Other
Textcaps Teste2 - GIT Large for Text-Aware Image Captioning
GIT-large model fine-tuned on TextCaps for generating image captions that include text visible in scenes.
About this model
Textcaps Teste2 is a GIT (GenerativeImage2Text) large-sized model fine-tuned on the TextCaps dataset for image-to-text generation. GIT is a Transformer decoder conditioned on CLIP image tokens and text tokens, trained with teacher forcing on image-text pairs. The model uses bidirectional attention for image patches and causal attention for text tokens when predicting the next word.
This checkpoint was pre-trained on 20 million image-text pairs and further fine-tuned on TextCaps, a dataset specifically focused on captioning images that contain text in the scene. This makes it particularly suited for describing images where reading and referencing visible text matters — signs, documents, UIs, and similar content.
The model is based on Microsoft's GIT architecture from the paper [GIT: A Generative Image-to-text Transformer for Vision and Language](https://arxiv.org/abs/2205.14100). It can also be applied to visual question answering and image classification. Licensed under MIT.
Use via HuggingFace Transformers with the `image-to-text` pipeline. Images are resized, center-cropped, and normalized with ImageNet statistics during preprocessing.
Project signals
- 10 Hugging Face downloads
- 3 Hugging Face likes
Topics
transformers · pytorch · safetensors · git · image-text-to-text · vision · image-to-text · arxiv:2205.14100 · endpoints_compatible