8.8k
2.1k





At the time of writing, Comfy-Manager may fail to properly install the SD Prompt Reader family. Manually install if it doesn't work from here: https://github.com/receyuki/stable-diffusion-prompt-reader
DONATE:
https://ko-fi.com/lopi999
(despite the name this workflow could very easily work for non-vpred noob/illu models)
NOTE: By default, V3 and upwards use a CFG++ sampler on the first KSampler node by default. CFG++ samplers NEED to have lower CFG values, around 1.0 <= x <= 2.0, or else they will produce oversaturated images.
V6:
What happens when you decide to push updates before testing them? Well, normally nobody sane would even dare to pose this question, but ComfyUI sure did and it answered for itself. A lot of stuff broke when Comfy pushed out the new frontend update, so, this forced me to have to scramble together a new workflow.
Replaced Effeciency Nodes LoRA Stacker because it was heavily broken in the new update
Removed prompt composer because without being able to save space by turning widgets into inputs, it lost it's purpose
Added a token counter
Changed the spacing of nodes and groups because now the minimization of nodes has been decreased
V5:
I began designing V5 because I wanted to add a bit more features to V4, while also trying to figure out how to make the overall workflow much cleaner.
Added a prompt saver section, that uses the SD Prompt Reader node family to function. This allows for images to be saved with a lot of metadata intact.
Efficiency nodes have been replaced with EasyUse nodes, and a global seed has been added. This prevents a lot of seed issues I was having with the Efficiency node family.
ModelSamplingDiscrete node has been removed, Comfy by default auto-detects the sampling method to use and already knows to use ZSNR for Vpred Noob models.
TextBox3 and a few other nodes have been gutted out, the node that originally fed into the positive input for the loader is now a Prompt Composer Node. The ugly mess of many individual textboxes with a switch for artist mixes has been gutted out and replaced with a node I made recently, compacting the area of the Artist Mixer group and allowing for a built-in way to index or randomize the mix used.
V4:
Was going to post this earlier but the site was down for whatever reason.
Added a LoRA section, ability to either specify specific LoRA or randomize, and a TextBox node for LoRA tags.
Added a number to int node because the random number generator node has issues just generating general integers, so now it will generate a float that gets converted into an int.
A few size adjustments have been made due to the new node structure of recent ComfyUI updates.
V3.1:
Not a big update, I just added a bit more to the upscale group section.
Now uses a concatenated positive prompt for the USDU node, which consists of only the artist tags and quality tags.
Now uses a ControlNet loader to load the NoobAI EPS Tile node.
V3:
This version is essentially the best I think I can get this workflow to be. I've removed a few things that I've never touched myself much, but still added a bit more to the mix.
Added an artist tag switcher, so you can easily switch between (or randomize) artist styles
Added an upscaling section to the workflow
Generally streamlined the workflow a bit more.
I've also noticed that the hires fix node native to the effeciency node family seems to not have issues now? I'm not sure if I want to make a version of this workflow that has it though. If you want, you may replace the hires sampler section with that node and see if it works for you, this should be easier to use XYZ plotting with as well.
V2:
Added v2 of this workflow. In the initial version, I noted that there were issues with the hires fix script node, that I initially had no clue about until I did some further testing. This updated workflow changes only one thing: replace the hires fix script with another implementation for hires fix. In short, there is a latent upscaler node that goes into another KSampler. Treat this KSampler like a hires fix with it's settings.
This workflow is designed specifically in mind for NoobAI VPred and any merges or finetunes based on NoobAI VPred. Designed around easyuse nodes for ease of use. Not meant to be really overly complicated.
Utilizes the Mahiro node to add more weight to the positive prompt for CFG
Has groups to accomplish variation seeds, XY plot, and Facedetailer for faces and hands
A straightforward setup for hires fix
Notes for different artist mixes to use
Easy to use latent size picker to set any standard SDXL resolution