This is what I'm doing:

• cloned the repo
• went with Option 1 and downloaded the zip, unzipped the "bins" folder
• opened the SimpleDiffusion app and changed the URL in the task modifier to point to the local path:

            let path = URL(fileURLWithPath: "/Users/xyz/Downloads/bins")
            do {
                try await sd.prepModels(localUrl: path)
            }

• additionally set the inputImage to my input image:

            inputImage = NSImage(contentsOf: URL(fileURLWithPath: "/Users/xyz/IMG_1076.JPG"))?
                .cgImage(forProposedRect: nil, context: nil, hints: nil)

When I run the app, I set the prompt to "cat" and start the conversion. The process begins and then crashes with:

The file “first_stage_model.encoder.conv_in.weight.bin” couldn’t be opened because there is no such file.

Thought you'd get a kick out the fact I used the app to auto generate it's own icon. I tried many times but the first one was the best by far so it feels like fate.

I'm getting the following error: failed assertion 'length(16640) must be <= 16384.'

Steps to reproduce :

• Create a new projet
• Add native-diffusion as a dependency (branch: master)
• Download pre-converted bins from here
• Run python web server that serves bin.zip from localhost:8080
• Use SimpleDiffusion example here

I launch the app, I type "astronaut cat" in the TextField. And after a few seconds I got this error :

-[MTLDebugComputeCommandEncoder setThreadgroupMemoryLength:atIndex:]:543: failed assertion `length(16640) must be <= 16384.'

Specs

Tested on Mac details here

Exported to

• iPhone 14 Pro Max
• iPad Pro 6th gen

Screenshot

Logs

ipad-log.log iphone-log.log

Steps to reproduce :

Created a new projet Added native-diffusion as a dependency (branch: master)

Download pre-converted bins here : https://drive.google.com/file/d/1EU_qxlF-p6XrxsoACvcI2CAmu45NphAC/view?usp=share_link

Copy/Pasted code from Readme.md, and chanded it to localUrl :

let path = URL(fileURLWithPath: "/Users/xyz/Downloads/bins")
do {
  try await sd.prepModels(localUrl: path)
}

I launch the app, I type "astronaut cat" in the TextField. And after a few seconds I got this error :

-[MTLDebugHeap newBufferWithLength:options:]:123: failed assertion `Buffer Validation
newBufferWith*:length 0x20000000 must not exceed 256 MB.

When you pass an input image to the generator, it currently doesn't check if the image fits the model (currently we only support SD1.5, so it's always 512x512)

This warning is for all users using img2img via an inputImage

The warning could be an assertion failure, but open to ideas.

Next: Resize off-size images, using a center crop. Feel free to go for this directly. Later: Support for SD2.0

2.0 just came out today. Blog post, repo

Highlights

• Generate 768x768 images (weights)
• New Upscaler model that can scale up to 4x in one go (weights)
• Depth to image for shape-preserving img2img
• Improved inpainting (weights)

We probably want to support both 2.0 and 1.5