VERSION 1: (experimental)
REQUIRES DETAIL DAEMON CUSTOM NODES.
REQUIRES FOOLHARDY 4X UPSCALE MODEL.
(both are available through the comfyui manager)
VERSION 2: Better Results
REQUIRES DETAIL DAEMON CUSTOM NODES.
REQUIRES REALESERGAN 2X UPSCALE MODEL.
REQUIRES KIKO TOOLS CUSTOM NODES.
(both are available through the comfyui manager)
version 2 is a 2-pass BASE model workflow focusing the detail daemon nodes pushing detail noise on the second pass as opposed to running it in 1 pass. this gives the model a chance to develop core details and overall composition BEFORE pushing the noise sampling detail daemon on top to finish the generation. This version works ALOT better! you can adjust both samplers step counts to whatever you feel necessary, the workflows values are just the best ive found so far after testing for a few hours. i do suggest keeping the second sampler lower than 15 steps. it creates TOO much detail when you give it too much room to throw noise over the image. trust me 10 to 12 is enough. if you want slightly more detail, lower the "start" value in the daemon node to .40 and lower the "end" value to .60 to begin filling the image with noise earlier which forces the model to clean it up with more detail. give it less steps if you do.
Alternative settings:
Detail Daemon Node
start - .40 (currently .60)
end - .60 (currently .80)
the workflow is set to run the detail sampling at the end of each step, these alternate settings begin with natural sampling then in the middle of the step it introduces noise with the daemon node then stops early to give the model a chance to smooth it back out. you can also drop the "start" to .20 and the "end" to .40 if you want to begin the sampling step with extra noise but i recommend the middle or end settings as it gives a better output.
PROMPTING MATTERS!
whatever you want to be scaled in detail, you need to prompt for. if you want their eyes to be detailed, PROMPT FOR DETAILED EYES. the sampler and model work together to focus prompt specific details.
works best on close-up portraits. characters far from focal point are sometimes blurry.
this is an experimental workflow for adding detail daemon and foolhardy 4x upscaler to z-image base workflow.
WARNING: this workflow is extremely violatile and doesnt always yield the best results. that being said, running multiple seeds can sometimes bring out some AMAZING generations. helps if you just want an extra pop in some iterations and recompression back onto smartphones hides the artifacting that occurs in some generations because of the upscaler. (youll see)
dont come for me, i warned you its not perfect lol