10 Stable Diffusion Models Compared!

All Your Tech AI
1 Mar 202410:35

TLDRIn this video, the host explores 10 generative AI art models, testing their ability to follow prompts and produce aesthetically pleasing images. Models like Proteus V2, SSD 1B, and Juggernaut XL are evaluated based on their output's adherence to the prompt and visual quality. The results vary, with some models excelling in photorealism and others in anime or surreal styles. The audience is encouraged to vote on their favorite model and try them out on their own.


  • ๐ŸŽจ The video compares 10 different generative AI art models using the same prompt to assess their performance and aesthetic quality.
  • ๐Ÿ–Œ๏ธ The models tested include Proteus V2, SSD 1B, Playground V2, Stability AI's stable diffusion XL, Juggernaut XL, anime XL, Kandinsky 2.2, real viz XL, and dream shaper X XL turbo.
  • ๐Ÿ† Proteus V2 demonstrated impressive results with high-quality images, adherence to the prompt, and fast generation times.
  • ๐Ÿš€ SSD 1B, a fine-tuned stable diffusion XL model, was found to be faster but with reduced quality and missed the Ruby eyes detail.
  • ๐ŸŒˆ Playground V2, trained with mid-journey images, showed higher aesthetic quality but had some focus and saturation issues.
  • ๐Ÿ“ธ Stability AI's stable diffusion XL served as the baseline model, producing softer, less saturated images that some found aesthetically pleasing after upscaling.
  • ๐ŸŒŸ Juggernaut XL versions 8 and 9 aimed to improve upon the base model for higher aesthetic scores, with varying results in eye color adherence and overall image quality.
  • ๐Ÿ’Ž anime XL, fine-tuned for anime and cartoons, delivered high-quality results with the desired Ruby eyes and distinct anime style.
  • ๐Ÿ”ฎ Kandinsky 2.2 produced surreal and unique images with a dark aesthetic, but did not fully adhere to the prompt's eye color requirement.
  • ๐Ÿ‘๏ธ Real viz XL version 2 provided high-quality images with a slightly odd eye detail and a lack of Ruby eyes, indicating room for improvement in prompt adherence.
  • ๐Ÿš€ Dream shaper X XL turbo, despite its name, did not meet expectations with overly stylized and less realistic results, but performed better with non-human subjects like dragons.

Q & A

  • What was the main objective of the video?

    -The main objective of the video was to test and compare 10 different generative AI art models using the same prompt to see how each model interprets and generates the image, and to allow viewers to vote on their preferred model based on the results.

  • Which model was used as the baseline for comparison in the video?

    -Stability AI's Stable Diffusion XL was used as the baseline model for comparison in the video.

  • What were the two key aspects the video focused on when evaluating the AI-generated images?

    -The two key aspects the video focused on when evaluating the AI-generated images were the models' ability to follow the detailed instructions in the prompt and the aesthetic quality of the final image.

  • How did the Proteus V2 model perform in the test?

    -The Proteus V2 model performed well in the test, generating high-quality images that closely followed the prompt, including the specific detail of having ruby-colored eyes. It also generated images quickly.

  • What was unique about the Playground V2 model's training?

    -The Playground V2 model was uniquely trained and fine-tuned with 30,000 images from mid-journey, aiming to achieve a higher aesthetic quality score than the Stable Diffusion XL model.

  • How did the Juggernaut XL models differ from each other in the video?

    -The Juggernaut XL models differed in their iterations, with version 8 providing a more natural-looking image, while version 9, despite following the prompt closely with ruby eyes, had a creepy aesthetic and some abnormalities in the teeth and skin areas.

  • What was the main advantage of the Anime XL model in the context of the test?

    -The main advantage of the Anime XL model was its ability to generate images with an anime aesthetic, providing high-quality results with the desired ruby eyes and other specified features, making it suitable for projects requiring a cartoon or anime style.

  • What aesthetic characteristic did the Kandinsky 2.2 model produce?

    -The Kandinsky 2.2 model produced images with a surrealist aesthetic, characterized by a darker tone and highly stylized patterns, which, while not adhering fully to the prompt, offered a unique visual style.

  • How did the Real ViZ XL version 2 model perform in terms of prompt adherence and aesthetic quality?

    -The Real ViZ XL version 2 model did not adhere as closely to the prompt, particularly missing the ruby eyes detail, and its aesthetic quality was considered slightly odd with eyes that appeared unrealistic and a pattern in the freckles that was too symmetrical.

  • What was the general conclusion about the different AI art models?

    -The general conclusion was that different AI art models excel in producing certain types of images based on their specific training data sets. For instance, some models are better for photorealism while others, like Anime XL, are better suited for anime-style images. The choice of model depends on the specific prompt and art style desired by the user.

  • How did the presenter plan to engage viewers in the video's content?

    -The presenter planned to engage viewers by providing a webpage where they could view all the generated images, participate in a poll to vote for their favorite model, and leave comments to share their preferences and opinions.



๐ŸŽจ Testing 10 AI Art Models - Introduction and Methodology

The video begins with the host introducing an experiment to test 10 different generative AI art models, including popular ones like Stability AI's Stable Diffusion XL and lesser-known models fine-tuned for specific aesthetics or textual embeddings. The goal is to compare the models' ability to follow prompts and produce visually pleasing images. The host plans to use an identical prompt for each model and will display the results on a website for viewers to vote on their favorites. The models to be tested include Proteus V2, SSD 1B, Playground V2, Stability AI's SDXL, Juggernaut XL versions 8 and 9, Anime XL, Kandinsky 2.2, Real Viz XL version 2, and Dream Shaper XXL turbo. Links to these models are provided for those interested in trying them out. The first image tested is of a red-haired girl with specific features, and the host discusses the importance of both prompt adherence and aesthetic quality in evaluating the models.


๐Ÿ” Analysis of AI Art Model Results - Observations and Comparisons

In this segment, the host analyzes and compares the results from different AI art models. The discussion begins with the Proteus V2 model, which generates high-quality images that closely follow the prompt, including the challenging detail of Ruby-colored eyes. The SSD 1B model, a faster version of Stable Diffusion XL, produces lower quality images missing some prompt details. Playground V2, trained with mid-journey images, delivers higher aesthetic quality but has some focus and saturation issues. The Stability AI's SDXL baseline model produces softer, less saturated images that the host sometimes enhances with an image upscaler. Juggernaut XL versions 8 and 9 show improvements in sharpness and aesthetic quality but have issues with prompt adherence and overall visual appeal. The Anime XL model, trained on anime images, provides a good alternative for those seeking an anime aesthetic. Kandinsky 2.2 offers a unique, surreal aesthetic that doesn't fully adhere to the prompt. Real Viz XL version 2 produces high-quality images with some pattern oddities. Dream Shaper XXL turbo, while fast, results in overly stylized images that may not be as realistic. The host emphasizes that different models excel in producing certain types of images based on their training data sets.


๐Ÿ“Š Conclusion and Viewer Engagement - Final Thoughts and Call to Action

The video concludes with the host summarizing the AI art model experiment. The host expresses surprise at the standout performance of Proteus V2 and encourages viewers to check out the images on the website, participate in a poll, and share their opinions in the comments. The host also reminds viewers to download their favorite models or use them on Pixel Dojo. The message concludes with a poetic reminder of the host's role in the tech community and a commitment to providing valuable AI insights.



๐Ÿ’กGenerative AI Art Models

Generative AI Art Models refer to artificial intelligence systems designed to create visual art based on input data or prompts. In the context of the video, these models are used to generate images from detailed textual descriptions, showcasing the diversity and capabilities of different AI art generators. Examples include Proteus V2, SSD 1B, and Stability AI's stable diffusion XL.


Fine-tuning in machine learning, including AI art models, is the process of adjusting a pre-trained model to better perform a specific task or improve its performance on a particular dataset. In the video, models like Juggernaut XL and anime XL are mentioned as being fine-tuned for higher aesthetic quality or specific styles like anime and cartoons.

๐Ÿ’กTextual Embeddings

Textual embeddings are representations of text in a numerical, vector format that capture the semantic meaning of words or phrases. These embeddings are crucial in AI art models as they help the AI understand and follow the textual prompts more accurately, generating images that better match the described concepts.

๐Ÿ’กAesthetic Values

Aesthetic values refer to the collective appreciation of beauty or good taste as applied to art, which can be subjective and varies among individuals. In the context of AI art models, it relates to how well the generated images align with the perceived beauty or visual appeal to the viewer.

๐Ÿ’กPrompt Adherence

Prompt adherence is the degree to which an AI model follows the instructions or details provided in a textual prompt. In AI art generation, it is crucial for ensuring that the output matches the user's intended concept or description.

๐Ÿ’กHyper-Detailed Photography

Hyper-detailed photography refers to images with an extremely high level of detail, capturing intricate and fine aspects of the subject. In the context of AI-generated art, it represents the goal of creating highly realistic and detailed images that resemble high-quality photography.

๐Ÿ’กSampling Scheduler

A sampling scheduler is a mechanism used in generative models to determine the sampling process during the image generation. It affects the quality and characteristics of the final output by controlling aspects like the number of iterations or the level of detail.

๐Ÿ’กPixel Dojo AI

Pixel Dojo AI is a platform mentioned in the video that allows users to access and utilize various AI models without the need for having the computational resources to run them on their own computers. It provides an accessible way for users to experiment with different AI art models.

๐Ÿ’กAI Upscale

AI Upscale refers to the process of increasing the resolution of an image using artificial intelligence. This technique is used to enhance the quality and sharpness of images, especially when starting from lower resolution outputs typical from some AI models.

๐Ÿ’กAnime Style

Anime style is a term used to describe a specific visual art style originating from Japanese animated television shows and movies. It is characterized by colorful artwork, stylized characters with large eyes, and exaggerated expressions.


Surrealism is an artistic and literary movement that seeks to express the unconscious mind by combining unexpected, dreamlike, or fantastical elements. In the context of the video, it refers to the unique and sometimes unsettling aesthetic of certain AI-generated images that evoke a surreal quality.


Testing 10 different generative AI art models with identical prompts to compare their outputs.

The use of models beyond Stability AI's stable diffusion XL, which have been fine-tuned for specific aesthetic values or textual embeddings.

The experiment's goal is to evaluate how well each model follows detailed instructions and produces visually pleasing images.

Proteus V2's impressive performance in both following the prompt accurately and producing high-quality, fast results.

SSD 1B, a fine-tuned stable diffusion XL model with fewer parameters and faster generation, but with a drop in quality.

Playground V2's training with 30,000 images from mid-journey, aiming for higher aesthetic quality than stable diffusion XL.

Stability AI's stable diffusion XL as the baseline model that others are compared against, with its softer, default image style.

Juggernaut XL's iterations attempting to refine the base model for higher aesthetic scores and sharper images.

Juggernaut XL Version 9's peculiar output with Ruby eyes and an overall creepy aesthetic.

Anime XL's specialization in anime and cartoons, producing high-quality results with the desired Ruby eyes and freckles.

Kandinsky 2.2's unique aesthetic with a surrealist touch and overly precise patterns.

Real viz XL version 2's high-quality output with a slightly odd portrayal of the eyes and a non-adherence to the prompt's Ruby eyes.

Dream shaper X XL turbo's overly saturated and stylized output, and its potential for better results with non-human subjects.

The importance of matching the AI art model with the specific data set or art style one is aiming to create.

Proteus V2 emerging as a leader among the tested models for its overall performance and adherence to the prompt.

A call to action for viewers to check out the models, participate in a poll, and download their favorite models.