Hacker News Reader

Comment by boppo1 | Hacker News Reader

Why isn't this a threat to Stable Diffusion? I see more super high quality stuff from midjourney that I struggle to ID as generated than I do SD.

View on HN

Stable Diffusion is open and fully deterministic: a given version of SD+tools+seed shall always give exactly the same output. The model is available to everyone so you run it locally.

Which means there are countless (free) amazing tools around SD.

StableDiffusion is threatened by exactly nothing.

(others have mentioned that SD shall happily generate porn: I don't care about that... But I care about SD being the actual "open AI").

Maybe I'm doing it wrong, but I didn't couldn't get SD+tools+seed to be deterministic.

Images generated, with the exact same settings (including seed), on m1 laptop are not the same as the images from my nvidia GPU desktop with the SD-webui.

Because SD runs locally and produces the kind of content people want.

true, fortunately there are many more things to create than nudity, sexually explicit, and political figures

“so for everything else, there’s Midjourney”

SD is also free and there are custom models for anything you want whether it’s creating landscapes or anime waifus. And those specific models are better than the general model that Midjourney uses which is “good enough” at everything.

Not only does it have custom models, you can train your own custom model. This is a big deal for many workflows that Midjourney and DALL-E just can’t accommodate (at least for now!).

Yep. 99% of what I do with SD is generate images of my wife, daughter, and dog.

Dall-E also doesn't allow you to say "in the style of X" and similar. Which I guess appease some concerns from artists, but also makes it harder to guide it where you want it.

So it's not just about wanting to make explicit images, it's about not having a different entity control what you can and cannot do.

I'm able to observe that as well, I'm also able to observe that caring about that isn't necessary in most use cases and the convenience still wins.

I do branding and presentations all the time with graphics from Midjourney and rapidly iterate in that discord chat in various branches compiled in parallel all day. Make something as convenient, cheaper, private/offline, and faster and I'm there. Otherwise the ideology is irrelevant to me, and the unsaid secret is that that's true for most everyone else too.

It was more a rebuttal to you claiming that it only has explicit images going for it. Almost like you were labeling those using SD to use it only for your claimed reasons.

Midjourney is good for inspiration, but as more people use it things that are just whole cloth generations are going to start to look same-y and cheap. Some iteration/post processing will be necessary not to look low rent.

with SD you can use LoRA, and "artist style" is, to my understanding, the easiest thing to make a model of. With LoRA (Low Rank Adaptation) you describe everything in an image that is not what you want the model to model. With a style, You can just use BLIP/CLIP or deepbooru to describe all of your images. At worst, you might have to remove other styles/artists in the tags. the model will only learn about the style. supposedly. I don't follow art enough to have a favorite style; i don't have enough source material on hand to do a style model, nor do i know enough to go out and grab a dataset.

As an aside, textual inversion also made some great inroads into this sort of thing. So smallest size to largest: textual inversion(megs), Locon/lycoris/lora(tens of megs), full model (gigs!). The accuracy range over all compositions follows the same respective order, as well.

Loras are 1 to hundreds of megs for sd15 (I've seen 1mb Lora and also 700mb Lora merges) and I think around 1gb+ for sdxl.

Hypernetworks exist as well for sd15 at least, they're up to I think around 80mb?

> the kind of content people want.

Well put. I was wondering how that aspect was going to be acknowledged & phrased here.

Stable Diffusion + fine-tuning is quite powerful for creating specific art styles that aren't easily described to DALL-E / MJ

Because SD has way less commercial restrictions and provides far more control than a text prompt alone ever could even with a real human on the other end.

Stable Diffusion is extensible - for example, ControlNet images have gone super viral on Twitter in the last few days. https://twitter.com/0xgaut/status/1702394230478360637 https://twitter.com/deepfates/status/1701055664603426970

Also, it's basically free if you own a Mac or an iPhone.

It is completely free even if you don't own an Apple product

SD is uncensored, you can generate anything you want with no "safety" rails to stop you.

Just take a loot at civitai to see the kind of finetuned models that are out there.

I do feel Midjourney is better in the "press button and pretty image pop out" but if you want to have control over the process and results then tools like Invoke are generations ahead.

Because people can make porn with locally hosted Stable Diffusion.

They can't do that with MidJourney and DallE.

Reply on news.ycombinator.com