DALLE-2 Statistics - Key Statistics 2023

How does DALLE- 2 work?

DALLE- 2 uses an encoder-decoder model, where textual input is first encoded into a machine-readable format and then processed by the model to generate an image. The resulting image is then decoded and made visible to the user. This process allows DALLE- 2 to create accurate and detailed images from textual descriptions, making it a powerful tool for various industries.

➤ Take a look at the image below

Source : GitHub

➤ Here’s a more detailed description :

DALLE- 2 uses a two-stage process to generate images. The first stage involves generating the "gist" of an image by using a model called the ‘Prior’ to create a CLIP image embedding from the given caption. The CLIP image embedding describes the essential features of the image. In the second stage, a diffusion model called unCLIP generates the actual image by filling in the remaining details based on the embedding generated in the first stage. During training, unCLIP receives a corrupted version of the image it is meant to reconstruct, as well as the CLIP image embedding of the clean image. The name unCLIP comes from its ability to reverse the mapping learned by the CLIP image encoder. Since unCLIP is trained to "fill in the gaps" necessary to produce a realistic image from the embedding, it learns to model all the information that CLIP deems irrelevant for its training objective and discards. This innovative approach enables DALL-E 2 to create highly accurate and detailed images from textual descriptions.

➤ In addition, DALLE-2 also has other capabilities like:

• Inpainting (Figure 1) : You can perform edits to an image using language;
  

  Source : Inpainting (Figure 1)

  
  
• Variations (Figure 1.1) : Generate novel images that capture the essence of a given reference image, while varying the arrangement of details, and also enhance the resolution.
  

  Source : Variations (Figure 1.1)

  
  
• Text diffs (Figure 1.1) : transform any aspect of an image using language.

How much does DALLE-2 cost?

Previously, DALL-E was available for use by those granted access by OpenAI, with a limit of 50 free image generations per day. Users were able to create a range of art using this text-to-image AI, such as restaurant mood boards, music videos, 3D renderings, product designs, and magazine covers.

However, on July 20th, 2022, OpenAI introduced the beta launch of DALL-E 2, along with a new pricing model. Users receive 50 free credits for the first month of use and 15 free credits each month thereafter. Each credit allows users to submit a text prompt to the AI, which generates four images at a time (three images for edits and variations).

Users can purchase additional credits to access more image generations, with a pack of 115 credits available for $15. If users exhaust their credits, they can purchase another pack for an additional $15. To put it in perspective, 115 credits can produce up to 460 images.

What do the users think of the pricing?

After OpenAI announced the pricing model for DALL-E 2, people have expressed a range of opinions. While some users are dissatisfied, others have shown their support. Let's take a look at some of the opinions and concerns raised:

• Some users believe that, compared to the prices charged by professional artists for their time and expertise, the cost of $15 for 115 prompts is not expensive at all.
• Users are required to pay for a minimum of 5,750 credits (worth $750) at once, which means they cannot start by purchasing just 115 credits.
• Since it often takes several attempts to get the desired image, some users are concerned about the credits running out before the perfect image is generated.
• Some users are unhappy with the fact that the paid credits have an expiration date.