Researching #ComputerVision at #GoogleDeepMind using JAX/Flax (http://github.com/google/flax). Views are my own.
Andreas Steiner
Looking for a small or medium sized VLM? PaliGemma 2 spans more than 150x of compute!
Not sure yet if you want to invest the time 🪄finetuning🪄 on your data? Give it a try with our ready-to-use "mix" checkpoints:
🤗 huggingface.co/blog/paligem...
🎤 developers.googleblog.com/en/introduci...
Attending #NeurIPS2024? If you're interested in multimodal systems, building inclusive & culturally aware models, and how fractals relate to LLMs, we've 3 posters for you. I look forward to presenting them on behalf of our GDM team @ Zurich & collaborators. Details below (1/4)
🚀🚀PaliGemma 2 is our updated and improved PaliGemma release using the Gemma 2 models and providing new pre-trained checkpoints for the full cross product of {224px,448px,896px} resolutions and {3B,10B,28B} model sizes.
1/7
Want to get started using PaliGemma 2?
🎤 developers.googleblog.com/en/introduci...
🤗 huggingface.co/blog/paligem...
💾 kaggle.com/models/googl...
🔧 github.com/google-resea...
7/7
In addition to the pre-trained checkpoints, we also release two checkpoints fine-tuned on the DOCCI dataset, which generate fine-grained captions with a great quality/compute trade-off – and no yapping!
5/7
After 🪄finetuning🪄 on your data, you can expect to see great results, like the sota we got on recognizing table structures, music scores, molecular structures, and text, and on radiography report generation.
4/7
If you want to know more, now is a good time to head over to the 31 page tech report.
Brought to you by an amazing team of collaborators from
@GoogleDeepMind
and
@GoogleAI
.
arxiv.org/abs/2412.03555
6/7
As the original PaliGemma, the pre-trained PaliGemma 2 models have segmentation and detection capabilities, and excel at OCR – which makes them extremely versatile for 🪄finetuning🪄. The original demo hf.co/spaces/big-v... gives you an idea of the capabilities.
3/7
Adding this new "model size" dimension unlocks substantial improvements for some tasks (blue, e.g. AI2D), and compounds with improvements from increased resolution for most tasks (green, e.g. InfoVQA).
2/7
