Draw Things: AI Generation

As the title says, the app doesn’t generate anything, it’s just a long wait even with a few steps (5 steps). The phone is just getting warm (iPhone 15 pro max, iOS 18.0). Please fix this 😊

an easy ai generative free/opensource app, however i have one problem with the flux model, it keeps crashing all the time on my macbook pro, maybe im missing something i have forgot to download a missing peace to the model ??

So this is first and foremost a very well thought out app. It’s probably as simple as you’re gonna get it, in terms of UI and so on. It’s probably also as stable as you’re gonna get it in iOS land. My 12.9” M1 (8gb) iPad Pro is remarkably stable despite all the load it’s under when generating complex SDXL or SD images. It’s a shame that the M1 heats up so fast and needs to be throttled. The only things I’d like to see are more samplers, if possible. On stable diffusion sites I see samplers not available here (like DPM++ 2M SDE Heun Karras, and similar names but with 3M and or Exponential in it). That together with some resolutions not being available makes it impossible to replicate the same results you see on sites like CivitAI etc. The other one is I’d like to be able to copy the trigger word from the Embeddings list and that it automatically adds < > around the trigger word as those signs needs to be there (at least in the positive prompt).

Things work in very unintuitive and non-standard ways, and functionality is hidden away in weird places. As an example of non-standard behavior, hovering over a button doesn't give you a tooltip, instead you need to press and hold the button. As an example of hidden functionality, consider LoRA training. You click the LoRa selector button, which opens a dropdown where you select "manage your models". Then you get a popup window listing your LoRAs. In the bottom right corner of this this window is a tiny weird symbol no human on Earth has seen before. Click on it, and the LoRa trainer window opens. But even worse than the wonky and illogical UI, is tha fact this app loves to squirrel away files, without telling you where it puts them, or how to delete them! Every image you create is added to a sidebar on the right. You can delete these images, but only by right clicking on them individually, and I have no idea where they are being stored so I can get rid of them. Since an avid prompter will soon generate tens of thousands of images this is obviously completely unacceptable. Another example of squirreling away files is when you open that LoRA training window. There are a few buttons there, and since this UI has taught you the only way to get a tooltip is to click on a button you will of course do that to try to figure out what the buttons are doing. When you click a button called "automatic", for no other reason than to try to figure out what it does, it automatically starts downloading over 2 GB of files for BLIP captioning. It didn't warn you this was goign to happen, it doesn't ask you if you want this, it just goes ahead and does it. And then it squirrels away those files somewhere without telling you where it put them. I've looked in all the places I could think of, and I couldn't find them. That's over 2 GB of wasted hard drive space for installing a functionality I do not want to use. The general feeling you get from this app, is that only the developer knows where things are and what they do. And there is no documentation, so you're stuck with trial and error to figure out what the non-standard behaviors are. I mean, have you ever seen a slider that does undo before? How is anyone going to guess that is what it does, especially when it has no caption? Using this app is like playing a game where you have to use the TFGH keys to move instead of the WASD keys used by every other game on the planet.

First, it’s pretty marvellous to be able to run Stable Diffusion locally on your iPad. Thumbs up for that! I have a crash problem though. If I try to render with SD 2 in 768px the app crashes. I have a maxed out 2020 ipad Air so it really should not crash.

I’ve been using Stable Diffusion on desktop for over a year on Nvidia based machines, and I never expected to be able to get the flexibility of using models, LORAs and even ControlNet that this app provides. It’s phenomenal that all that functionality is shoehorned into a mobile app, and I applaud Liu Liu for pulling off a great feat, well done. Keep up the good work! Would be great to see M4/M4 Pro optimizations for the desktop app as well.

This works on just 3gb ram flux,sd3 large 8bits I’m learning the app and I’m having a blast😂👌🏽😊

This is the best on-device AI image generator for still images for both text to image and image to image generation. It is significantly faster and easier for beginners than ComfyUI. It is easy to import models from HF and civitai, and is compatible with SD(1.x, XL, 2.x, 3.x), flux, and all the loras that are available as well. I am running this on a M4 mac, and generation times for non-turbo models are around 1 minute, over or under depending on size and step count.

I absolutely love this app. Any checkpoint, lora, TI, etc that you want to use can be easily imported. Generation is pretty quick- SD1.5 images generate super quick and SDXL don’t really take much longer, especially if you use lightning or turbo checkpoints / loras. You can write your own scripts if you’re a real go getter. Any subpar output you get from this app is almost certainly due to bad input (prompting, config, etc). I have one gripe and it’s really stupid, so if you’re on the fence for some reason and reading this just stop reading now. It’s a 5/5 apply complaint is pretty dumb and I may be the only person who has this complaint- why is tiled decoding on by default for SDXL checkpoints? Sure, I can save config but every time I switch to a different model, I either forget to load a saved config or tweak other settings only to forget that one and start a generation with it on. Since it’s default on, it doesn’t matter if it was off last time I used that model, when it’s reloaded, tiled decoding is back on again. And the toggle is all the way at the bottom. None of that other stuff down there (hires fix etc) is on by default, why this? It’s not like the default image generation is THAT big that it’s necessary, it’s only 1024x1024. So I don’t understand that part. Anyway great app, love it and thanks for your work!

I can ideate on the go. Thank you, Liu Liu. You are a blessing, and an asset to the community!