search

shiimizu / ComfyUI-PhotoMaker-Plus

PhotoMaker for ComfyUI
GNU General Public License v3.0
219 stars 21 forks source link

how is this different from diffusers implementation? #28

Open albseb511 opened 5 months ago

albseb511 commented 5 months ago

Hi i was trying to convert a workflow from comfy to use it on diffusers, but i am seeing that i am not able to get the same output.

I tried with juggernaut xl v7, 10 LCM scheduler lcm lora cfg - 2.0 steps 8, start merge step 1

I tried, euler ancestral, with cfg 5, and 24 steps with 1 also.

So a few things what i observed, was from a different layer. 

shared_id_embeds = self.vision_model(id_pixel_values)[2]

when i move this to 2 in the original implementation, i get out of index error.

In comfy i am getting great outputs, when i try it on diffusers, with similar configuration i am not getting the same output.

Wanted to understand how i can get similar outputs from the node here.

shiimizu commented 5 months ago

diffusers generates noise on the GPU. I did a comparison at https://github.com/shiimizu/ComfyUI-PhotoMaker-Plus/issues/3#issuecomment-1900543421 There are nodes that allow you to generate noise on the GPU like https://github.com/ltdrdata/ComfyUI-Inspire-Pack