Qwen edit can take an image as a sample, and work with that. “The character in a victorious pose” would get whatever character you have, and reproduce it in a victorious pose. Couple of examples:
And a little janky because it IS generative AI after all…
Edit: and a bonus screenshot showing how little effort I had to put into this lol
Those are quite the minimum specifications to run it! Damn! Have they ever heard of “optimization”? Because in 4 minutes, SD 3.5 medium got
out of my iPhone 13 Pro, running locally… with a total of 6gb of system ram. So I’m going to say that 32 gb AND a RX 7900 are a little laughable (even if much faster no doubt).
I have a few examples that I hope retain their metadata.
Seed mode is… basically, I stopped using Automatic1111 a long time ago and kinda lost track of what goes on there but in the app I use (Draw Things) there’s a seed mode called Scale Alike. Could be exclusive, could be the standard everywhere for what I know. It does what it says, changing resolution will keep things looking close enough.
Edit: obviously at some point they had to lose the bloody metadata….
“Better quality” is an interesting concept. Increasing steps, depending in the sampler, changes the image. The seed mode usually changes image with changes in size.
So, what exactly do you mean with “better quality”?
This isn’t what you asked specifically, but it’s related enough… have a look into https://apps.apple.com/it/app/draw-things-ai-generation/id6444050820?l=en-GB as it’s free, ad free, free from tracking and really well optimized. With that I can run Schnell on my iPhone 13 Pro!
I enjoy this 1.5 Lora https://civitai.com/models/165876/2d-pixel-toolkit-2d it’s pretty neat!
The floor is a carpet, and the shoes is harder to tell but might be a similar situation? Velvet maybe?
I’m guessing there’s a mix. The smallest version is 700 million, possibly the one used to generate the time data reported, but the largest (or not?) still runs with 8gb. If I remember correctly SD3 is supposed to have multiple versions, starting from 800 millions and going up, so this is going to be interesting.
Cool, looks simple enough. Can’t test it on my phone, but for things with the A12 and up (although ram can and will be a problem if less than 6gb) there’s https://apps.apple.com/it/app/draw-things-ai-generation/id6444050820?l=en-GB
Can I offer what I believe is a better option? 1.5 LCM models. 5 steps for a good image, and they’re 5 steps at 1.5 speeds :)
I like this but obviously you can find other LCM models.
How old of a Xeon? Because it won’t be a fast result, but maybe you are fine with it. Back when I tried this, SD 1.5 could do 20 steps at 512x512 on my Ryzen 5 3600 in roughly 7 minutes…
Unless they aim for a specialized model? I don’t have insight on the matter, just a guess.
Is it? The authors all have names that (in my ignorance!) all sound Russian, and Kandinsky was a Russian painter…
Yup, same reason why you can ask for a fox using a crocodile as a mech and get a good result. The model has the concept of all things requested and mixes them (with varying success).
Much simpler way, even if not as cool: try https://apps.apple.com/app/id6444050820 that is free and great. First download a model, if you already have one you like just use the Civitai.com link, otherwise there’s a selection of them and I’d try… Realistic Visions? The 8 bit one saves you a little space. Then use the photo icon in the bottom to import a photo, you want something with the face clearly visible. Use the eraser to get rid of everything other than the face. Now try prompting for an astronaut! Should work already but you can try adding Controls: Tile, Inpainting and Pose specifically.
Performance wise you can go at the bottom of the left side, click on the cpu looking icon and select (if it’s not selected already) Use CoreML: Yes and for the CoreML compute units go with All.
My M2 Mini takes a dozen seconds for a normal generation like that. With the Control modules and Inpainting is going to be a little slower obviously but hey, it’s a fast app.
Not going to lie, I almost gave up a few times. But I can also be stubborn… anyway since this is apparently the first confirmation it works, it’s probably be helpful if I mention that it’s a 12gb 3060. :)
My request is dumb, the UI is glitching a little but hot damn 12 iterations per second! Impressive.
… I’ll check later, but I do remember grabbing the “right one” as I had version 12, so this might very well be it.
Now knowing where to look, I did some fixing by myself! Main issue is that I had CUDA 10 and 12, no 11. Then after going insane about that tiny difference… I landed on something I lack the knowledge to decipher: “PyInstallerImportError: Failed to load dynlib/dll ‘C:\Program Files\NVIDIA GPU Computing Toolkit\TensorRT-8.6.1.6\lib\nvinfer_plugin.dll’. Most likely this dynlib/dll was not found when the application was frozen.”
All I can say is that the file is there.
Yes-ish. The base is Draw Things and the relevant bits are https://github.com/drawthingsai/draw-things-community?tab=readme-ov-file#cuda-capable-linux that isn’t too difficult to setup. The app with the pretty interface is Apple only (the developer one day decided to cram the full 1.5 on his iPhone and that was the start of this. The app has feature parity between the iOS, iPad and Mac versions, the gRPC server is “just” the generation parts decoupled from the app) but there’s a Comfy plugin to use the server.
BTW on Apple’s hardware Comfy is poorly optimized, while Draw Things is optimized. The iPhone XR is the oldest hardware capable of on device generation, and (with the right settings) could do a SDXL 1024x1024 generation. 13 minutes mind you for 8 steps, but also 3gb of total system memory. On the other hand, the iPhone 17 Pro is a third of the speed of my RTX 3060. There’s also a friendly Discord, and the dev clearly enjoys adding support for new, cool models because he’s quick at it but doesn’t share roadmaps of any kind.
Yeah. I really, really like that thing.