Also an attempt at generating fakes Van Gogh's The Starry Night (for better or for worse) ;)

Once again OCR misinterpretations ahead...

Very interesting results, for example for Raphael's The School of Athens' description - especially when compared with the results obtained with Florence-2-base and Qwen2-VL-2B.

Just posted on @huggingface.bsky.social : a VLM Visual Arts analysis with DeepSeek Janus-1.3B.

Very interesting results once again especially when compared with those obtained with Florence-2 and Qwen2-VL

Sequel to OCR processing and text in images analysis: this time with new recently released model DeepSeek Janus-1.3B (still on @huggingface.bsky.social of course)

Very interesting results both for HTR and printed content.

Spoiler alert: Qwen2-VL-2B reveals itself as a Jane Austen fan…

Once more with Florence-2-base and Qwen-VL-2B.

VLM analysis part 2, as always on @huggingface.bsky.social : this time it’s based on textual content in images whether handwritten, typed/printed, or even from an illuminated medieval manuscript.

Comparative analysis of the results obtained when analyzing a large number of different artworks (from the Medieval era to Abstract Contemporary art).

It’s clear the models are more in favor of figurative art than Kandinskys - when detecting objects especially.

First of my VLM study series (posted on @huggingface.bsky.social) :

An in-depth analysis of artworks with Microsoft Florence-2-base and Alibaba Cloud Qwen2-VL-2B