Really cool - keep at it.

Could you give an example of these datasets?

I think they should be easy to find (I never actually used one, but I keep on seeing references...) here's one

https://huggingface.co/datasets/cognitivecomputations/Wizard...

Some starting points:

- Find old datasets (hn.algolia.com "datasets", use huggingface, search arxiv)

- Use weird search engines, e.g. exa.ai (searches based on embeddings vs. google's pagerank/keywords) + google dorks. weird input = weird output

- mix 2 ideas you see - look at showcase channels on discord and pages on random frameworks, things people are building at Buildspace

- Find old facebook groups with lots of grumpy members posting regularly

These are just random strategies I use before I make things I post on twitter (https://twitter.com/joshelgar), but they work pretty well for coming up with fun projects.

Thanks, these sound good. I don’t understand the one about “finding fb groups with grumpy members” - what then?

Often 40+ yr old people complain about problems that you can solve e.g. vet technicians don't like veterinary software - you can fix that somehow.

How do you find grumpy men facebook groups? I would love a strategy

1. GPT4 "500 tough professions w/ hardship involved" 2. Multiply that with "500 shifts in <profession> or <industry> regulations> since 2000" 3. Get inspired to come up with - grumpy fisherman mad about mooring fees, dog walkers upset about walking limits, 10x increases in flytipping in blackpool 4. As your ideation improves, train a better gpt.

Frankly all those FB groups (min. names) are probably in the GPT training set, just a case of finding good prompts to get inspiration.

Why are they benchmarking it with 20+10 steps vs. 50 steps for the other models?

prior generations usually take fewer steps than vanilla SDXL to reach the same quality.

But yeah, the inference speed improvement is mediocre (until I take a look at exactly what computation performed to have more informed opinion on whether it is implementation issue or model issue).

The prompt alignment should be better though. It looks like the model have more parameters to work with text conditioning.

in my observation, it yields amazing perf at higher batch sizes (4 or better 8). i assume it is due to memory bandwith and the constrained latent space helping.

However, the outputs are so similar that I barely feel a need for more than 1. 2 is plenty.

I think that this model used consistency loss during training so that it can yield better results with less steps.

...because they feel that at 20+10 it achieves a superior output than at 50 steps for SDXL. They also benchmark it against 1 step for SDXL-Turbo.

Sharing another weekend project here - every "Launch HN" post ever made, coloured according to how many points they received.

Code: https://github.com/JoshElgar/HNLaunches

Data: https://github.com/JoshElgar/HNLaunches/blob/main/src/app/da...

I've open-sourced it here: https://github.com/JoshElgar/foodpixel

The tech stack was NextJS, Tailwind + Vercel.

Hey, had a similar idea, would be great to chat - can reach me at my username on google’s mail.

1. An infinite crafting game: https://foodformer.com

2. An embeddings-based job search engine: https://searchflora.com

3. I used LLMs to caption a training set of 1 million Minecraft skins, then finetuned Stable Diffusion to generate minecraft skins from a prompt: https://multi.skin

I love the skin generator

Inspired by Neal's "Infinite Craft" (https://news.ycombinator.com/item?id=39205020), I made a food-based version. You can use the sidebar to create a new food and click or drag the example ones.

Saw this on Neal's twitter a couple of days ago, it inspired me to make the food version :)

https://twitter.com/joshelgar/status/1750141793377686000