|
--- |
|
license: gpl |
|
--- |
|
|
|
# STORYTIME -- **teaching Stable Diffusion to use its words** |
|
|
|
Greetings, Internet. The purpose of the Storytime model is to create a Stable Diffusion that "communicates" with text and images. My focus currently is on the English language, but I believe any of the techniques that I apply are applicable to other written languages as well. Inasmuch as "language" encompasses a huge range of notions and ideas (many of which I am probably ignorant of as a hobbyist), the purpose specifically is to use undifferentiated image data to "teach" Stable Diffusion about various language concepts, then see if it can adopt them and apply them to images. |
|
|
|
To give you a sense of what I mean, this model is Stable Diffusion v1.5 trained on a small dataset of alphabet flashcards, simple word lists, grammatical concepts presented visually via charts, images of pages of text, and images of text along with picutres, all in English. However, the letters, words, sentences, paragraphs, and other mechanics are not specifically identified in the captions. A flashcard showing the letter "A", for instance, will have a generic caption such as "A picture of a letter from the English alphabet". But the letter itself is not identified. An image of a list of common sight-words will have a caption similar to "A picture showing common words in the English language written using the English alphabet". And so on. At some point in time I will publish the data I use for training, but I want to be sure about sourcing and attribution and an issues around those (and rectify any that may be problematic) before publishing. |
|
|
|
In a similar manner lists of words are presented with the concept/class "word" included in the caption, but the words themselves are not spelled out. Can Stable Diffusion learn to put together legible letters, words, and sentences simply by "learning" from data presented in this manner? This is what I aim to understand. |
|
|
|
The sample images here include those that contain legible words, including words (usually) from the image generation prompt. I suspect these are words that are "abundant", visually, in the dataset, but at this point I cannot and will not draw any conclusions about whether or not, for instance, the Stable Diffusion model has learned to identify the word "tree" with its visual representation, both in picture form (a picture of a tree) and "written" form (a picture of the word "tree"). I have found by interrogating CLIP/BLIP to generate captions for some of these images that it picks up on the words written in the picture, even nonsense words and/or misspellings and those with letters whose form isn't well-defined. Several examples of such captions are shown below. |
|
|
|
v1 of Storytime was trained using 130 512x512 images of alphabet letters, flashcards, word charts, pages of text, and pages of text with images on a computer running Windows 10 with a single Nvidia RTX a4000 GPU for 300 epochs with a batch size of 8. No prior preservations images were used. |
|
|
|
For best results (most legible text in the image), use the DPM++ 2M with a CFG around 7 and 90 sample steps at a resolution of 640w x 768h. |
