NEW Stable diffusion 2.1 RELEASED!

Sebastian Kamph
7 Dec 202210:15

TLDRStable Fusion 2.1 has been released as an improvement over the poorly received 2.0 version, addressing issues such as restrictive data filtering and image quality, particularly for images of people. The update introduces a more diverse data set, better handling of prompts, and support for a wider range of aspect ratios. Users have adapted to the new model by using negative prompts, and the developers have acknowledged these improvements in version 2.1, promising better results for architecture, landscapes, and character images. The release aims to combine the best of both worlds, delivering high-quality images across various styles and subjects.


  • πŸš€ Stable Fusion 2.1 has been released as an improvement over the poorly received 2.0 version.
  • πŸ’‘ The 2.0 version was criticized for its significant changes in model functionality, leading to unsatisfactory results for most users.
  • 🌟 Users found ways to improve results with 2.0 by using negative prompts and learning the model's ins and outs.
  • 🎨 Stable Fusion 2.1 supports a new prompting style and brings back many prompts that were previously effective.
  • πŸ“ˆ The new version includes more data, more training, and less restrictive filtering of the data set.
  • πŸ–ΌοΈ There was a focus on improving the diversity and range of the data set, particularly in architecture, interior design, wildlife, and landscape scenes.
  • πŸ‘€ Version 2.1 aims to fix issues with generating images of people by reducing the restrictive filtering that previously cut down on the number of people in the data set.
  • πŸ™οΈ The model now promises better rendering of architectural concepts, natural scenery, and fantastic images of people and pop culture.
  • πŸ“Š The release delivers improved anatomy and hands, and is better at a range of art styles compared to Stable Fusion 2.0.
  • 🌐 Stable Fusion 2.1 is an open-source release available on Hugging Face for those interested in exploring and using the model.

Q & A

  • What is the main issue addressed in the release of Stable Fusion 2.1?

    -The main issue addressed in Stable Fusion 2.1 is the improvement over the previous version, 2.0, which had a total Fiasco of a release due to the way the model worked, resulting in most users getting terrible results.

  • How did users adapt to the changes in Stable Fusion 2.0?

    -Users adapted to the changes in Stable Fusion 2.0 by learning new ways to prompt the model, using negative prompts and other techniques to achieve better results.

  • What were some of the improvements made in Stable Fusion 2.1 based on user feedback?

    -Stable Fusion 2.1 brought back support for the new prompting style, eased up on the restrictive filtering of the data set, improved anatomy and hands, and became better at a range of art styles compared to version 2.0.

  • What was the impact of the data set filtering on the image quality in Stable Fusion 2.0?

    -The data set filtering in Stable Fusion 2.0 resulted in a big jump in image quality for architecture, interior design, wildlife, and landscape scenes but dramatically cut down the number of people in the data set, making it harder to generate images of people.

  • How did the developers address the issue of generating images of people in Stable Fusion 2.1?

    -The developers addressed the issue by working hard to give the model a more diverse and wide-ranging data set, easing up on the restrictive filtering, and fine-tuning the model to capture the best of both worlds, allowing it to render beautiful architectural concepts and natural scenery with ease, as well as produce fantastic images of people and pop culture.

  • What new features were introduced in Stable Fusion 2.1 regarding image resolution and aspect ratio?

    -Stable Fusion 2.1 introduced the ability to render non-standard resolutions, which helps in creating extreme aspect ratios for beautiful vistas and epic widescreen imaging.

  • How do different tools handle negative prompts?

    -Different tools handle negative prompts in various ways. For example, Dream Studio uses a vertical bar or pipe, Automatic 11 11 uses a special box for negative prompts, and Invoke uses brackets for negative prompts.

  • Where can users find the weights and checkpoints for Stable Fusion models?

    -Users can find the weights and checkpoints for Stable Fusion models on Hugging Face, an open-source platform.

  • What is the significance of the YAML file for Automatic 11 11?

    -The YAML file is needed to use Automatic 11 11 with Stable Fusion models, but it was not found by the speaker during the discussion.

  • What is the purpose of the negative prompt in the context of Stable Fusion?

    -The negative prompt is used to reinforce the visual fidelity and style of the generated images, blocking certain elements that are not desired in the final output.

  • How can users try out Stable Fusion 2.1?

    -Users can try out Stable Fusion 2.1 by downloading the models and testing them in different UIs or by using Dream Studio, which is a Stability AI platform available at



πŸš€ Introduction to Stable Fusion 2.1 and Its Improvements

This paragraph introduces the release of Stable Fusion 2.1, reflecting on the shortcomings of the previous 2.0 version. It highlights the improvements made in the new version, such as better support for the new prompting style, reinstatement of various prompts, and a more diverse data set. The speaker expresses optimism about the changes, particularly in generating better images of people and easing the restrictive filtering that was an issue in version 2.0.


🌟 Enhanced Features and User Feedback in Stable Fusion 2.1

The second paragraph delves into the specific features that have been enhanced in Stable Fusion 2.1, such as improved anatomy, better handling of art styles, and the ability to render non-standard resolutions. It also discusses user feedback and how the developers have listened and adjusted the filters to produce better results while still excluding adult content. The paragraph emphasizes the model's versatility in creating architectural concepts, natural scenery, and images of people and pop culture.


πŸ“’ Conclusion and Encouragement to Try Stable Fusion 2.1

In the final paragraph, the speaker concludes the discussion on Stable Fusion 2.1 by encouraging viewers to try out the new version and share their experiences. The speaker acknowledges that some may still prefer older versions like 1.4 or 1.5 due to their flexibility, but suggests that the improvements in 2.1 could be worth exploring. The paragraph ends with a prompt for feedback and a sign-off until the next video.



πŸ’‘stablefusion version 2.1

This refers to the updated version of a software or model, which is an improvement over the previous version 2.0. The term suggests that the developers have learned from the shortcomings of the earlier release and have made necessary adjustments to enhance its performance and user experience. In the context of the video, stablefusion 2.1 is presented as a solution to the issues faced by users with version 2.0.

πŸ’‘negative prompts

Negative prompts are a technique used in AI models where the user provides instructions on what not to include in the generated output. This helps guide the AI to produce more accurate and desired results by avoiding certain errors or unwanted elements. In the video, the use of negative prompts is mentioned as a workaround to improve the outcomes with the previous version of stablefusion.

πŸ’‘data set filtering

Data set filtering refers to the process of selecting and refining the data used to train an AI model. This ensures that the model is trained on high-quality, relevant data and avoids including inappropriate or irrelevant content. In the context of the video, the developers have adjusted the filtering to address user concerns about the restrictive nature of the previous data set, which led to the removal of many images.

πŸ’‘image quality

Image quality is a measure of how well an image looks in terms of its resolution, clarity, and the accuracy of its representation. High image quality is important for users who want to utilize AI models for generating visual content. In the video, the improvement in image quality is highlighted as one of the key advancements in stablefusion 2.1, particularly in areas like architecture, interior design, wildlife, and landscape scenes.

πŸ’‘anatomy and hands

Anatomy and hands refer to the accurate and realistic representation of human anatomy and hands in generated images. This is a challenging aspect for AI models, as it requires precise details and proportions. In the context of the video, the developers of stablefusion 2.1 claim to have made improvements in this area, which was a problem in the previous version where the generation of people often resulted in poor quality images.

πŸ’‘art styles

Art styles encompass the various visual languages and techniques used in creating artwork. These can range from realistic to abstract and from traditional to modern. The ability of an AI model to generate images in a wide range of art styles is a testament to its versatility and creativity. In the video, the new version of stablefusion is praised for being better at producing images in a variety of art styles compared to the previous version.

πŸ’‘non-standard resolution

Non-standard resolution refers to image dimensions that do not conform to common or widely used sizes. This can include both extremely high resolutions, like 4K or 8K, and unique aspect ratios that create panoramic or vertical images. The ability to render non-standard resolutions opens up new possibilities for creative expression and technical applications.

πŸ’‘open source release

An open source release means that the source code of the software is made available to the public, allowing anyone to view, use, modify, and distribute the software freely. This encourages collaboration, innovation, and community involvement in the development and improvement of the software. In the context of the video, the open source nature of stablefusion allows users and developers to access and contribute to the model.

πŸ’‘Dream Studio

Dream Studio is likely a platform or tool associated with the stablefusion model, designed to facilitate the use of the AI for creating images. It may provide a user-friendly interface and additional features to enhance the image generation process. In the video, Dream Studio is presented as a way for users to interact with the stablefusion model.

πŸ’‘yaml file

A YAML (YAML Ain't Markup Language) file is a human-readable data serialization format often used for configuration files, data exchange, and as a markup language for cloud services. It is known for its simplicity and readability, making it a popular choice for defining parameters and structures in various applications. In the context of the video, a YAML file is mentioned as a requirement for using the stablefusion model with a specific tool.


Stable Fusion 2.1 release announced following the problematic launch of version 2.0.

Improvements in 2.1 aim to address user feedback and enhance model performance, especially in image quality.

Version 2.0 faced criticism for its drastic changes in model behavior leading to poor user results.

2.1 supports new and old prompting styles, reintegrating familiar features for users.

The update includes more data, extended training, and less restrictive data filtering.

Adjustments made to NSFW content filtering to maintain image quality without compromising safety.

Enhanced diversity in dataset aims to improve person image generation, addressing a major issue in version 2.0.

Negative prompts remain necessary for fine-tuning output, despite criticisms.

Architecture, landscapes, and environmental imagery significantly improved in 2.1.

Support for a wider range of aspect ratios introduced, catering to diverse creative needs.

Less aggressive filtering leads to fewer false positives and better content generation.

Improved rendering of people and pop culture references in the latest release.

Version 2.1 delivers better anatomy depiction and supports a variety of art styles more effectively.

The new update facilitates better use of negative prompts and integrates seamlessly with various tools.

Stable Fusion 2.1 is an open-source release, available on Hugging Face with support for popular UIs like Automatic 1111.

Feedback on version 2.1 is encouraged, with a focus on community-driven improvements and continued innovation.