Florence 2 - The Best Small VLM Out There?

There is a new VLM on the scene and it comes with a dataset of 5Billion labels. The new model can do a variety of old world tasks like bounding boxes and segmentation along with newer LLM style captioning etc.
Paper: arxiv.org/pdf/2311.06242
HF Spaces Demo: huggingface.co/spaces/gokaygo...
Colab : drp.li/fGyMm
🕵️ Interested in building LLM Agents? Fill out the form below
Building LLM Agents Form: drp.li/dIMes
👨‍💻Github:
github.com/samwit/langchain-t... (updated)
github.com/samwit/llm-tutorials
⏱️Time Stamps:
00:00 Intro
00:13 Florence-2 Paper
02:19 Florence - 2 Architecture
03:20 Florence - 2 Detailed Image Captioning
03:41 Florence - 2 Visual Grounding
04:09 Florence - 2 Dense Region Caption
04:24 Florence - 2 Open Vocab Detection
06:01 Hugging Face Spaces Demo
10:41 Colab Florence - 2 Large Sample Usage

[@parkerspitzer]

Thanks for your work on sharing this information. Much easier to watch your content than keep my ear to the ground all day trying to keep up. Much appreciated, sir.

[@danielmz99]

Thanks for the great content. A video going through the fine-tuning process on this one would be amazing. I am not sure how this could scale to a video implementation (probably passing a frame each time).

[@coolmcdude]

I also would love a video/notebook for a Florence 2 fine tune

[@IsxaaqAcademy]

It's also good at OCR for hand written documents

[@jeremybristol4374]

I'm enthusiastic about these smaller models. Thanks for covering this!

[@mukkeshmckenzie7386]

Vqa tutorial would be nice!

[@IanScrivener]

Thanks Sam!!
Please keep up the great work...

[@micbab-vg2mu]

Thank you - it looks interesting:)

[@aa-xn5hc]

Great, yes, fine tune would be very interesting.

[@unclecode]

This is what people should call "small", anything below 1B! Thanks for your video. By the way, I played around with the quantized version, the result is unbelievably good! I shared a post on Twitter and mentioned you and shared the Colab. Take a look at it. I tried 8 bits and 4 bits. It's odd how 4 bits is almost the same as the base model!

[@samwitteveenai]

I saw you tweet and retweeted it, very cool stuff. I will check it out. just been knee deep in Gemma stuff for last few days

[@unclecode]

@@samwitteveenai Thanks, and yes, it's Gemma2's turn. Waiting for your CZcams notification about the Gemma video!

[@jefframpe5075]

Thanks, Sam! I always appreciate your videos.
I would love your take on how Florence-2 compare with Apple's 4M-21.

[@ALEXPREMIUMGAME]

awesome, thanks

[@GiovaniFerreiraS]

I'd love seeing a fine tuning video, specially if it's not question answering, just so it's a different use case from the documentation. Maybe with a quick intro talking about what are possible scenarios where fine tune would be specially helpful.

[@samwitteveenai]

Noted!

[@marcoscipioni132]

Yes, I'm trying to use it for table extraction out of scanned pdfs with little success so far. Would love to see how you implement that.

[@ranu9376]

I've tried this model, describing the image is great. I've also tried the docvqa, but giving only one word answers and not getting even simplest questions right. i had hoped to do some classification and compare with other models.

[@SaiManojPrakhya-mp4oe]

It would be great if you can show a finetuning example!

[@sohitshivhare1541]

Thanks for the information this is great.
Can i fine tune it for certain specific images like few short learning. Can you put a tutorial for the same it will be great full.

[@ShravanKumar147]

what would you pick for fine-tuning ?
Any specific application ideas?

[@ariramkilowan8051]

I think fine-tuning for OCR would be a good demo. OCR in the real world with images of documents is much harder than OCR on electronic documents so would be cool to see how a small model like this does as an alternative to Claude/GPT4.

[@MH-ke2wi]

I tried the OCR and OCR with region on images converted (no scanned) from PDF pages. Nothing fancy, standard text with some titles, sections, lists... it is absolutely unusable. When it detects something, it usually got it right, but it could only see around 25% of the text.

[@ariramkilowan8051]

@MH-ke2wi yeah also been struggling to get decent results with OCR

[@tonyrungeetech]

Hi Sam. Thank you for the videos. I've been playing around with some of the smaller vision models and trying to implement batched inferencing with little success. If you were trying to accomplish running multiple VQA style questions against the same image quickly, how would you go about that goal? Is batching even in the right direction I should be looking?

[@srk5702]

We request you to do fune tuning on object detection. Because, all llms are useful generating text oupit only. Thanks in advance

[@JustEmbraceTheChallenge]

Please do fine-tuning for Object detection

[@AbhishekKotecha]

Hi Sam, thanks for the video. What do you think about how does it compare with Phi3-V? My take is that this is more raw and better for fine tuning, do you also think so?

[@Walczyk]

this is completely better and more advanced than phi 3 v crap image detection

[@mshonle]

I wonder how much performance would be affected when something so distilled then gets quantized?
Also, it seems amazing that it can handle segmentation for an unspecified set size! With Phi3 Vision you would need to provide a token to represent, say, each giraffe you want to identify.

[@samwitteveenai]

quantization is a good question! I would expect it to suffer more than a big model. Might give it a test tomorrow.

[@SinanAkkoyun]

Where is the dataset? I couldn't find the release

[@toadlguy]

Would be interested on how much memory is required to run these models. they seem pretty small even unquantized. Maybe I will try it later on my 8GB M1 Mini. One thing I am curious about: at 3:38 , the description for the image is wrong in ways that seem odd. The title is described as being on top with the "20 Years of ..." underneath and Ron's tie is described as red and hair blonde. I wonder if this is just vagaries of the model (placement data would be strange) or over reliance on training data. Or a straight up mistake in 'creating' the paper (which would probably be the most disturbing😉).