Conversation

Notices

1. Embed this notice
   Sick Sun (sun@shitposter.world)'s status on Saturday, 21-Sep-2024 03:53:37 JST Sick Sun
   you use dall-e to generate high quality images in seconds
   
   I fight comfyui for multiple days to upscale a single god damn image
   
   we are not the same
   • Embed this notice
     Sick Sun (sun@shitposter.world)'s status on Saturday, 21-Sep-2024 03:54:57 JST Sick Sun

     in reply to
     it wasn't working but it was generating fucked up images
     
     now it's generating blank images and I don't know what I did.
     
     welp
   • Embed this notice
     Sick Sun (sun@shitposter.world)'s status on Saturday, 21-Sep-2024 04:01:53 JST Sick Sun

     in reply to

     • Vitor Hugo
     @vhns because I am experimenting and it makes experimenting extremely fast.
     
     When I figure out how to do it in comfyui I understand how to do it in python.
   • Embed this notice
     Vitor Hugo (vhns@yuga.surf)'s status on Saturday, 21-Sep-2024 04:01:54 JST Vitor Hugo

     in reply to
     @sun serious question, why don't you just use something like Tensorflow instead of fighting some GUI program?
   • Embed this notice
     Sick Sun (sun@shitposter.world)'s status on Saturday, 21-Sep-2024 04:03:01 JST Sick Sun

     in reply to

     • Vitor Hugo
     @vhns the truth is I don't know what I am doing and am trying to learn.
   • Embed this notice
     Sick Sun (sun@shitposter.world)'s status on Saturday, 21-Sep-2024 04:11:12 JST Sick Sun

     in reply to

     • Vitor Hugo
     @vhns I'm generating the original image at a specific size. Then, attempting to double the dimensions.
     
     First I am taking the latent image output from the first sampler, and upscaling it so x and y are doubled. Then I am adding noise to it. Then I am feeding it back into a sampler again with the same prompts and controlnet conditioning. the samplers are based on Flux with a couple of loras applied to them.
     
     as I mentioned, the first step works so the models don't seem to be fucked up.
   • Embed this notice
     Vitor Hugo (vhns@yuga.surf)'s status on Saturday, 21-Sep-2024 04:11:13 JST Vitor Hugo

     in reply to
     @sun are you trying to use an autoencoder or something for upscaling? I could try and link some study(ies) on existing networks that work and have existing data on their efficiency instead of hitting your head against the table like that. It would also make it easier to implement them.
   • Embed this notice
     Sick Sun (sun@shitposter.world)'s status on Saturday, 21-Sep-2024 04:11:28 JST Sick Sun

     in reply to

     • Vitor Hugo
     @vhns I am trying to upscale an image but really I'm doing that to learn concepts.
   • Embed this notice
     Sick Sun (sun@shitposter.world)'s status on Saturday, 21-Sep-2024 04:14:18 JST Sick Sun

     in reply to

     • Vitor Hugo
     @vhns so like I know this works because people do it. the problem is that _I_ cannot get it to work.
     
     The best I could do was generate a ghostly mess. I changed something, not sure what, and now I'm just getting uniform noise output.
   • Embed this notice
     Sick Sun (sun@shitposter.world)'s status on Saturday, 21-Sep-2024 04:18:09 JST Sick Sun

     in reply to

     • zonk
     @zonk when you first load it up it is soooo intimidating, but I basically get it now. the problem is "getting it" is just the step before understanding what you're even doing with all the moving parts
   • Embed this notice
     zonk (zonk@shitposter.world)'s status on Saturday, 21-Sep-2024 04:18:10 JST zonk

     in reply to
     @sun i tried for several hours to get it to work. it did not cooperate at all. not very comfy tbhtbh