madebyollin/megalith-10m

alfredplpl

Jul 3

I’m going to do captioning the images with Florence-2.
What do you plan?

madebyollin

Owner Jul 3

@alfredplpl That would be great! Honestly, I am hoping someone else will caption Megalith-10m so I don't have to 😅

Florence-2 captions looked okay in my initial test:

Some other captioners that might be worth trying:

SD3 used CogVLM
CommonCanvas used BLIP-2
AuraDiffusion seems to be using MoonDream2
PixArt started with LLaVA then switched to Share-Captioner

madebyollin

Owner Jul 4

I was curious and did a quick comparison of the different captioners here https://gist.github.com/madebyollin/ea2a80be58bd4e910a2cd73e20e5ee0c. Based on my initial impressions, I think Florence-2 is good but CogVLM and Moondream2 seemed a bit better (slightly more detail, slightly fewer hallucinations on the three images I tried)

animepfp

Jul 4

I can caption the entire dataset in about a week if you forward it to me with llava-next 8b, which I mostly use because it's fast. However, the dataset is missing metadata which is valuable for captioning ie the image title and descriptions. I can write a script in beautifulsoup to fetch all that, but it will take a bit. If you already have the images downloaded and could share them privately, that would help too.

In general florence2 does poorly with data integration, because it is a tiny model and lacks most context aside from identifying things with bounding boxes. CogVLM2 and InternVL are better but are slow. I can do multi-label classifier output for open-images and booru tags as well.

Let me know if you're interested! Communication here is fine.

alfredplpl

Jul 5

•

edited Jul 5

Thank you for your discussion.

But, I couldn’t download the images from Flickr because of the access limit.

Could you upload the images to Hugging Face?
If you do so, I can download the images and caption them.

madebyollin

Owner Jul 6

@animepfp Thanks for your interest! The image titles / tags / descriptions on Flickr seemed very challenging to use when I looked at them (mostly empty / useless, highly multilingual & non-standard formats, often irrelevant or misleading), but maybe there's some signal there. If you can fetch album names as well you might be able to get better results.

@alfredplpl Flickr will rate-limit you if you try to download too many images at once from one machine, but if you lower the number of parallel downloading threads I think the dataset should be able to download in a day or two.

I don't have the full-resolution dataset downloaded myself (apparently it's around 2TB), but I do have low-resolution (256x256) jpgs saved. Would that be enough for your captioning script?

animepfp

Jul 6

Unfortunately that is probably a little small for the llavanext architecture. I can try to download it independently. Thank you.

madebyollin

Owner Jul 6

@animepfp Sounds good! If you upload results anywhere (raw images, captions, or both), let me know and I can add a link to the README 🤝

alfredplpl

Jul 14

•

edited Jul 14

@madebyollin I also think that the 256x256 images is small for dense captioning. But the 256x256 images are useful for text-to-image training on first stage. I'll try caption the 256x256 images if you upload them.

It is hard for me to download the full-size images.
Someone also said:
https://x.com/IMG_5955/status/1812331380657070521

Thank you for your cooperation.

madebyollin

Owner Jul 14

@alfredplpl Regarding downloading performance, I did a quick test downloading the first 1m images with 4 threads at 512px resolution, and it seems to take < 1 day per million

img2dataset --url_list megalith-00001-of-00010.parquet --input_format "parquet" \
   --url_col "url_highres" --output_format files \
   --save_additional_columns '["url_source"]' \
   --output_folder megalith-10m-00001-of-00010 --processes_count 1 --thread_count 4 --image_size 512 \
   --resize_mode keep_ratio --resize_only_if_bigger true --min_image_size 512

I don't remember what the rate limit is, so you might be even able to increase this to 8 or 16 threads and download faster.

Regarding the tweet:

Megalith-10M images are resized to 1024px on the long edge, making them unsuitable for HD image generation model training. Full-size images are available via Flickr API, but strict limits make it extremely difficult to obtain 10M images this way...

This is true; b (1024px) is the largest size that uses a shared secret (https://www.flickr.com/services/api/misc.urls.html) so downloading higher resolutions probably requires more work (I haven't attempted it yet, but it might require having an API key at download time or something).

alfredplpl

Jul 16

Thank you for your comment.

I tried the script. Then, I got the next 1m images.

$ img2dataset --url_list megalith-00002-of-00010.parquet --input_format "parquet"    --url_col "url_highres" --output_format files    --save_additional_columns '["url_source"]'    --output_folder megalith-10m-00002-of-00010 --processes_count 2 --thread_count 8 --image_size 512    --resize_mode keep_ratio --resize_only_if_bigger true --min_image_size 512
92it [26:35:48, 1040.74s/it]
worker  - success: 0.965 - failed to download: 0.013 - failed to resize: 0.022 - images per sec: 5 - count: 10000
$ df -h
Filesystem      Size  Used Avail Use% Mounted on
/dev/sda1       197G  143G   47G  76% /

So, I will try caption them because the size is enough for dense captioning.

alfredplpl

Jul 22

I have captioned 1m images by Florence 2 and T4 x4. The all of captions will be released in the next week from my company.

madebyollin

Owner Jul 22

@alfredplpl That's great news!

animepfp

Jul 22

InternVL2 short and long captions should be done by the end of next month.

madebyollin

Owner Jul 25

Thanks all! Some updates:

@liuliu87 from DrawThings.ai has uploaded some preliminary captions made with ShareCaptioner here as well as archived raw images which should make Megalith-10m captioning more convenient.
I've started a section in the README for linking to the Megalith-10m captioning efforts I'm aware of

alfredplpl

Jul 26

Thanks, @liuliu87 .

I continue captioning the images because we can avoid overfitting.

According to the NeurIPS paper[1], multiple caption is valid for overfitting.

[1] https://arxiv.org/abs/2305.20086

madebyollin

Owner Jul 26

@alfredplpl Yes, having multiple captions available would be ideal! Both to reduce memorization of images, and to reduce the model's reliance on any specific captioning style/format.

animepfp

Jul 29

•

edited Jul 29

No worries. I can get the florence and sharegpt4v captions after, make shortened versions of each, and then put them in my repo for a total of 6x captions per image and link/credit all the other repos.

InternVL2 is a little slow, but the results so far are pretty good.