guernikacore / guernikakit Goto Github PK

any chance for Guernika to support ipadapter / ipadapter plus?

At the moment I stopped experimenting with LoRAs, as it's crucial for us to "hot-swap" them. E.g., have one SD model (~1Gb), and multiple LoRA models (~30Mb), and pick which one to use. Baking LoRAs into the SD model works great for testing, but having multiple heavy models for each LoRA in the project sucks, so I'm still waiting for some info on official LoRA support.

apple/ml-stable-diffusion#206 (comment)

Hey, there! I'm curious about how prompt weights should be formatted. Does it follow the usual format? For example: a man standing in a field (of wildflowers: 1.5).

Hello - first of all many thanks for these kits - very much appreciated. Just a simple question: what re your plans for long term support of these repos and do you have a roadmap? I know you must be really busy with your published app(s) ....

thanks again. Patrice

Hi, I have been trying to run StableDiffusionXLPipeline on my ipad. However, it seems that there are some issues in TextEncoder.

I found that StableDiffusionXLPipeline TextEncoder1, StableDiffusionXLPipeline TextEncoder2 and StableDiffusionMainPipeline use the same TextEncoder. Does it also support SDXL?

        // Stable Diffusion 2.X TextEncoder
        self.hiddenSize = 1024
    } else {
        self.hiddenSize = 768
    }

I remembered the hiddenSize of TextEncoders in SDXL, one is 768 and one is 1280.

Please advise. Thanks!

Is there any place I can find documentation on this project?

I have been trying to get guenika kit working with different models and keep running in to this error through different pipelines..
let inputImageShape = metadata.inputSchema[name: "z"]!.shape
GuernikaKit/Encoder.swift:41: Fatal error: Unexpectedly found nil while unwrapping an Optional value
My models, don't work in the App Store Guernika either, which looks like is behind guernika kit on samplers

Models I had trouble with are turbo models, which were either converted by someone else for sdxl-turbo, or I couldn't get to convert with your program so I used terminal with sd-turbo, which I have running on an apple coreML stable diffusion with lcm. I am pretty sure that the issue is with the way that Guernika looks for metadata, that my models seem to have in a different format. Is there currently a way to bypass the metadata check and give the variables manually? or is there a way I could add metadata that Guernika likes to the models manually?

here are the models that are causing issues
https://huggingface.co/collections/BloggsMr/coreml-models-6586d877bbb04840e35a8d5c

Here is the code I was using, which works fine with your models.

        do {
            let sampleInput = SampleInput(
                size: CGSize(width: 512, height: 512),
                prompt: "a pretty bottle",
                negativePrompt: "",
                seed: 123456 ,
                stepCount: 2 ,
                guidanceScale: 1.0 ,
                scheduler: .lcm //tried others too
            )
            print("0 - settings: \(sampleInput)")
            let filepath = "\(vars.ModelFolder)/\(vars.CurrentModel)"
            let baseUrl = URL(fileURLWithPath: filepath)
            print(filepath)
                let specificPipeline = try GuernikaKit.load(at: baseUrl) as!  StableDiffusionXLPipeline
                if let image = try specificPipeline.generateImages(input: sampleInput) {
                    vars.GeneratedImage = image
}```

Edit- The answer was to only use models from Guernika huggingface or convert using Guernika Model Converter, But this is useful info that will probably affect other people so I'll leave it here.