InvokeAI - SDXL Getting Started

Invoke
29 Jul 202306:40

TLDRInvokeAI 3.0.1 introduces support for SDXL in its linear UI, simplifying the process of inputting SDXL prompts. The update includes a positive and negative style prompt box, with a recommendation to use the concatenate basic prompt with style button for optimal results. The style prompt is crucial for coherence and should be concise and stylistic. For FP16 precision, a fixed VAE is advised to avoid image aberrations. The video recommends starting with a 1024x1024 image size and adjusting the refiner model settings carefully for best results. The presenter also suggests trying different schedulers and anticipates more SDXL fine-tunes becoming available.

Takeaways

  • 🔧 InvokeAI 3.0.1 supports SDXL with a user-friendly linear UI.
  • 📝 The SDXL model introduces new prompt fields: positive style prompt and negative style prompt.
  • 🔀 Using the concatenate button in the UI links the subject and style for better results.
  • ⚠️ Over-prompting in the negative style field can reduce image coherence in SDXL.
  • 🎨 Too much focus on the style prompt may result in loss of subject clarity.
  • 🛠️ For those using FP16 precision, a fixed VAE is recommended to avoid visual aberrations.
  • 📏 It’s crucial to maintain image resolutions similar to the ones SDXL was trained on, such as 1024x1024 pixels.
  • 🖼️ The refiner model can enhance image details but may require experimentation with settings.
  • ⏳ The refiner should be applied at around 75% completion of the base model's steps for best results.
  • 🚀 Euler and Euler Ancestral schedulers tend to work better with SDXL, compared to other schedulers.

Q & A

  • What is the significance of the new 3.0.1 update in InvokeAI?

    -The new 3.0.1 update in InvokeAI introduces support for SDXL in the linear UI, making it easier for users to input SDXL prompts and utilize the refiner directly from the UI.

  • What changes occur when switching to an SDXL model in InvokeAI?

    -When switching to an SDXL model, users will notice two additional prompt boxes: the positive style prompt and the negative style prompt.

  • What is the recommended approach for using the subject and style prompt boxes with SDXL?

    -It is highly recommended to use the 'concatenate basic prompt with style' button to link your subject and style prompts when passing them to the second encoder for optimal results.

  • Why is the style prompt considered important with the SDXL model?

    -The style prompt is crucial with the SDXL model as having too much content in it can lead to a loss of coherence, and it should be focused and stylistic in nature.

  • What should users be cautious about when using the negative style prompt?

    -Users should be cautious about not heavily prompting in the negative style prompt field, as the SDXL model tends not to perform well with too much content in the negative prompt.

  • Is there a specific VAE file recommended for use with SDXL when running in fp16 precision?

    -Yes, when using fp16 precision with SDXL, it is recommended to download a fixed VAE file that was released with the original SDXL base 1.0 model to avoid image aberrations.

  • What are the recommended dimensions for images when using SDXL?

    -For SDXL, it is recommended to start with a 1024 by 1024 image or a width and height that has roughly the same number of pixels to ensure coherence.

  • What is the role of the refiner model in SDXL, and how should it be used?

    -The refiner model in SDXL can add nice details to an image but can be finicky. It is recommended to get accustomed to it and play around with settings to achieve good results.

  • How does the number of steps selected in the UI affect the refiner in SDXL?

    -The number of steps selected in the UI will determine how many steps are run using the refiner. It is suggested to have the base model complete roughly 70-80% of the steps, leaving the finishing touches for the refiner.

  • What are some recommended schedulers to use with SDXL?

    -For SDXL, Euler or Euler ancestral schedulers are recommended as some DPM plus plus schedulers may result in weird-looking images.

  • Where can users find SDXL prompt styles for easy reference?

    -Users can find SDXL prompt styles on GitHub, where a contributor named keyturn has put together a reference. The link to this resource is included in the details below.

Outlines

00:00

🚀 SDXL 3.0.1 Update: New Prompting and VAE Support

The video introduces the 3.0.1 update for the software, which now includes SDXL support in the linear UI. It highlights changes to the prompting system, including two new prompt boxes for positive and negative style prompts. Users are encouraged to use the concatenate basic prompt with style button to combine subject and style effectively. A key focus is ensuring that style prompts aren’t overloaded, especially in the negative prompt field, as too much information may lead to incoherent results. There’s a reminder about using a fixed VAE file for FP16 precision to avoid image artifacts. Additionally, users are advised to maintain image sizes around 1024x1024 or similar pixel dimensions for optimal results.

05:03

📏 Image Dimensions and Refiner Model Tips for SDXL

This section covers more technical aspects of working with SDXL, particularly image dimensions and using the refiner model. It suggests starting with 1024x1024 dimensions but adjusting width and height proportionally if resizing. The refiner model, which adds final details to images, can be turned off if the base model delivers acceptable results. However, it requires experimentation to achieve good outcomes. The refiner functions like a denoising tool, and the number of steps should be set to around 75% of the base model steps. Tips include using Euler or Euler Ancestral schedulers for better results while refining, though there’s still room for personal adjustment.

🌟 Upcoming SDXL Fine-Tunes and Prompting Resources

The final section anticipates the release of several SDXL fine-tuned models. It mentions the new Dream Shaper XL model and how users can easily convert models using the model manager. The video also references a helpful SDXL prompt styles guide created by a contributor named Keyturn, available on GitHub. As a quick demo, the host showcases combining positive and negative prompts, adding stylistic terms. The video concludes by teasing more updates and improvements, encouraging viewers to stay connected for future releases.

Mindmap

Keywords

InvokeAI

InvokeAI is the name of the software discussed in the video script. It is a tool that has been updated to version 3.0.1, which includes support for SDXL. The software allows users to input prompts and refine images through a user interface (UI). In the context of the video, InvokeAI is central to the discussion as it is the platform on which users will be implementing SDXL features.

SDXL

SDXL stands for 'Stable Diffusion XL', which is a model that has been integrated into InvokeAI 3.0.1. It represents a shift in how users interact with the software, as it introduces new ways of prompting the AI. The video emphasizes the importance of using the style prompt effectively with SDXL to achieve coherent results.

Linear UI

Linear UI refers to the user interface of InvokeAI, which has been updated to be more user-friendly. It allows users to input SDXL prompts and use the refiner directly from the interface. The script mentions that the UI has been made 'super easy' to use, indicating that it is designed for intuitive interaction.

Positive Style Prompt

The Positive Style Prompt is one of the additional prompt boxes introduced with SDXL in InvokeAI. It is used to specify the stylistic elements that users want to see in the generated images. The video suggests that this prompt is crucial for achieving the desired output from the SDXL model.

Negative Style Prompt

The Negative Style Prompt is another new feature in InvokeAI that allows users to specify what they do not want to see in the generated images. The video advises users to keep this prompt focused and stylistic, as overloading it can lead to a loss of coherence.

Concatenate Basic Prompt with Style

This is a feature in InvokeAI that automatically concatenates the subject and style prompts when passing them to the second encoder. The video suggests that using this feature is highly recommended for users who are not familiar with how prompts should be structured for the best results with SDXL.

FP16 Precision

FP16 Precision refers to a type of numerical precision used in AI models. The video mentions that when using FP16 precision with SDXL, users should download a fixed VAE to avoid issues with image generation, such as the 'weird scan line look' mentioned in the script.

VAE

VAE stands for 'Variational Autoencoder'. In the context of the video, it is a component that needs to be downloaded and installed for InvokeAI to function correctly with SDXL, especially when using FP16 precision. The video suggests that a fixed version of the VAE should be used to prevent image aberrations.

Refiner

The Refiner is a part of the InvokeAI tool that can add details to images generated by the base model. The video explains that while the refiner can enhance images, it can also be 'finicky' and may require users to experiment with settings to achieve the best results.

Scheduler

In the context of the video, a Scheduler refers to an algorithm that controls the process of image refinement in InvokeAI. The video mentions that some schedulers, like DPM++, may produce unusual results, while Euler schedulers are recommended for use with SDXL.

Denoising Strength

Denoising Strength is a setting in InvokeAI that determines how much noise is left in an image after the base model has finished processing. The video suggests setting this to around 0.7 to 0.8 to allow the refiner to work on the image and add the final touches.

Highlights

InvokeAI 3.0.1 introduces SDXL support and makes it easy to use through their linear UI.

SDXL models require different prompting strategies, using both positive and negative style prompts.

InvokeAI's linear UI concatenates subject and style prompts for better results with SDXL.

Avoid overly complex negative prompts in SDXL, as too much negativity reduces image coherence.

For FP16 Precision, users must download a fixed VAE to avoid visual aberrations or watermark-like effects.

Recommended image sizes for SDXL models are roughly 1024x1024 pixels or pixel dimensions that maintain a similar total pixel count.

InvokeAI's refiner model can be toggled on and off; it adds extra detail but may be finicky with certain schedulers.

The refiner model's effectiveness depends on choosing the right number of steps, usually around 0.7 to 0.8 for denoising.

Using 75% of the base model’s steps for SDXL and refining the last 25% often gives the best results.

Schedulers like Euler or Euler Ancestral are recommended for SDXL models due to their stability compared to DPM++ schedulers.

More SDXL fine-tuned models, like Dream Shaper XL, are expected to be released in the coming weeks.

New SDXL models can be easily converted with the Model Manager tool in InvokeAI.

For beginners, SDXL Prompt Styles reference is available on GitHub to guide effective prompting.

InvokeAI offers integration with community resources, including Discord for more support and updates.

InvokeAI continues to evolve, with more features and improvements coming soon.