so Looking Glass lets you "finetune" the training of ruDalle's general purpose model toward the style of your inputs. Train it on 100 Impressionist landscapes, and it will produce images in that style. Give it one portrait photo of Joe Biden, and it will produce many nearly identical slightly melty Joe Bidens.
I've finetuned it to produce basically Corporate Memphis style abstracts before. Could probably do it better now. But using a finetuned model to produce stylized outputs with specific subject matter derived from photo or text inputs (like another photo from iraq, or just "abu ghraib illustration") is another step of challenge I haven't tried yet
there is a Google Colab notebook called "Looking Glass" that lets you load the ruDALL-E XL model "Malevich" and train it with image inputs.
a Colab notebook lets you run code on Google gpus via your browser.
ru-dalle is a Russian project comparable to Openai's dalle. :putin-wink: (the thing in OP uses the lightweight version of dalle, dalle mini)
so Looking Glass lets you "finetune" the training of ruDalle's general purpose model toward the style of your inputs. Train it on 100 Impressionist landscapes, and it will produce images in that style. Give it one portrait photo of Joe Biden, and it will produce many nearly identical slightly melty Joe Bidens.
I've finetuned it to produce basically Corporate Memphis style abstracts before. Could probably do it better now. But using a finetuned model to produce stylized outputs with specific subject matter derived from photo or text inputs (like another photo from iraq, or just "abu ghraib illustration") is another step of challenge I haven't tried yet