Hacker News

Automatic1111 for Apple metal, 40% speed up sd1.5

47 points by dmikey831 5 hours ago | 18 comments

vanillax 4 hours ago

SD 1.5 is so insanely old at this point im not sure the point of this when you have Mini Max H3 on metal https://news.ycombinator.com/item?id=49252179

dmikeyanderson 3 hours ago

old but relevant, it's like a production system. We can continue to tune and squeeze 1mp pictures out of it in high quality on relevant localized hardware without breaking a sweat. 1-3 seconds on a m3 system, and then cascade of tooling below stays the same.

SV_BubbleTime 3 hours ago

A1111 also had its time. Comfy takes a couple hours to learn, and then it’s vastly superior. Like, if you are already on Mac there is no reason to also kick yourself in the face trying to go back to 1111.

qclibre22 3 hours ago

There's also sdcpp ( https://github.com/leejet/stable-diffusion.cpp ) . It is for people that know command line shells like bash or Powershell. SDCPP is like llama.cpp but it is for using text to image AI models. Not everyone likes the node/graph programming interface in Comfy or a Webpage interface. Some people need these GUI/web interfaces and get lost and frustrated in shell scripting.

dmikeyanderson 3 hours ago

Totally and sdcpp and drawthings were inspiration for this work. It's not just about the UI, there's a whole stack of plugins, APIs etc. It would be a bigger shift, than just use other UI. Composition etc.

Plus this was fun to learn where the attention and time spent is eating up. I plan to git into the UNet, and then do some Multi-Model Memory Attention tuning as well, for Refining and Inpaint Swaps for the 8Gb Macs.

vunderba 2 hours ago

A1111 (or rather Forge Neo at this point) can still be pretty useful (for a Windows user) if somebody just want a batteries-included approach because it makes it trivial to move back and forth between masking, manual inpainting, tacking on a high‑res fix and/or upscale, and swapping things out on the fly quickly.

Caveat: No idea how well A1111 and its variants function on Mac though.

These days I mostly use ComfyUI because after building out a node‑based workflow, it’s easy to export as a JSON file with websocket outputs that I can integrate into my programs.

https://github.com/Haoming02/sd-webui-forge-classic/tree/neo

jurgenburgen 4 hours ago

Trigger warning: AI content.

trencedamp 3 hours ago

If anyone is triggered by AI content they surely avoid HN

selectodude 2 hours ago

I love AI content. I hate content written by AI.

It's not serious. It shows that the author doesn't care about showing what he's up to.

He cares about making content to post on a blog.

And that's disrespectful to the reader.

dmikeyanderson 2 hours ago

I'm not an editorialist, I spend my time on the software side of things, but want a detailed technical journal of what happened to achieve the gains. Also why the github link is near the top of the intro which I wrote. I'm not trying to disrespect the reader, but the time of the user of the software. Appreciate your view. I've made adjustments.

jurgenburgen 2 hours ago

I spend all day reading AI generated text in one form or other in my day job. When I come to unwind in HN I’m not interested of more of the same. Seeing “load-bearing”, “landing things”, “it’s not x, it’s y” makes me think of work. I honestly would rather read human text with typos and warts.

IronWolve 3 hours ago

Automatic1111 wasnt being updated, people moved to Forge Neo, ComfyUI and Maestro.

vunderba 2 hours ago

Invoke AI is also really good if you prefer working in a canvas-style environment, in a more traditional graphical way. It easily has the nicest workflows for outpainting/regional/inpainting by far.

https://github.com/invoke-ai/invokeai

dmikeyanderson 3 hours ago

Yup! It was a fun experiment though!

Der_Einzige 3 hours ago

What a crazy repo. Some cyber archeologists need to do a case analysis of the entire diffusion model ecosystem circa the original Stable Diffusion up to now.

smallerize 4 hours ago

The speedup includes running LoRAs, right? Which ones are you using?

dmikeyanderson 3 hours ago

Yes it includes LoRA composition as well, a bunch from the internet, and they work just fine!

yogorenapan 2 hours ago

A1111 is such a throwback. It's been ages since that & AI image generation was cool. Back before the massive amount of slop and when things were just a fun experiment

dmikey831 5 hours ago

[flagged]