Comparison with other Diffusion Models

As the results of our experiments show, our diffusion model has the following features:

Consistently high-quality images are generated

This includes factors such as resolution, realism, and visual fidelity. High-quality images are essential for ensuring that the diffusion model is capable of generating realistic and visually appealing content. We have invited many industry clients （ie. Art Center College of Design) to test and evaluate the quality of our image generation. We have the same quality as market leaders like Dalle2, MJ, and Stable Diffusion. It is better than other models depending on how it is used and what criteria are used to compare it. Please see the images generated using the exact same prompts on the left for a comparison. (Left: AIGC Chain; Right: Stability AI).

High relevance of the generated images to the prompt

This involves evaluating how well the model is able to interpret and respond to the input prompt, and how well the generated images align with the intended subject matter or theme of the prompt. The ability to generate relevant and faithful images from the input prompt is largely dependent on the design of the CLIP model, which is used to convert the input text or image into an embedding. The CLIP model is trained on a large corpus of data and learns to map words and images to a common latent space. This allows it to effectively capture the semantic relationships between words and images. Our model is equipped with native English and Chinese CLIPs, which allow it to generate high-quality content that accurately reflects the input prompt in both languages.

Faster model training speed

To train and fine tune the same image set, a market comparable model needs 1 hour with 4 V100 GPUs, while AIGC Chain only needs 0.6 hour with 1 V100 GPU.

Faster content generating speed

Our diffusion model has significantly faster generating speed compared to existing methods. This allows us to generate high-resolution images in a shorter amount of time, making it a more practical and efficient tool for image generation. Using V100 GPU, to generate the same resolution image, our generating time is between 10% and 50% of market comparable models.

Consistent artistic style while allowing for creative variation

This means that the artwork produced within a style will have a cohesive and recognizable aesthetic, while leaving room for individual creativity and variation. For example, a consistent artistic style could include specific colour palettes, composition techniques, or subject matter, while leaving allowance for individual expression and variation within these constraints.

The use of this technology is essential for the creation of non-fungible tokens (NFTs), as it provides a cohesive and recognizable body of work that can be easily identified and valued by collectors. This approach allows for artistic freedom and creativity while ensuring that the artwork is desirable to collectors. For example, in the below NFT development, AIGC Chain was used to train a small model for the project. Once the training was done, the small model became the virtual artist, which could be used to make a set of NFT images for the project in a style that was easy to recognize and consistent.

With AIGC Chain models, users can quickly and easily make a series of high-quality images that all have the same artistic style. This allows for the creation of an engaging visual narrative with well-defined characters and settings for a picture book. In the demo below, AIGC Chain made a series of images automatically, with each image matching the text story.

Flexibility, allowing it to be adapted for different vertical industries and IPs

For example, our framework can be customised for the fashion industry, allowing merchandisers and designers to quickly and efficiently generate a wide range of high-quality images for clothing and accessories to be used in prototypes for manufacturing. Our model also has the ability to capture market intelligence and learn from past data and trends. This allows it to generate designs that are tailored to the specific market and audience of the fashion brand, ensuring that the generated designs are relevant and appealing to customers.