Fooocus Update 2.4.0 - Custom VAEs, Hyper-SD, Random Styles, and More

Kleebz Tech AI
26 May 202408:06

TLDRIn the Fooocus Update 2.4.0 video, Rodney from Kleebz Tech covers new features like Hugging Face file downloads, Interposer upgrade, reconnect button, custom VAE selection, random style option, updated anime model, full prompt visibility in log files, auto image description, NSFW filter, LoRAs in prompts, new schedulers, inpainting brush color change, clip layer bypass, and Hyper SD option. He also discusses potential live streaming and thanks Mashb1t for updates and Patreon supporters.

Takeaways

  • 🆕 Fooocus has released version 2.4.0, introducing several updates and new features.
  • 📡 Supports downloading Hugging Face files from a mirror site for backup purposes.
  • 🚀 Interposer has been updated from version 3.1 to 4.0, enhancing compatibility between SD 1.5 and SDXL models.
  • 🔄 A new reconnect button allows users to resume their work after a crash or accidental closure.
  • 🎨 Users can now select a custom VAE for decoding from the latent space, which can improve image quality.
  • 🌟 Introduces the option to select a random style for image generation, adding variety to outputs.
  • 🎭 The default anime model has been updated to a newer version, improving the quality of anime-style images.
  • 📑 Log files now display the full raw prompts, providing more transparency into the generation process.
  • 🖼️ An option to automatically describe an image when dropped into the variation and upscale section has been added.
  • 🚫 A new NSFW filter can be enabled to prevent explicit content from being generated.
  • 🔤 LoRAs can now be included directly in the prompt for more control over image generation.
  • ⏱️ New schedulers have been added for different image generation speed preferences.
  • 🖌️ The inpainting brush color can now be changed, making it easier to distinguish the mask from the image.
  • 👁️ The ability to bypass clip layers has been added, allowing users to skip certain layers in the clip model for faster generation.

Q & A

  • What is the latest version of Fooocus mentioned in the video?

    -The latest version of Fooocus mentioned in the video is 2.4.0.

  • What is the new feature related to downloading Hugging Face files?

    -The new feature allows support for downloading Hugging Face files from a mirror site, which serves as a backup for the primary download site.

  • What is the significance of the Interposer update from 3.1 to 4.0?

    -The Interposer update to version 4.0 is designed to enable interoperability between SD 1.5 and SDXL Stable Diffusion models, allowing them to generate images from random noise despite their different architectures and latent spaces.

  • What does the new reconnect button in Fooocus do?

    -The new reconnect button allows users to reestablish their session with Fooocus after a crash or accidental closure of the command window, without losing any settings they were adjusting.

  • How can users select a custom VAE in the updated Fooocus?

    -Users can select a custom VAE by downloading one and placing it in the required directory, then choosing it from the dropdown in the debug menu.

  • What is the new style option that has been added to Fooocus?

    -The new style option allows Fooocus to randomly select a style in addition to any styles the user has chosen, resulting in a different style for each image generated.

  • What update has been made to the default anime model in Fooocus?

    -The default anime model has been updated from Anime Pencil XL version 1 to version 3.1.

  • How has the log file been improved in the Fooocus update?

    -The log file now displays the full raw prompts, allowing users to see exactly what was added to the prompt from the styles.

  • What is the new feature related to image description in Fooocus?

    -Users can now enable a feature that automatically describes an image if an image is dropped into the variation and upscale section with a blank prompt. This requires editing the bat file to add a flag.

  • What is the purpose of the new NSFW filter in Fooocus?

    -The NSFW filter prevents explicit content from appearing on the screen. When enabled, if NSFW content is detected, the saved image will be completely black.

  • How can users include LoRAs in their prompt in the updated Fooocus?

    -Users can now include LoRAs in their prompt using a specific syntax, although this does not work with wildcards.

  • What are the new scheduler options available in Fooocus?

    -The new scheduler options can be found in the debug menu under the scheduler dropdown, though the technical details are not discussed in the video.

  • What minor change has been made to some of the image input areas in Fooocus?

    -The minor change is the removal of the label to drag the image to the input area, as it was considered redundant.

  • What new feature has been added to the inpainting area in Fooocus?

    -The new feature allows users to change the color of the inpainting brush for masking images, as well as mix multiple colors on one mask.

  • What does the ability to bypass clip layers in Fooocus entail?

    -The ability to bypass clip layers allows users to adjust the number of layers used in the clip model through a slider in the debug menu, which can be useful for customizing image generation.

  • What is the new Hyper SD option in Fooocus and how does it work?

    -The new Hyper SD option uses the HyperSD four-step lure behind the scenes, allowing users to use any SDXL model with the HyperSD mode for faster image generation on lower-end systems.

Outlines

00:00

🔄 Fooocus 2.4.0 Update Overview

Rodney from Kleebz Tech introduces the latest Fooocus update, version 2.4.0, which includes quality of life improvements and backend changes. The update is automatic and offers new features such as support for downloading Hugging Face files from a mirror site, an update to Interposer from version 3.1 to 4.0 to facilitate interoperability between different Stable Diffusion models, and a new reconnect button to recover from crashes or accidental closures. Additionally, users can now select a custom VAE for decoding from the latent space, and there's a new style section option to select a random style for image generation. The default anime model has been updated, and the log file now shows full raw prompts. An optional feature to automatically describe an image when dropped into the variation and upscale section requires editing the bat file.

05:00

🎨 Enhanced Features and User Interface Improvements in Fooocus

This paragraph discusses additional updates in Fooocus 2.4.0, including the ability to see the full raw prompts in the log file, an NSFW filter to prevent inappropriate content, the inclusion of LoRAs in prompts, new scheduler options, and minor UI changes such as the removal of redundant labels. A significant feature is the ability to change the color of the inpainting brush for better mask visibility. Users can now adjust the clip layers via a slider in the debug menu. Lastly, the new Hyper SD option allows the use of any SDXL model with the HyperSD mode for faster image generation on lower-end systems. The video concludes with Rodney thanking Mashb1t for his work on Fooocus, mentioning his Patreon and other support options, and considering live streaming for interactive content creation.

Mindmap

Keywords

💡Fooocus

Fooocus is a software application designed for image generation, particularly with the use of AI models like Stable Diffusion. In the context of the video, Fooocus is being updated to version 2.4.0, which introduces several new features and improvements aimed at enhancing user experience and functionality. The script discusses various updates to Fooocus, such as support for downloading models from a mirror site and the introduction of a reconnect button.

💡Hugging Face

Hugging Face is mentioned as a source for downloading model files, with the update allowing for downloading from a mirror site. Hugging Face is a company that provides a platform for developers to share and collaborate on machine learning models, including those used in natural language processing and computer vision.

💡Interposer

The term 'Interposer' refers to a software component that acts as an intermediary between two other components. In the video, an update to Interposer from version 3.1 to 4.0 is highlighted. This update is crucial for enabling compatibility between different AI models, specifically SD 1.5 and SDXL Stable Diffusion models, which have different architectures and latent spaces.

💡VAE

VAE stands for Variational Autoencoder, a type of neural network that is used for encoding and decoding data, in this case, images. The script mentions the ability to select a custom VAE in the updated Fooocus, which is significant for users who want more control over the image generation process, especially when upscaling images to avoid faded-looking results.

💡Reconnect Button

The 'Reconnect Button' is a new feature in Fooocus that allows users to reconnect to their previous session if the application crashes or is accidentally closed. This is beneficial as it preserves the user's settings and progress, as illustrated in the script where it replaces the generate button upon relaunch.

💡Random Style

The 'Random Style' feature, as discussed in the script, allows Fooocus to select a random style in addition to any styles the user has chosen. This can result in a diverse set of images being generated, with each image having a unique style. It adds an element of unpredictability and creativity to the image generation process.

💡Anime Model

The 'Anime Model' refers to a specific model within Fooocus used for generating anime-style images. The script mentions an update to this model from version 1 to version 3.1, indicating an improvement in the model's capabilities or the quality of the images it produces.

💡Log File

A 'Log File' in the context of Fooocus is a record of the actions performed during an image generation session. The update allows for the log file to display the full raw prompts, which was not previously possible. This feature enhances transparency and allows users to better understand the process behind their image generation.

💡NSFW Filter

The 'NSFW Filter' is a feature that prevents the generation of Not Safe For Work content. The script explains that this filter can be enabled in the debug section of Fooocus, and if enabled, it will not display a preview of generating images and will save a black image if NSFW content is detected. This is important for users who want to avoid explicit content.

💡LoRAs

LoRAs, or Low-Rank Adaptations, are a technique used in machine learning to adapt a pre-trained model to new tasks with fewer resources. In the script, it's mentioned that users can now include LoRAs in their prompts, which suggests that Fooocus is providing more flexibility in how users can utilize AI models for image generation.

💡Inpainting Brush

The 'Inpainting Brush' is a tool used in image editing to fill in or mask parts of an image. The script highlights a new feature that allows users to change the color of the inpainting brush, which can improve visibility and precision when working on images with similar color backgrounds.

💡Bypass Clip Layers

In the context of the video, 'Bypass Clip Layers' refers to a feature that allows users to skip certain layers in the CLIP model during image generation. This can affect the style and detail of the generated images. The script explains that users can adjust a slider in the debug menu to control this feature, which can be useful for achieving specific visual effects.

💡Hyper SD

Hyper SD is a mode in Fooocus that utilizes a four-step process for image generation, allowing the use of any SDXL model. As mentioned in the script, this mode is beneficial for users with lower-end systems as it offers faster image generation. However, the presenter notes a preference for higher quality images generated by more steps.

Highlights

Fooocus Update 2.4.0 introduces quality of life improvements and new backend features.

Support for downloading Hugging Face files from a mirror site added.

Interposer updated from version 3.1 to 4.0 for better model interoperability.

New reconnect button to recover settings after Fooocus crashes or closes.

Ability to select a custom VAE for improved image decoding.

Option to select a random style for image generation.

Default anime model updated from Anime Pencil XL version 1 to 3.1.

Log file now shows full raw prompts for better transparency.

Feature to automatically describe an image when dropped into variation and upscale section.

NSFW filter added to prevent inappropriate content.

Ability to include LoRAs in prompts for more control over image generation.

New schedulers available for different image generation needs.

Image input areas have had redundant labels removed for a cleaner interface.

Inpainting brush color can now be changed for better masking.

Option to bypass clip layers for more control over image generation.

New Hyper SD option for faster image generation on lower-end systems.

Mashb1t has been managing the Fooocus project and contributed to the updates.

The creator is considering live streaming to demonstrate Fooocus features.