In case you don't wish to read a lot, this ended up being well made and worth picking up if you ever want an option of making go go.

I mostly just use webui as I just play to learn tricks and understand the design of AI image generation but have spent the last week losing my shit over comfyui's god awful interface design choices and it's node limitations. But I am slowly making a good all-in-one workflow I plan to release someday on here. While I was testing the multiple LoRa setup, I came across one LoRa I obtained (not this one) but didn't test before that ended up required a weight of 1.7 to even effect the generation (and tested today on webui with the same issue). I used additional networks on it and seen that it was made by someone who thought you tagged everything in captions to train a lora even tho it does the opposite. While I was there, I used webui to test out some various char lora to see which worked well enough for my comfyUI testing and I decided to try this one. It works great, at least in webui so far, even using a more realistic (2.9D or whatever people call it) checkpoint.

So I'll say this for those wanting to use it locally at least. Set your main prompt for the first pass up with around a :1.5 LoRa weight or higher if desired. Setup your 2nd pass/hires.fix with a 0.75 or so weight. Setup your main prompt in a minimal fashion then copy that over to the 2nd pass's prompt. Then add all your insane prompts to that first pass that probably just added +600 tokens to the total size. I didn't mean to, but I got a realistic result that looks like it could be real selfie of a person. If you're attempting to duplicate that yourself, I use euler on the first pass, and DPM++ 3M using Karras on the hires.fix & adetailer. and I didn't even need adetailer at that point but I use it anyhow because that's the default work flow. Just be sure to edit your config so the max steps taken can go beyond 150 (I use 1500) and jack up the size of adetailer to 250 on face as it is required for some work but this one would work with 50 or less easily. I'm not sure if forge or others have the same issue with adetailer decreasing step counts depending on quality but that makes a difference (tho shouldn't on this LoRa).

So good work with this LoRa jollyboy. Works well with lower and higher weights without shitting itself as most do. I didn't check your captioning work but it doesn't matter much when the training results just work.

Note: I know someone will probably end up asking so here, I used iLustMix v5.5 as the checkpoint which produced insanely good results. Newer version(s) looked less realistic so I'll probably keep 5.5 and the newer versions if I were to upgrade it so I also have a less realistic option so don't just get the latest one. I also used the lazypos embedding and also did it with lazympos, lazyreal, and lazynsfw embeddings as well since they sometimes help, hurt, or do nothing and it didn't seem to alter things in any negative way so left em in as they played well with this lorand prompts. And lazyneg is in the negative prompts. If I get a chance, I'll upload some samples gallery since I don't see it being shown off well on here.