@Ikena I apologize for my short terse initial comment. Normally that prompt is a placebo prompt, or an unnecessary prompt. Such as "best quality, ultra high res, hdr, 4k, 8k, official art, extreme detailed, highest detailed, intricate details". Prompts like those had worked with Midjourney and many people kept using them with newer models. Others copied those prompts not understanding they had no effect, and/or causing generation degradation.

@GM_MD So that's where those clues came from - MJ.

I tried a lora first time with 2.0 and it didn't turn well. Is it recommended to use less prompting or other useful tricks make it work better? I tested it with a popular Stabilizer IllustriusXL Lora and had no luck with that.

@Chris2000SP well since i haven't tried this checkpoint and dont know what lora you are talking about i can only speculate, but loras normally use strength between 0.5 and 1, if you use 2 most of the time you will get random noise as a result

@mike090y9 Yes, i know that. But i had more Issues than fixes using the Lora i mentioned on the Hassaku XL (Illustrius) 2.0 Checkpoint. I had conjunctions that i put in negative prompt together with this Lora i Mendioned. It has fixed some issues and did on other parts mistakes that wasn't a problem before.

@Chris2000SP well to be fair without knowing what lora you are talking about there isn't much anyone can do to help you, since each lora behaves different, and you also need to consider the compatibility of styles, for example an anime checkpoint like this will have low compatibility with a lora that was trained on a non anime style of images, if the lora base it's too different from the lora checkpoint you may want to use style transfer nodes instead of a lora

@mike090y9 thx. And for the info, i used a Anime and Illustrius compatible Lora.

@Ikena he say he is a newer in stable diffusion, and he want to ask you some advise because this is first times he use your model but nothing he found in relevant posts. i haven't used used sd a long time either.when i use it before, it was still at version 1.5. so i hope to learn about the current situation of the latest model.

When you make an image, you generate a seed, Wich is a serial of numbers let's say, just copy that or a remix using the same seed.

Well yeah if you want more precise accuracy and is not a famous character, Lora for it will help a lot.

Usually at least 20 images I would say , the more images for the Lora the better

Yeah, for me 1.2 still superior

Imo, this is what works best for me; Upscaler: the one that ends with + anime6b, denoise: 0.35-0.45, hires steps 15-20 upscale: 1.5 to 1.95 if I have 832x1216 as base res

Illustrious is better than Pony in every way. But it's been like that for 6 months mate, time to wake up.

Good to know, I'll stick to 1.3b

@Ikena same here, using comphy was able to generate images only with euler_ancestral + sgm_uniform, Hires fix melts an image, and detailer for hands and faces misses colours. Guess there is some kind of an incompatibility with loras i used.

Tested without loras at all - all good, but with styling loras all goes south

@2125152512551 use on tensor i tested. image size is 896x1152

Some stuff is better with 2.1, but 1.3a does seem like a better model overall, this is rather unfortunate (we apprecate your work though Ikena)

yeah, V2.1 gives me a lot of blurry/oily images (with karras scheduler mostly but with other settings too) I'm downgrading to 1.3A

@new7ype yeah

it is okey, I trained some on both base illustrious and hassalu 1.3. they are both good but give different style

@TOF_enjoyer thanks for the feedback, for now I still have good results with noobaiXLNAIXL_epsilonPred11, I will stick with that for the moment

@ToonGenAI is Lora trained with noob will working on models based on illustrious without noob merge like hassaku ? so far i make lora with base illus, im quite curious about this

@Hanekawa01 they work 50% of the times, I will just train for them separately, since training for vpred require additional stuff

@Hanekawa01 I train my lora with noobaixl. It will indeed work with other illustriousXL models to generate images. I usually prefer "prefectious" GoofyAi checkpoints. But Amanatsu checkpoint is not bad too. Generating image with Hassaku is different for me, the color are kind of washed out, maybe it's my workflow.

Try using Sg Uniform and see if that corrects it

@cesa210 I'll try it, thanks. Been mostly using Euler A, Normal

@FondantViper for what i test is either SG Uniform, Simple or Beta

Instead of using the word "lust", I would recommend using Danbooru tags like: "seductive smile, wavy mouth, blush, flustered, half-closed eyes, furrowed brow, nervous sweating, heart, spoken heart, heart-shaped pupils, lovestruck"

If you go to danbooru, click to the wiki tab and search tag_group:face_tags it will give you a list of expressions. Edit: You could also possibly try negative prompting for fullmetal alchemist.

@CHROME2D0ME Ah, thank you so much!

@person012345 Got it, thank you for the tips! I didn't know about that

bro, you feel it too? I've been using hassakuXL v1.3A more than the newest. the outcomes are much better than 2.1

@zetaplus92 yep, still

Simple and Beta also works

I have the feeling is over-trained or something on training process went wrong.

I also notice that i wants to put door frames and doors in several images, without being in the prompt.

If you mean hair intakes you can put tag "hair intakes" in negative to fix it

@MV_QT I re-ran many of my generations with "hair intakes" in negative and unfortunately it made near to no difference. I do appreciate the tip, though, will be adding that every time from here on

@kilvoctu Have you tried increasing tag strength like this (hair intakes:1.4)? There is also Hair Intakes Eraser Lora for Illustrious, but I haven't tested it - https://civitai.com/models/543355?modelVersionId=1132477

@MV_QT lol, there's a LoRA for it? Increasing the tag strength wasn't doing much, but this LoRA is working quite nicely. Thanks so much for your time!