Exciting Updates on Fooocus in SDXL

AD

APIMart

APIMart

AI API, 99.9% SLA. Your AI, Always On.

AI API, 99.9% SLA. Your AI, Always On.

Try Now

Table of Contents

1. Introduction
2. The Power of AI Animation and DALL-E3
3. The Importance of Stable Diffusion
4. Exploring SDXL - The Key to Stable Diffusion
5. Pros and Cons of SDXL
6. The Evolution of Fooocus and SDXL
7. Updates in the Anime Version of Fooocus
8. Introducing Newly Introduced Image Prompts
9. Comparing SD1.5 and SDXL
10. Generating Realistic Images with SDXL
11. Comparison of Different Refiner Models
12. Using LoRA to Customize Image Styles
13. Creating Beautiful Women with SD1.5 and LoRA
14. The Future of SDXL and Fooocus
15. Exclusive DALL-E Anime Style LoRA Images

Article

Exploring the Power of SDXL: The Key to Stable Diffusion

In the world of AI animation and image generation, the focus has primarily been on cutting-edge technologies like AI animation and DALL-E3. However, amidst all the hype, the concept of stable diffusion has taken a backseat. It's high time that we shift our Attention to SDXL, an open-source tool that may hold the key to stable diffusion and take it to the next level.

SDXL (Stable Diffusion eXtended Learning) is gaining popularity among users who understand the importance of stability in image generation. While the power of DALL-E3 is undeniable, SDXL brings a unique AdVantage to the table with its open-source nature and commitment to avoiding any inappropriate or offensive content. This ensures a safer environment for users, making it a promising tool for widespread diffusion.

I agree with the overwhelming admiration for DALL-E3's capabilities. However, it should not overshadow the importance of stable diffusion. Stable diffusion, by its nature, focuses on maintaining stability and consistency in image generation without compromising on quality. SDXL embodies this philosophy and has already proven to be a valuable asset in various applications.

One of the main contributors to SDXL's success is the rapid evolution of Fooocus, the software that is closely associated with SDXL. Fooocus has played a pivotal role in pushing SDXL towards practicality and accessibility. The latest updates in Fooocus have made SDXL a highly useful tool, both for anime-style and real-life image generation.

For those fortunate enough to have the necessary hardware setup, utilizing VRAM is highly recommended to enhance the overall performance and quality of image generation. However, it's important to note that a minimum of 12GB of VRAM is required for comfortable usage. Users with lower VRAM capacity may experience slower generation times, as demonstrated by the comparison between a 3060 laptop with 6GB VRAM and an RTX4090.

Now, let's Delve into the content we'll be exploring today. In the first part, I will introduce the updated anime version of Fooocus and the real-life version of the software. The Second part, scheduled for a future post, will focus on newly introduced image prompts and other exciting features. By watching both parts, You will be able to clearly witness the difference between SD1.5 and SDXL, prompting you to consider incorporating SDXL into your own projects.

I must say that the images I have shared on X have received positive feedback. However, I have saved the best for last: LoRA, a powerful model I created exclusively for SDXL. LoRA is specifically designed for generating DALL-E3-style anime images, and it has undergone extensive training for over 4 hours, resulting in a collection of 200 high-quality images. Although creating LoRA required 24GB of VRAM and might not be easily achievable for everyone, it will be exclusively distributed to our community members as a token of appreciation.

To ensure that every community member benefits from LoRA, I will provide the URL along with detailed instructions on how to use it. Make sure to check out Alice's X page for daily updates and access to the exclusive LoRA model.

Before we delve further into the updates in Fooocus, let's address some common questions. How can you update Fooocus? The good news is that Fooocus updates automatically. Once you launch the software, it will automatically start updating without the need for any manual intervention or executing the git pull command. This user-friendly feature saves time and makes the updating process seamless.

Two additional batch files, run anime bat and run realistic bat, will be available in the Fooocus_win64 folder once the UI starts for the first time. These files are part of the Fooocus webUI and come with preset settings specifically tailored for anime and realism. While you have the option to customize these settings, the pre-configured UI provides an optimal starting point that has been meticulously adjusted by the author.

Let's dive deeper into the anime series. Double-clicking the run anime bat file will launch the Fooocus UI with a variety of pre-installed presets for anime image generation. Upon startup, dedicated checkpoints, such as the Blue Pencil version 0.5 for SDXL and dreamshaper ver8 for SD1.5, will be available. The inclusion of blue pencil from Alice's previous recommendation indicates Fooocus's commitment to integrating the latest advancements in SDXL.

To expedite the image generation process, Fooocus eliminates the need to download checkpoints if both blue_pencil, SDXL, and dreamshaper are already present in the specified folder. Simply place the necessary models in the specified folder to save time and streamline the generation process. As a personal preference, I classify them into folders, requiring SD1.5's Dream Shaper to be in the folder for SDXL models, although it may seem a bit cumbersome. Unfortunately, specifying multiple folders to include all files doesn't work at the moment, but it would certainly be a welcome addition for seamless model management.

It's worth noting that SDXL's blue_pencil has been continuously updated, with the Current version being 0.8. The default model downloaded may change in future updates, so make sure to check the command prompt for any important notifications during startup. Now, you may be Wondering about the significance of the checkpoint dreamshaper in Stable Diffusion 1.5. This recent development allows the SD1.5 checkpoint to be utilized as a refiner model, enabling users to experiment with different styles and enhance the overall quality of their generated images.

Speaking of style, let's explore the wide range of options available in the Anime Style tab. Compared to the previous version, Fooocus now offers a greater selection of styles related to anime, such as SAI anime and SAI fantasy. To my surprise, a new masterpiece called Fooocus SAI has also been introduced.

In the settings tab, the inclusion of negative prompts demonstrates Fooocus's commitment to providing a comprehensive toolset. Negative embedding and prompts are already included, significantly reducing the time spent on manual input. The thoughtful inclusion of these presets showcases the author's dedication to user convenience and satisfaction.

Now, let's generate an image using the default settings to examine what Fooocus produces. Let's keep the prompt simple: "1girl." With that prompt in mind, let's generate an image. The result is a slightly more realistic anime image compared to SD1.5's blue_pencil, thanks to SDXL's advancements. Every Detail in the generated image appears meticulously crafted, demonstrating the impressive speeds at which Fooocus operates. Generating two images in step 30 speed mode took approximately 16 seconds with an RTX4090, a significant improvement compared to the 2 minutes and 40 seconds required by the 3060 laptop.

Next, let's compare the images generated with and without using the SD1.5 model as a refiner. By fixing the seed value and using all default settings, we can observe the impact of the refiner. The image on the left is generated without any refiner models, while the image on the right incorporates Dream Shaper as the refiner model. Comparing these images, the ones utilizing Dream Shaper exhibit softer and more refined features, primarily due to the thin Outline around facial features. However, individual preferences may vary since the sharper impressions achieved without a refiner also retain a distinctive appeal. It's important to note that these observations remain subjective, emphasizing the need to experiment and find the optimal style for each individual's preferences.

Upon closer inspection, the differences become more evident. In the image generated with the refiner model, Dream Shaper reduces the amount of detail around the blue eyes and results in a softer overall look. This effect extends to other aspects, such as the thin outline of the hair, contributing to an overall Sense of softness and refinement. Based on these observations, it becomes apparent that not every detail needs to be explicitly defined in the generation process.

Now, let's explore a comparison between SD1.5's blue_pencil and the images generated using refiner models. Using control net tiles, I created a similar image with SD1.5, showcasing the capabilities of this model. While SD1.5 produces more detailed output, the variations in line thickness and application lead to an inconsistent and slightly rough appearance. Though deliberate to achieve a distinct anime-like style, the differences between SD1.5 and SDXL become more pronounced, especially when examining details like wings.

Moving on, let's shift our focus to the refiner models themselves. Starting with SD1.5's blue_pencil for comparison, we can observe its impact on SDXL-generated images. By applying blue_pencil as the refiner model for SDXL, the resulting image appears significantly brighter, showcasing a VAE-related issue. It's important to exercise caution when combining certain models, as they may not produce the desired outcome. Although it may be possible to utilize the combination by baking VAE with the Stable Diffusion Model toolkit, this time, let's explore alternatives with different models.

One such model is counterfeit, which introduces vibrant colors and clear contrasts. By comparing the resulting image to the standard SDXL output, one can appreciate the enhanced vividness and edginess achieved by the counterfeit model. Building on that, meinamix appears to have a more subdued difference, with slightly thinner and fluffier lines. Although the details may seem subtle, they contribute to a more solid and refined overall impression. However, it's worth noting that the difference between using a refiner model and generating an image without one may be the least noticeable in meinamix.

The ability to combine SD1.5 and SDXL enables users to experiment with different painting styles and fine-tune the final output. By exploring various styles and adjusting advanced settings, users have the opportunity to Create their own templates that match their artistic vision perfectly.

Shifting our attention to the real-life system, let's explore the marvel of SDXL's incredibly smooth skin. To witness this feature in action, simply click on the real batch file to initiate the software. The realistic stock photo checkpoint, known as version 1 of the realistic Stock Photo, will be automatically downloaded.

For the realism-focused features, Fooocus offers a range of presets such as Fooocus photograph and Fooocus negative. These presets are designed to enhance the overall realism of the generated images. In terms of the model, the base model includes the previously downloaded realistic stock photo, providing a solid foundation. Additionally, a unique LoRA variant called SDXL film photography style is integrated with a strength of 0.25, further enhancing the realism. Advanced settings, including sharpness and guidance Scale, can also be adjusted to achieve desired effects. Overall, these features are carefully designed to push the boundaries of realism and bring images to life.

The settings tab within the real section offers various negative prompts, ensuring that users can explore a wide range of possibilities within the realistic domain. These meticulous settings further demonstrate the commitment to achieving the highest level of realism possible.

To generate an image, simply select the desired prompt and click on the generate button. For the purpose of comparison, we'll generate an image of a Western woman, utilizing the prompt of a French woman with blonde hair in a French braid enjoying coffee at a café in the morning. The resulting image is truly remarkable, with skin that looks and feels incredibly realistic. It's difficult to believe that the image was generated and not captured with a camera.

While the quality of the generated image is impressive, subtle imperfections such as the size and positioning of hands are still noticeable. Nevertheless, the overall appearance remains stable and Cohesive, further affirming the advancements brought by SDXL.

Zooming in on the image, the level of detail becomes strikingly evident. Every strand of hair is intricately rendered, the clothes appear highly realistic, and even the background boasts an impressive level of detail. The realism achieved by SDXL is truly remarkable and opens up new possibilities for various applications.

Let's Continue the exploration by generating an image of a Japanese woman. The prompt suggests a woman in her 20s wearing a black dress with a plunging neckline. The resulting image is equally impressive, showcasing a distinct change in facial features compared to the previous Western-style image. The Japanese woman's face exhibits softer and more delicate qualities, capturing the nuances of Asian features accurately. The attention to detail extends to accessories and other elements in the image, further enhancing the sense of realism.

To further explore the capabilities of SDXL, let's upscale the image we generated by 2x. By utilizing the upscale 2x feature with a focus on quality, we can witness the transformative power of SDXL. Upon examination, even in the upscaled image, the level of detail and realism is maintained, demonstrating the robustness of the model. The intricate depiction of eyelashes, the accurate portrayal of accessories, and the realistic rendering of facial features all contribute to the final output.

It's important to note that the size of certain features, such as the eyelashes, may vary depending on the prompt's emphasis on different ethnicities. In the case of the Japanese woman, the eyelashes appear shorter, aligning with the characteristics commonly found in the Asian population. The resulting images exhibit a natural balance, ensuring authenticity and realism.

Now that we have explored the capabilities of SDXL, let's address some common concerns. Some users may wonder if it's possible to create even more beautiful women, similar to the majicmix style. While various methodologies have been attempted, one promising approach is to utilize SD1.5's model of your preferred style as the refiner. By comparing images generated using majicmix as the refiner and the default mode, one can appreciate the effectiveness of this approach.

However, it's essential to be aware that the application of SD1.5 models in the real system may impact the benefits of SDXL. The unique characteristics of SDXL may be diluted or Altered when SD1.5 models are incorporated. Another option to create more visually appealing faces is through LoRA, a specialized model I created after extensive experimentation.

LoRA was developed by generating a large number of images using majicmix and modifying various settings to match the desired style. The training process involved feeding LoRA with 350 images, adjusting parameters such as dim and alpha. After a five-hour training session that required 24GB of VRAM, LoRA was born.

To demonstrate LoRA's capabilities, I prepared a prompt that emphasizes the features present in majicmix. By generating an image using this prompt with a default LoRA strength of 0.1, we can observe the subtle yet noticeable changes in facial features. However, it's worth mentioning that the created faces may not always meet expectations, as personal preferences and artistic choices play a significant role. While some faces may appear less appealing, there is a reasonable chance of generating a visually pleasing image by utilizing LoRA. It's important to experiment with different strength values to strike the right balance and achieve the desired outcome.

For those interested in utilizing LoRA, it will be made available for download in the summary section. Please keep in mind that the file size is quite substantial, and it may not be feasible for everyone to utilize it due to hardware constraints. Nevertheless, I encourage you to explore the potential of LoRA and fine-tune the strength parameter to personalize your image generation process.

In the next video, I will delve deeper into additional features such as image prompts and control nets. Given the comprehensive nature of these topics and the extensive content covered here, it's best to dedicate a separate discussion to them.

As we wrap up our exploration of SDXL and Fooocus, let's reflect on the remarkable images that can be achieved using these tools. In my opinion, SDXL, in combination with Fooocus, is currently the best choice for image generation. The possibilities are endless, and the results are truly captivating. If your GPU supports it, I highly recommend incorporating SDXL into your workflow. For those without the necessary hardware, I encourage you to explore alternatives and consider investing in an RTX3060 or equivalent GPU to fully experience the potential of SDXL.

I would like to express my gratitude to the community for their continued support and engagement. As a token of appreciation, I will exclusively present DALL-E anime-style LoRA images to our valued members. These images were generated using the SDXL blue_pencil model, which played a crucial role in the training process. LoRA represents the culmination of the art style and will be made available to members as a testament to their unwavering support.

This concludes today's video, and I sincerely thank you for watching until the end. If you enjoyed the content, please consider subscribing to the Channel and hitting the like button. Your support means the world to me. Look forward to the upcoming video, where I will delve deeper into the topic of SDXL and Fooocus. Goodbye for now, and see you soon!

Highlights

• Discover the power of SDXL in stable diffusion for image generation.
• Explore the updates in anime version Fooocus and real version Fooocus.
• Understand the importance of using VRAM for optimal performance.
• Compare the differences between SD1.5 and SDXL in image refinement.
• Generate realistic images with SDXL and explore different refiner models.
• Customize image styles using LoRA and its impact on facial features.
• Witness the transformative power of SDXL in upscaling images.
• Enhance image generation by utilizing SD1.5 models as refiners.
• Receive exclusive DALL-E anime-style LoRA images as a community member.

FAQ

Q: Can Fooocus be updated automatically? A: Yes, Fooocus updates automatically upon launching the software, eliminating the need for manual intervention.

Q: How can I use SD1.5 as a refiner model with SDXL? A: By specifying SD1.5's Dream Shaper checkpoint as the refiner model, you can enhance the quality and style of the generated images.

Q: Can I incorporate my preferred painting style in SDXL? A: Yes, by utilizing SD1.5 as a refiner model, you can customize the painting style and experiment with different styles to personalize your images.

Q: How can I generate more visually appealing faces with SDXL? A: You can utilize LoRA, a specialized model created for SDXL. LoRA can be trained with specific prompts to achieve your desired style and facial features.

Q: Are there any exclusive benefits for community members? A: Yes, exclusive DALL-E anime-style LoRA images will be made available to community members as a token of appreciation. These images represent the culmination of the art style and will only be accessible to members.