Friday, September 23, 2022

Midjourney and Stable Diffusion

Ben Thompson:

The initial roll-out of large language models seemed to confirm this point of view: the two most prominent large language models have come from OpenAI and Google; while both describe how their text (GPT and GLaM, respectively) and image (DALL-E and Imagen, respectively) generation models work, you either access them through OpenAI’s controlled API, or in the case of Google don’t access them at all. But then came this summer’s unveiling of the aforementioned Midjourney, which is free to anyone via its Discord bot. An even bigger surprise was the release of Stable Diffusion, which is not only free, but also open source — and the resultant models can be run on your own computer.

Divam Gupta (via Hacker News):

Introducing Diffusion Bee, the easiest way to run Stable Diffusion locally on your M1 Mac. Comes with a one-click installer. No dependencies or technical knowledge needed.

Steve Troughton-Smith:

Stable Diffusion on my M1 is kinda neat, but slows the system to a crawl for a minute or two while it processes. Stable Diffusion on my 3080 Ti, however, is blisteringly fast. 3 seconds to produce an image in a batch.

It sounds like it doesn’t yet use the M1’s Neural Engine.

Andy Baio (Hacker News):

As AI-generated art platforms like DALL-E 2, Midjourney, and Stable Diffusion explode in popularity, online communities devoted to sharing human-generated art are forced to make a decision: should AI art be allowed?

See also: Dithering, which had an interesting discussion of Midjourney’s Discord-based interface.

Author:

Previously:

5 Comments RSS · Twitter

Beatrix Willius

For giggles I tried Diffusion Bee. On the first start the interface was white-in-white. The developers have odd ideas about interfaces. For instance there is no cancel button when downloading a model.

On my MacBook Air M1 the software was super slow. The results were quite underwhelming.

"The results were quite underwhelming"

That hasn't been my experience with Stable Diffusion. It's amazing. The results it produces typically fall into one of two categories: hilariously wrong (prompted to generate a hot dog eating a hot dog, at one time it made a dog with sausage ears), or absolutely incredible (in my RPG friends chat, we one-upped each other trying to generate pictures for a campaign, and the results were gorgeous).

Obviously an M1 laptop won't perform as well as a top of the line graphics card. They're two different beasts all together. I guess the confusion stems from Apples intentionally misleading slide where they illustrate that an M1 is more energy efficient than a PC graphics card.

Think of it like this, my VW Golf MkV needs less gas than a F1 car when doing 80km/h. That doesn't mean my car is on par with the F1 when it comes to top speeds, acceleration, taking corners etc.

It's still way more convenient for every day use though. Imagine having to park an F1 car outside the grocery store... let alone pack a week worth of groceries in it and then drive home.

Running Diffusion Bee on a M1 Max system with 64 GB RAM and upgraded graphics cores, a single image at maximum resolution (768x768 pixels) and with the maximum number of steps (75) takes about a minute or two. It's a RAM-hungry application, consuming about 30 GB while rendering. The processor cores are not hit too hard; I don't currently have a way to measure GPU core utilization but the temperatures go up by 17˚C for all of the GPU temperature sensors that are reporting, so I know the GPU cores are getting used more heavily. System responsiveness is unaffected while using the program.

Results are a mixed bag; when I tried on my own it seemed underwhelming, but when I took some example prompts I'm able to generate some really amazing scenes. I've been having the system generate images whenever I can; only wish there were a way to have it generate even more at a time so that I could have it running all day. Thanks to the author for putting the software together to make it so accessible!

It wasn't as easy to install as DiffusionBee, but this Stable-Diffusion fork* was a lot faster on my M1 Pro MBP at ~1m30s per image (50 iterations/image) vs 5 minutes for DiffusionBee.

David, you can view GPU utilization in Activity Monitor. Check the Window menu.

If you try DiffusionBee you should know that it dumps a 4.6GB model in ~/.diffusionbee.

*https://github.com/invoke-ai/InvokeAI

Leave a Comment