A. Methodology B. Components of Stable Diffusion
The methodology for implementing the proposed text-to- Stable diffusion is a system with several components and
image generation framework involves a systematic process concepts. There are multiple models. When we look at the
that integrates Stable Diffusion with the Diffusers library. internal mechanisms, the first thing we identify is the
The following steps outline the approach: presence of a text-understanding component that transforms
1) Model Initialization the textual data into a numerical representation that captures
The framework begins by loading a pretrained Stable the ideas in the text.
Diffusion model using the Diffusers library. This library
provides an optimized and modular interface for deploying
diffusion-based generative models, ensuring efficient
execution and scalability. The model is initialized with
default parameters, which can be adjusted based on the
application requirements, such as resolution or
computational constraints. Fig. 3. Stable Diffusion Components

2) Text Encoding The text encoder is a special Transformer language model.

A text encoder, typically based on a pre-trained A list of numbers (a vector per token) is used to represent
transformer model (e.g., CLIP or BERT), is employed to each word or token in the input text. The information is
convert input text prompts into latent representations. These subsequently sent to the Image Generator, which consists of
representations act as semantic guides for the diffusion several components.
model, ensuring the generated image accurately reflects the Steps that are taken by the image generator:
textual description. The text encoder is fine-tuned, if 1) A creator of image information
necessary, to improve alignment with the visual domain. This component, which generates image information
3) Latent Diffusion Process by running for numerous steps—the steps option in Stable
The Stable Diffusion model performs the image Diffusion interfaces and libraries usually defaults to 50 or
generation in the latent space, significantly reducing 100—is the key to Stable Diffusion.Technically speaking,
computational overhead compared to pixel-space diffusion this component is made up of a UNet neural network and a
models. The process begins with random noise in the latent scheduling algorithm. The process of processing
space, which is iteratively denoised under the guidance of information step-by-step to produce a high-quality image
the encoded text representation. Each iteration refines the (by the subsequent component, the image decoder) is
image, progressively moving closer to the desired output. referred to as "diffusion" in this component.
4) Image Constraints and Customization .
To enable user-defined control, the framework
incorporates a mechanism for specifying image constraints.
These constraints may include parameters such as style,
color schemes, resolution, or object placements. This step
enhances the flexibility of the system, allowing it to cater to
diverse application requirements.
5) Image Generation
After the diffusion process completes, the model
generates a high-resolution image that aligns with the input Fig. 4. Image Information Creater
text and adheres to the specified constraints. The final
output is decoded from the latent space back to pixel space, 2) Image Decoder
producing a coherent and detailed image. Using the information it received from the information
6) Performance Optimization producer, the image decoder creates a picture. It just
To improve efficiency, the framework employs executes once to create the final pixel image at the
techniques such as mixed-precision computation and model conclusion of the operation.
quantization. These optimizations reduce memory usage and
computation time, making the system suitable for
deployment on resource-constrained environments.
7) Evaluation and Analysis
The generated outputs are evaluated based on criteria
such as image quality, semantic alignment with the input
text, and adherence to user-defined constraints. Metrics like
Fréchet Inception Distance (FID) and human perceptual
evaluation are used to assess the performance.
This methodology ensures that the proposed framework
is efficient, scalable, and capable of producing high-quality, Fig. 5. Image Decoder
• To gradually process and distribute data in the latent
End
End
information space, use U Net + Scheduler.
The inputs are an initial multi-dimensional array made up of
text embeddings and noise (organized lists of integers, Fig. 7. Implementation of the latent diffusion process
sometimes referred to as a tensor).
The result is an array of processed information.
• An autoencoder decoder that generates the final image VI. RESULTS AND DISCUSSION
using the information array that has been processed. An image created in response to the supplied text prompt is
The array of processed data is the input. the end result of a text-to-image. The purpose of this image
is to illustrate the information that is stated in the given
prompt. The evaluation of text-to-image models was carried
out using the CLIP Score to measure text-image alignment
and the FID Score to evaluate image realism as outlined in
Fig. 9.

Fig. 6. Encoding and decoding the information

V. IMPLEMENTATION
The foundation of the system is formed at the beginning of
the startup phase by loading a pretrained Stable Diffusion
model. The next step is text encoding, which uses a Fig. 8. Generated Image from text was kids dancing in rain
transformer-based paradigm like CLIP to convert the input
text prompt into latent representations that control the In comparison, GAN (CLIP:0.70, FID:18.2) and VAE
creation of images. The flowchart includes a decision point (CLIP:0.60, FID:25.4) demonstrated low text coherence and
to ascertain whether the user has established constraints, image quality. DALL-E 2 (CLIP:0.88, FID:12.5) displayed
such as style, resolution, or certain features. If limitations strong text alignment but had slightly lower realism. Stable
are specified, they are utilized to ensure that the final image Diffusion models surpass those based on GAN and VAE in
conforms with them. Other than that, the process remains both metrics. Stable Diffusion of the model 2.1 has the
unchanged. The core mechanism of the system is the latent metrics of (CLIP: 0.87,FID 9.5) and similarly for SD 2.1 has
diffusion process, which iteratively refines random noise in the metrics of (CLIP:0.90, FID:8.2) achieved the incredible
the latent space under the influence of the encoded text, comes around, showcasing prevalent realism and semantic
progressively transforming it into a meaningful image. Once alignment.
the diffusion process is complete, the generated output from
the latent space is decoded back into pixel space to produce
the final high-resolution image.

Start
Start

Load Pre-trained Stable Diffusion Model

Fig. 9. CLIP Score vs FID Score for various models

Text Encoding with Transformer

User
User
Specifie
Specifie
dd
Constrai
Constrai
nts
nts

Apply user Constraints Skip to Next Step Fig. 10. Speed and Quality for various models

Latent diffusion Process

Fig.10 shows the VRAM Usage vs. Inference Time
Comparison, Stable Diffusion models accomplish a balance
between effectiveness and speed, offering moderate VRAM
Generate Final Image
utilization with reasonable deduction times. In contrast,
GANs provide quicker generation but require higher [6] Chen, T. Q., Li, X., Grosse, R. B., & Duvenaud, D. (2018). Isolating
sources of disentanglement in variational autoencoders. Advances in
memory, whereas VAEs are slower despite lower VRAM Neural Information Processing Systems (NeurIPS), 5594-5603.
usage. This demonstrates Stable Diffusion's ideal trade-off [7] Karras, T., Aila, T., Laine, S., & Lehtinen, J. (2018). Progressive growing
for scalable text-to-image generation. of GANs for improved quality, stability, and variation. arXiv preprint
arXiv:1710.10196.
[8] Chen, Yankun & Liu, Jingxuan & Peng, Lingyun & Wu, Yiqi & Xu, Yige
VII. CONCLUSION AND FUTURE SCOPE & Zhang, Zhanhao. (2024). Auto-Encoding Variational Bayes.
The use of Stable Diffusion in a text-to-image production [9] M. Mehralian and B. Karasfi, "RDCGAN: Unsupervised Representation
Learning With Regularized Deep Convolutional Generative Adversarial
system shows how diffusion-based models may effectively Networks," 2018 9th Conference on Artificial Intelligence and Robotics
generate high-quality, semantically aligned, and [10] Wu, J., Zhang, C., Xue, T., Freeman, B., & Tenenbaum, J. (2019).
configurable images. This technique overcomes the Learning a probabilistic latent space of object shapes via 3D generative-
drawbacks of traditional approaches, including instability in adversarial modeling. Advances in Neural Information Processing
GANs and low fidelity in VAEs, by utilizing the Diffusers Systems (NeurIPS), 82-92.

library and latent diffusion processes. Furthermore, adding [11] R. D. Hjelm, A. Fedorov, S. Lavoie-Marchildon, K. Grewal, A. Trischler,
and Y. Bengio, "Learning deep representations by mutual information
user-defined constraints increases the framework's estimation and maximization," in Proceedings of the International
adaptability, making it appropriate for use in a variety of Conference on Learning Representations (ICLR), 2019.
industries, like advertising, media, and education. The [12] S. D. McDermott and M. W. Mahoney, "Adaptive importance sampling
outcomes demonstrate the system's utility by demonstrating for diffusion-based generative models," arXiv preprint arXiv:2102.02760,
2021.
its capacity to produce images that not only follow textual
[13] A. Brock, J. Donahue, and K. Simonyan, "Large scale GAN training for
instructions but also satisfy user needs. high fidelity natural image synthesis," in Proceedings of the International
Furthermore, adding multimodal input support to Conference on Learning Representations (ICLR), 2018.
the model, integrating text with pre-existing photos or [14] Y. Zhang, Y. Zhang, J. Wen, and Y. Li, "Self-supervised learning for
sketches, can lead to new creative applications. Finally, as image synthesis and manipulation," in Proceedings of the IEEE/CVF
generative technologies become more widely used, it will be Conference on Computer Vision and Pattern Recognition (CVPR), 2020.

essential to investigate ethical issues and make sure they are [15] C. Ledig, L. Theis, F. Huszár, J. Caballero, A. Cunningham, A. Acosta, A.
Aitken, et al., "Photo-realistic single image super-resolution using a
used responsibly. This work lays the groundwork for future generative adversarial network," in Proceedings of the IEEE Conference
developments in text-to-image generation. on Computer Vision and Pattern Recognition (CVPR), 2017).
[16] A. Brock, J. Donahue, and K. Simonyan, "Understanding and improving
REFERENCES interpolation in autoencoders via an adversarial regularizer," in
[1] Generative Models and the Stabilizing Diffusion" by T. Q. Chen, et al., Proceedings of the International Conference on Learning Representations
(ICLR 2021) introduced a novel method for image generation using the (ICLR), 2019.
stable diffusion process. [17] G. Yang, F. Fu, N. Fei, H. Wu, R. Ma and Z. Lu, "DiST-GAN:
[2] Diffusion Models Beat GANs on Image Synthesis" by D. Grathwohl, et Distillation-based Semantic Transfer for Text-Guided Face Generation,"
al., (ICLR 2021) compared the performance of GANs and diffusion 2023 IEEE (ICME), Brisbane, Australia, 2023, pp. 840-845, doi:
models for image synthesis. 10.1109/ICME55011.2023.00149.
[3] A.I., S.: Stable diffusion public release, https://stability.ai/blog stable- [18] Nick babich, How to generate stunning images using stable diffusion 09
diffusion-public-release jan 2023.
[4] Alaluf, Y., Tov, O., Mokady, R., Gal, R., Bermano, A.H.: Hyperstyle: [19] Rinon Gal, Yuval Alaluf, Yuval Atzmon, An Image is Worth One Word:
Stylegan inversion with hypernetworks for real image editing. Personalizing Text-to-Image Generation using Textual Inversion
arXiv:2111.15666 [cs] (03 2022) 2Aug2022 arXiv:2208.01618 Cornell university.
[5] Brock, A., Donahue, J., & Simonyan, K. (2019). Large scale GAN training [20] Yossef hosni, Getting Started With Stable Diffusion Nov11 2022.
for high fidelity natural image synthesis. Proceedings of the International [21] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser,
Conference on Learning Representations (ICLR). Björn Ommer, High-Resolution Image Synthesis with Latent Diffusion
Models 13 Apr2022 arXiv:2112.10752.