Hi everyone I'm just sharing my workflow, I don't have the energy right this moment to do a full writeup here, but I did add some super messy Notes as instructions for the chaos, in the workflow.
This workflow uses the new L Clip model from zer0int on HF.
This Workflow is designed to give you better photorealistic outputs when using your Character LORA's trained on civitai. I find I am able to generate much more interesting and consistent images by using a Split Sigma setup, which allows me to use the extra noise from the High sigma to do the first pass generation, this gives me extremely detailed skin at times. I will edit this article soon and add pictures just wanted to share this with my Facebook friends.
Here's an example image I made, one of the first. I don't have to cherry pick like I normally do when using a standard model only lora workflow! every image just like, works. Short prompts I noticed get me very nice amateur style pics (i didn't use any style lora to make this!)
Basic how to use it:
STAGE 1 prompt: describe what you want in few words for L clip. don't include your LORA trigger word. you can even disconnect lora from this stage if it's really messing up prompt adherence. You can also add some words to T5.
Stage 2 prompt: start with your trigger word and class (cause this is how I trained my LORA, i don't know better okay until I train differently and SEE better results for myself) so I start with "LilaLaRoux girl, " for my L clip 2A then I leave 2A T5 blank. This is because I am also doing DUAL PROMPTING for stage 2 !!! Why?
Because, now I am setting a FLUX guidance of 4 for the L clip, and leaving T5 blank. BUT on the 2B prompt, I leave the L Clip blank. and here you may write a detailed T5 as long as you'd like, this 2B t5 prompt has a flux guidance of 2 for photo realism. I find setting L clip prompt to Flux guidance 4, and T5 prompt to flux guidance 2, makes for an awesome photograph!
TLDR - I made a LORA of myself on civitai. The results rarely look nice. With this workflow - basically EVERY generation looks great. it's just harder to do cause of all the different prompt boxes lol.
NB: this new L Clip only takes 77 tokens! so keep your promp around 15 words hard limit (excluding "of" and "the" basic words). if you REALLY need to describe more, use the multiple T5 prompt boxes at your disposal.
PS - sorry about the workflow I know most of you hate such pasta. I am super ADHD and get extreme satisfaction putting everything as small and close together as comfortably possible so it all fits on my 1440p monitor lol. I will make a nice version soon when got time so newer users can learn how it works.
Also credits to the random reddit posts I found that taught me things like dual sigma setup, and the L clip guy zer0int if I didn't see you combining conditioning I would've never set this up! Any to many more along the way whos work integrated into my brain. thanks folks. Please roast me :D
More gender-bending examples of my stupid face: (all native generations in 720p no upscale!)
Example 3:
Example 4:
Example 5:
Example 6:
Example 7:
Example 8 (flux clef chin aaaah):
Example 9:
Example 10 (this is BAD, i included it to show a flaw in my LORA! so you know how you should label everything, except what you want it to learn? ... so I have a face tattoo under my eye. a small hollow heart tattoo. I never captioned it on purpose thinking it would learn it as part of my skin. nope. often this LORA makes little black marks under my eyes most times barely noticible. cause it doesn't know what it is! I should have instructed "she has a small hollow black heart tattoo under her eye" so it could learn it! also i only did 20 images lora, some of which were nudes. looking back now I now know to train these things separately. So I'm going to retrain my face, and train my body separately (with a LOT more images and steps and lower learning rate to learn the concept of what I look like nude too)
This is the bad one ^^ it happens when you don't caption your tattoo's! in fact all of my tattoos are broken cause I purposefully removed captions of them that JoyCaption made. Like I have a tattoo of a rose on my hand too... so not OFTEN it draws just random ass pictures on the back of my hand. whereas if I had captioned "she has a tattoo of a rose on the back of her hand" it model would QUICKLY learn what that specific rose tattoo looks like. so YES it's true you don't want to caption what it should learn, but that doesn't help if it is something weird like a tattoo etc. For example don't caption the hair color or style or eyes and it learns those successfully, cause it already knows how to make other similar things. tattoo's on weird body areas are all new concepts I guess.
Anyway big fat ramble cause I take vyvanse for my ADHD and it makes me YAP so I hope you enjoy my workflow, toodaloo! :D
PS -- I'm training a Tgirls lora for flux! release in next few days. realistic girl dick and puffy nipples incoming