Attempts to unblur the background of flux images somewhat, it's not fully sharp, but makes the blur more realistic, use strengths from 0.1 to 1.2, experimental. use unblurred background in your prompt.
It will not always make the background fully sharp, but instead it uses more appropriate, lesser blur for things like portraits, and reducing it fully on things like selfies, etc.
Description
FAQ
Comments (7)
Flux seems to have a serious fetish for bokeh. We get it, you learned how to do depth of field with the camera, relax crazy photog.
the reason why it's not working better is because blur in Flux is largely on model, so you need to amplify the network_dim and network_alpha, 64,64 will be good to solve this
Thanks for the hints! I'll see what I can do, this is only the very first test alpha
I have never properly understood what these settings do other than making LoRAs bigger. Please could you explain and give a very rough guideline on what good defaults for different things being trained would be? Characters, styles, concepts, etc?
@272727 all i know is that it train a bigger part of the model.
# = parts related to camera unfocused
$ = parts covered with focus dataset (train)
full model:
---##-###------#-------##-----###
small LoRa cover:
$$$$$$$$$$----#-------##-----###
LoRa with full model size:
$$$$$$$$$$$$$$$$$$$$$$$$$$$$
@rns96 Okay, I think I'm following. Thanks for your response!
i have tryed same ... here my conclusions and a lora too
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.