Model was trained by Sber AI
text2image generation1.3 B120 million text-image pairs & 2749 text-emoji pairs😋 Emojich is a 1.3 billion params model from the family GPT3-like, it generates emoji-style images with the brain of ◾ Malevich.
The main goal of fine-tuning is trying to keep the generalization of ruDALL-E Malevich (XL) model on text to emoji tasks. ruDALL-E Malevich is a multi-modality big pretrained transformer, that uses images and texts. The idea with freezing feedforward and self-attention layers in pretrained transformer is demonstrated high performance in changing different modalities. Also, the model has a good chance for over-fitting text modality and lost generalization. To deal with this problem is increased coefficient 10^3 in weighted cross-entropy loss for image codebooks part.
Full version of training code is available on Kaggle: 
All examples are generated automatically (without manual cherry-picking) with hyper-parameters: seed 42, batch size 16, top-k 2048, top-p 0.995, temperature 1.0, GPU A100. For making better generative emojis should use more attempts (~512) and select the best one manually.
Remember, the great art makers became "great" after creating just only one masterpiece.
Model was trained by Sber AI
text2image generation1.3 B120 million text-image pairs & 2749 text-emoji pairs😋 Emojich is a 1.3 billion params model from the family GPT3-like, it generates emoji-style images with the brain of ◾ Malevich.
The main goal of fine-tuning is trying to keep the generalization of ruDALL-E Malevich (XL) model on text to emoji tasks. ruDALL-E Malevich is a multi-modality big pretrained transformer, that uses images and texts. The idea with freezing feedforward and self-attention layers in pretrained transformer is demonstrated high performance in changing different modalities. Also, the model has a good chance for over-fitting text modality and lost generalization. To deal with this problem is increased coefficient 10^3 in weighted cross-entropy loss for image codebooks part.
Full version of training code is available on Kaggle:
All examples are generated automatically (without manual cherry-picking) with hyper-parameters: seed 42, batch size 16, top-k 2048, top-p 0.995, temperature 1.0, GPU A100. For making better generative emojis should use more attempts (~512) and select the best one manually.
Remember, the great art makers became "great" after creating just only one masterpiece.