Hacker News Reader

Comment by precompute | Hacker News Reader

This is hardly an "advanced algorithm", anyone can do this on a free colab instance. You need a few images of yourself for training on, and you can make a LORA for Stable Diffusion, and use it for anything you wish. The pricing is only for the ease of usage. People have made LORAs for anime characters, celebs, etc and they work pretty well. See Civitai, it has a large collection of models / LORAs / text embeddings.

These days you don't even need to suss out the negative prompts, you can use a negative text embedding (bad-hands, easynegative) to get good quality images.

Dreambooth is practically ancient now. You don't need to lug around huge converged models trained on a few images and a few tags. You can download a much smaller LORA and include it in your prompt and it just werks.

View on HN

Peak HN comment right here.

Most programmers have never made a LORA and have never used Stable Diffusion. For 99.9% of people outside the programmer class, this is absolutely an "advanced algorithm."

Just because you understand what it's doing under the hood, and just because it can be abstracted into a handful of discrete steps, doesn't mean it isn't advanced.

>Peak HN comment right here.

Where do you think we are?

>this is absolutely an "advanced algorithm."

When someone puts that on his product's page, it's assumed that he developed said algorithm, or at least had a hand in it. Instead, here the real product is the pipeline and not the "algorithm". The product would do better if it was honest about what it did.

And like I mentioned, Dreambooth is pretty "old" now. The service could probably be much cheaper if the OP moved to using LORAs, and it would give better results, because it wouldn't clash with the tokens in the underlying model, and could be used with any model.

Guide to LORA:

https://imgur.com/a/mrTteIt

https://old.reddit.com/r/StableDiffusion/comments/11vw5k3/lo...

Dreambooth gives better results than LORA. Especially if you can train on Dreambooth with EMA enabled, which you need around 20GB of VRAM for.

Though, I agree with you that the real product here is the pipeline. The parent comment is also correct though in that the majority of people aren't in a place where they can train their own models yet despite it being possible for them. This product is taking advantage of the current arbitrage opportunity.

Step 2: draw the rest of the fucking owl

If that comment is Peak HN, then this is a great example of why I love HN. Brutal technical analysis, combined with the hacker spirit of teaching others (thanks for the links in the sibling comment GP!)

GP isn't saying "this product would never work" (which is reminiscent of the classic Dropbox comment), GP is saying "this isn't 'advanced' and you can do it yourself" which is really helpful and exactly the sort of thing I love to see on HN.

Reminds me of the dropbox comment many years ago https://news.ycombinator.com/item?id=9224

Getting LoRa/Dreambooth to work reliably without human intervention is HARD, and Dreambooth still works better than LoRa in most situations.

Sure it’s "easy" to do it by hand, on a few persons, tweaking the parameters. Doing it at scale at good quality is still a challenge.

I see people spend HUNDREDS of dollars for a professional headshot. I would totally pay $9.95 if I was fairly confident the results would be worth the time.

Reply on news.ycombinator.com