Show HN: Generate webpage summary images with DALL-E mini | Hacker News Reader

Show HN: Generate webpage summary images with DALL-E mini

colab.research.google.com

75 pointstxtai25 comments

Meta: I know that official Google Research content (e.g. their AI blog) redirect to the `research.google` domain, but maybe there should be a special rule for the `colab.research.google.com` subdomain, so that it's more obvious that submissions are coming from a Colab notebook.

(don't get me wrong, it would be very funny if Google started using AI to generate summary images for webpages, given how haphazard their Featured Snippets text extraction can be)

This link describes a workflow to create webpage summary images with DALL-E mini. The workflow extracts text at a specified article, builds a summary and then generates an image for the summary text.

This comment describes a workflow to create Hacker News summary comments. The workflow extracts text at a specified article, builds a summary and then posts as a comment the summary text.

Smells snarky. OP isn't allowed to summarize?

I'm joking about how the technique seems to relate to the recent trend in generated text in general and how the comment is somewhat self referential in how it looks like such generated text while describing a use for it.

The OP's comment is useful, and while I probably came across snarky I don't think I came across in a manner where I diminish the value of the comment or the content it's describing.

That picture of "artificial intelligence" it made is hilariously derpy. It looks like frame -1 on the galaxy brain meme.

Agree, that one stood out to me too. Though they are all fascinating in a way.

For example, if you read the Wikipedia article on the War of 1812 and then look at the image generated, it makes sense why it chose frigates at sea. If you've been to the Epcot, the image looks similar to the world showcase around Germany/Italy, but it's original in it's own way.

Any cosmologists who can say if the generated 'galaxy' image follows the mass distribution rules?

When re-running the notebook I get the same results with every execution. I'd like to see some alternatives for the same text. Any pointers on how to achieve that?

You can change the seed. If you search the notebook for 1024, you can replace that argument with a different seed to get different results.

You can also change the prefix (or remove it) to change how the drawings are created. Right now it's set to "Illustration of ". You can try different things like "Sketch of ", "Oil painting of ", "3d animation of " and "Mosaic of ".

You can also change the min/max summary length, which is defaulted to 0 and 60 respectively.

I realize this point is to use AI to generate something new—-but you could also submit these keywords to the Unsplashes API and get some relevant, useful images.

That is a good point if you're looking for an existing image. This post is looking towards DALL-E mini as a start in generating descriptive images for any text.

What’s the cost to use this on a larger scale?

Everything in the notebook is open-source. It mainly uses the following projects:

https://github.com/neuml/txtai https://github.com/kuprel/min-dalle

txtai workflows can be containerized and run as a cloud serverless function - https://neuml.github.io/txtai/cloud/

I assume he meant the computational cost.

OK, I misunderstood there. Running the code, which generates 15 images, on a standard GPU Colab environment takes about 2 minutes. It may be possible to submit a single batch of text summaries to the DALL-E mini model, which would improve performance a good deal.

We looked into it. At current TPU/GPU prices it's about $.60 to $.30 an image.

Was this a cost just for the DALL-E piece or the full text extraction, summarization and image generation workflow?

Just DALL-E

Thanks, good to know.

DALL-E mini not DALL-E, just making sure everyone has the right expectations.

I love collab & I love all these cool little toy models coming out :)

cool, I like it

Reply on news.ycombinator.com