inference is so slow about photomaker HOT 15 CLOSED

The speed is amazing. It only takes me less than 10 seconds to generate an image on the A100 with 10 input images.

from photomaker.

10 sencond? but I cost 4 minutes, and I'm sure the code runs on gpu

from photomaker.

In principle, it is not much slower than sdxl.
How long does it take you to run sdxl?

from photomaker.

same config, run 40 steps, generate 4 img, cost 40s

from photomaker.

Hummm... on a 3090 on windows it generate 4 images in 66 seconds @ 50 steps

from photomaker.

Hummm... on a 3090 on windows it generate 4 images in 66 seconds @ 50 steps

Hi bmaltais, thanks for your work with this project. Do you know why it takes me 3+ hours to generate 2 images with default settings on a 2060 gpu? Any way to reduce the time.. by like 4000% haha?

from photomaker.

This is certainly not normal... have you followed the installation instructions in my fork or are you using the default install instruction in the main repo?

Try following my instructions and see if it make a difference...

https://github.com/bmaltais/PhotoMaker/?tab=readme-ov-file#-dependencies-and-installation-windows

from photomaker.

Try following my instructions and see if it make a difference...

https://github.com/bmaltais/PhotoMaker/?tab=readme-ov-file#-dependencies-and-installation-windows

Right! It´s very weird and I´m so hyped to generate stuff.. ! I just tried a clean install of your fork AND python, not using anaconda environments this time even tho it should technically work as well.

Install works perfectly. Generations however, seems to take forever. At least the gpu is at 100% as far as I can tell.

I did a quick test in comfy ui with a different standard model 20 steps and 512px res takes 30 seconds.

Something fishy is certainly going down.

from photomaker.

Try following my instructions and see if it make a difference...
https://github.com/bmaltais/PhotoMaker/?tab=readme-ov-file#-dependencies-and-installation-windows

Right! It´s very weird and I´m so hyped to generate stuff.. ! I just tried a clean install of your fork AND python, not using anaconda environments this time even tho it should technically work as well.

Still seems to take forever. At least the gpu is at 100% as far as I can tell.

I did a quick test in comfy ui where 20 steps and 512px res takes 30 seconds.

Something fishy is certainly going down.

Yeah... this is odd... I have pushed a PR to both ComfyUI nodes to add support for width and height... this should bring the ComfyUI support on par with Gradio. The nodes are still immature but hope things will improve so they become more standard ComfyUI nodes with proper batch image input support and multiple image outputs...

StartHua/Comfyui-Mine-PhotoMaker#4

ZHO-ZHO-ZHO/ComfyUI-PhotoMaker-ZHO#5

from photomaker.

Try following my instructions and see if it make a difference...
https://github.com/bmaltais/PhotoMaker/?tab=readme-ov-file#-dependencies-and-installation-windows

Right! It´s very weird and I´m so hyped to generate stuff.. ! I just tried a clean install of your fork AND python, not using anaconda environments this time even tho it should technically work as well.
Still seems to take forever. At least the gpu is at 100% as far as I can tell.
I did a quick test in comfy ui where 20 steps and 512px res takes 30 seconds.
Something fishy is certainly going down.

Yeah... this is odd... I have pushed a PR to both ComfyUI nodes to add support for width and height... this should bring the ComfyUI support on par with Gradio. The nodes are still immature but hope things will improve so they become more standard ComfyUI nodes with proper batch image input support and multiple image outputs...

StartHua/Comfyui-Mine-PhotoMaker#4

ZHO-ZHO-ZHO/ComfyUI-PhotoMaker#5

Thanks a bunch. Oh are you able to use ComfyUI somehow with photomaker? (might be a nooby question since im an amateur enthusiast) ^^. To clarify i used comfy with a standard and different model than this project. BUT would be pretty interesting to see if the performance is normal with Comfy and photoMaker!!

from photomaker.

Yes, you can use either of those ComfyUU photomaker nodes. The install is not straightforward but it works for me.

from photomaker.

Yes, you can use either of those ComfyUU photomaker nodes. The install is not straightforward but it works for me.

Oh damn thats super dope if it works as well as this project somehow.

Doesnt seem like they wrote how to install clearly in the projects. 🥇 😶‍🌫️ I will report back if this made a difference performance wise later.

EDIT Solved it: If others want to install: Use a git clone into the custom nodes folder within ComfyUI. Then a requirements pip install. 👍 You then have to create the workflow yourself by adding nodes. Unfortunately theres some errors upon trying to run it.

from photomaker.

@akk-123 @chille9 An important note: For those GPUs that do not support bfloat16, please change this line to torch_dtype = torch.float16, the speed will be greatly improved (1min/img (before) vs. 14s/img (after) on V100). The minimum GPU memory requirement for PhotoMaker is 15G.

from photomaker.

@akk-123 @chille9 change this line to torch_dtype = torch.float16

Thank you I thought the RTX series would support stuff like that, great that you clarified this. I can confirm that its indeed faster.
2 generated images at 1024px and 34 steps = 37minutes.

from photomaker.

just for information. rtx 2070: 800 sec for 1 output image with 20 steps. with torch_dtype = torch.float16.

UPD: after upgrade rtx 2070 8gb to rtx 4080 super 16gb speed increased from 800+ seconds to 5 seconds.

from photomaker.