Thanks for keeping us macOS users up to date! Great tutorial. Much appreciated.

Keep up the good work Laura!

Thank you so much for this Laura! Exactly the video I was looking for as I was about to try Vlad for my Mac. No, thank you. I really hope we have Mac silicon optimized versions coming out soon!

Nice Review!!! Keep going Laura!!!

SD options for the Mac are always improving, and some are better and faster than others. I’ve been trying all the different options lately but this is the first time I’ve heard about Vlad. I’ll be sure to try it out soon! Most recently I got Mochi Diffusion, a native Mac front-end that provides options to use CPU plus Neural Engine, CPU+GPU, or all three together. The CPU+GPU option was the fastest and used very little RAM. A1111 is still the winner for generation speed (3.4it/s), which is surprising, considering that Apple CoreML Diffusion, Mochi, and DiffusionBee all do extra conversion to optimize the models specifically for CoreML. DiffusionBee is the clear winner in terms of the least memory usage, but A1111 isn’t bad. It uses only a little more memory than InvokeAI 3.0.1. The biggest user of memory is (surprisingly) Apple CoreML Diffusion, which apparently preloads the whole model into RAM. It also takes 4 minutes to start up, so it’s not very useful except as a tech demo. The main downside to DiffusionBee, Apple CoreML Diffusion, and Mochi is they all use converted models so they require a lot of extra storage, whereas A1111, InvokeAI, and ComfyUI can all share the same checkpoint or safetensors model files. InvokeAI encourages converting models to diffusers format so that doubles its storage usage. Again, Apple CoreML is the worst because it uses huggingface-cli to download diffuser models, which creates a huge cache folder, and then it converts them to .mlpackage format. Overall my favorite is still InvokeAI, but all of these options are interesting and always fun to use. I look forward to seeing what everyone makes with these incredible new tools.

great vid! grazie

Thanks for the video. I also had the issue with python, but I was able to solve it differently. Btw, Vlad‘s Version is in lack of the API. Yes, it is faster (on my test), but I cannot use it programmatically.

Great video Laura! I have tried running AI applications on low-power computers, with Linux on my outdated PC, and it's only worth running codes based on really small models if the results are of decent quality. As for GPU acceleration, it's possible with some compatible AMD cards through ROCm, with better performance on Linux than on Windows, although it's not comparable to equivalent Nvidia cards. From what I understand, most CUDA-based codes are compatible without modification with ROCm, both for PyTorch and TensorFlow.

As you're using warp terminal, you have zsh as default. So you don't need to type cd (change directory) you just type ".." if you want to go under one more folder, you type "..." same as cd ../.. Also if you want to check the current folder path you are the command is "pwd", ls is to list the files and folders.

Grazie, sto infatti notando una lentezza notevole sui Mac serieM. Lo installerò e lo provo, non ho capito se supporta controlnet allo stesso modo. grazie.

Hi Laura, you compared Vlad vs A1111 just at the very first run of your prompts? What I'd seen here is 1st time is slow, after that the speed increase a lot. Perhaps the ideal would be run the prompts four or five repetitions, and take the average time. Thanks

Thanks Laura! do you know what it means if all my prompts are coming out as just black squares?

Hello Laura, thanks a lot for our videos. I tried running sd with control net and all on mac intel 16gb ram, at the end I could see the stable diffusion interface with all extensions but couldn’t run it , I guess because it wanted M1 chip at least. Will this vlad would run on mac intel laptops?

hi laura i get this error "Cannot convert a MPS Tensor to float64 dtype as the MPS framework doesn't support float64. Please use float32 instead." can you help me ?

Dear Laura, it never worked for me, can you teach us how to unisntall everything we downloaded, please?

If having OS issues why not just run docker containers? I have a few open source CUDA/huggingface/transformers Dockerfiles for this

Very clear tutorial, it doesn't work for me still have a smae error: " RuntimeError: "upsample_nearest2d_channel_last" not implemented for `Half`. Any issue for that's ? thank

Great video. I also have a MacBook Air M1 and run Automatic1111 on it. And also I use (the paid version of) Colab. Thanks for doing the testing for us Mac users :) Are you happy with runPod? How is it compared to Colab? That would be an interesting video, at least for me :)

I’m having issues installing gif2gif script- it shows up, but only converts 1 frame of the gif. Also not able to see control appear on Auto 11 11 - can you make a video Laura? Help!!!

I'm a PC user, I can't seems to see any significant improvement from VLAD over A1111, the startup is slower and for some reason the image generation quality is not as good as A1111 even though I am using all the same settings, seeds and model... most probably there is some additional settings behind in VLAD which makes the output quality downgrade.

When I run stable diffusion on google colab it disconnect immediately. Please help.

I keep getting SSL Verify certificate error how do I fix this ?

Thanks for making this comparison! What are the specs on your Macbook Air, is it the first generation M1?

Dear Laura! Thank you for your reviews on SD, they are very useful) I would like to ask for your help - what am I doing wrong 🤦🏻‍♀? You say that the average generation time in Automatic1111 takes around 40 seconds. When I first installed it, about similar generation parameters I had similar figures, but slightly slower - 43 to 53 seconds. Now 1 generation of 512*512 takes at least 2 minutes and it's horrible🤯 Maybe something is wrong with the settings. Can you tell me where and what exactly I should check to get back to these 40-50 seconds? 🙏🏻

questionL how much does collab to create 5200 frames ?

Hi, can you make a video about trainning SD Automatic1111 model on mac by using dreambooth? I will be very grateful to you. Please.

I am on Mac OS Catalina is there a way I can run Stable Diffusion ?

Can you tell me how to package this project into an integrated package? Let others download it and use it directly, without downloading Python.

omg! why did not say it in the beginingggg

it's automatic eleven eleven, but one one one one sounds like om nom nom nom 😆

comunque puoi chiamarlo Automatic "eleven eleven" come il 99,9% del resto del pianeta 😇

sis just say eleven eleven instead of one one one one... hehe