FairyTailor: A Multimodal Generative Framework for Storytelling | Hacker News Reader

FairyTailor: A Multimodal Generative Framework for Storytelling

github.com

55 pointslnyan13 comments

Open articleView on HN

Hmm, the dog seems to like it: https://fairytailor.org/321876ae4da353989aff10887ea94cdc

Seems to have a little difficulty coping with changes of tone ;-)

https://fairytailor.org/c1a5820376965428a9c8dcf531713851

I get a ‘We’re sorry but we do not support smaller screens’

Works in Firefox on Android for me - if I toggle "Desktop site" on.

Ed: the text at least displays - no images though?

Ed2: ah, tab to autocomplete images, I guess original link didn't use visuals? (works with hacker's keyboard on android - otherwise tab key might be tricky..)

original:

https://fairytailor.org/321876ae4da353989aff10887ea94cdc

with a couple of images:

https://fairytailor.org/327740dbfc0e5469a769677138a0e29f

The HANDSOMEST DOGGIE is going to be upset abuot these pictures...

This is so annoying. How about you simply display your desktop site at an acceptable zoom level and let me decide if it is usable or not.

Related paper: https://arxiv.org/abs/2108.04324

Online demo: https://fairytailor.org/

Neat! I'm having fun so far.

Seems like some sort of student project with IBM research? From the bottom of their website:

"Stories are generated using a fine-tuned GPT-2 Medium model provided by huggingface. The stories may contain inaccurate, inappropriate language. MIT and IBM make no warranties or guarantees, expressed or implied, assume no responsibility or liability of any kind and specifically disclaim all responsibility for any liability incurred as a direct or indirect result of using this website. This site may be taken down at any point without further notice."

IBM is doing a lot of interesting stuff around discourse and computational narratology. In particular their Debater program is pretty impressive.

https://www.research.ibm.com/artificial-intelligence/project...

I would be very interested in integrating this into an upcoming art production.

How would one go about finding engineers that can produce functionality similar to this, but recommending sentences from select news articles instead?

You might need something like salience ranking. And you don't need a heavy modern transformer model to do it. You can get by with yesterday's algorithms. This is LexRank. I think it's interesting because it uses graph-based centrality ranking, in a similar way that PageRank did.

Edit: Or create sentence embeddings, and cluster them or rank their distance against an input sentence, or something.

https://www.cs.memphis.edu/~vrus/teaching/nlp/papers/summari...

Looks very very cool

Wouldn't mind using a hosted version , I don't really feel like spending an hour setting this up.

Reply on news.ycombinator.com